1

Mobile Money, Rural Household Welfare and Remittances: Panel Evidence from Uganda

Ggombe Kasim Munyegeraƒ

and

Tomoya Matsumoto℘

February 2014

Abstract

Mobile money service in Uganda has expanded rapidly, penetrating as much as over

30 percent of the adult population in just four years since its inception. We

investigate the impact of this financial innovation on household welfare, using

household survey panel data from rural Uganda. Results from our preferred

specification reveal that adopting mobile money services increases household per

capita consumption by 69 percent. The mechanism of this impact is the facilitation of

remittances; user households are more likely to receive remittances, receive

remittances more frequently and the total value received is significantly higher than

that of non-user households. Our results are robust to a number of robustness checks.

Key words: Mobile money, financial inclusion, household welfare.

JEL Classification: O16, O17, O33, I131

ƒ Corresponding author. National Graduate Institute for Policy Studies. 7-22-1 Roppongi, Minato-ku 106-

8677 Tokyo Japan. Contact: phd10104@grips.ac.jp ℘

National Graduate Institute for Policy Studies. 7-22-1 Roppongi, Minato-ku 106-8677 Tokyo Japan.

2

I. Introduction

Financial inclusion1 plays an integral role in reducing rural poverty as it facilitates

saving and borrowing as well as empowering the poor to smooth consumption and

insure themselves against a number of vulnerabilities in their lives (World Bank,

2012). However, a large fraction of the population in developing countries lacks

access to the basic financial services (Asli and Klapper, 2012). Lack of access to

basic financial services restricts the ability of the rural poor to make savings and

investments and engage in both formal and informal insurance mechanisms aimed at

smoothing consumption and curbing poverty (Dupas and Robinson, 2008).

The prevailing low rate of financial inclusion has attracted the attention of

scholars to investigate its driving factors (Asli and Klapper, 2012; Kumar, 2006;

Collins et al., 2009; Susan and Zarazua, 2011). Among the commonly cited limiting

factors is the relative concentration of formal financial institutions in urban centers

with limited penetration among rural communities. This urban concentration poses

high monetary and opportunity costs involved in accessing and using financial

services, especially by the rural poor in remote locations. In their analysis of financial

access and exclusion in Kenya and Uganda, Susan and Zarazua(2011) re-defined

financial inclusion to include semi-formal and informal financial services like

Rotating Saving and Credit Associations (ROSCA) and Savings and Credit

Cooperative Organizations (SACCO). They found that exclusion is associated with

agro-ecological and socio-cultural characteristics of the region, rather than the mere

urban-rural status.

Mobile banking, a recent innovation in the financial sector, is expected to

bridge the financial service access gap, thus allowing for socio-economic

improvements especially among the financially excluded rural communities in many

1Financial inclusion, as synonymous to financial access, will be used interchangeably with financial

exclusion. Access and inclusion will be used to refer to a situation where an individual has access to

the services of a formal financial institution like a commercial bank, Micro-finance institutions and

insurance companies. Financial exclusion is used in this paper to refer to the involuntary lack of

access to formal financial services.

3

developing countries. Mobile banking allows users to make, deposits and transfers of

funds as well as purchase of some limited range of goods and services using their

mobile phone. This provides a relatively cheap and convenient means through which

family members and friends exchange financial assistance in the form of remittances

especially in remote areas with limited or no access to formal financial institutions

like banks. Empirical studies have illustrated the developmental role of mobile

banking. One such popular channel of this impact is the change in the pattern of

remittances (Mbiti and Weil, 2011). The benefit of mobile money extends beyond the

individual and household levels to businesses and organizations. Aker et al. (2011)

demonstrated that the welfare program that distributed financial assistance for people

to cope with the adverse effects of a severe drought in 2008 was implemented

cheaply through mobile money, relative to conventional transfer mechanisms. This,

they argue, owes to the relative inexpensiveness and convenience of mobile banking.

Jack and Suri (2011) provided evidence that access to mobile money services

facilitates risk sharing by significantly reducing the transaction costs of remittances

among family member and friends in Kenya. They found that households which

subscribe to M-Pesa - Kenya’s most popular mobile money service - were able to

cushion themselves against consumption volatilities when struck by income shocks,

by receiving remittances from a wide pool of members in their social networks.

Despite the relative importance of mobile banking in the lives of the rural poor, less

is known about its impact on their welfare. Specifically, there is scanty empirical evidence

on how financial access affects the lives of the rural poor in developing countries. To the

best of our knowledge, there is no empirical study that analyses the socio-economic impact

of mobile banking in the Ugandan context, most of the recent works are based on the

Kenyan experience (Mbiti and Weil, 2011; Jack and Suri, 2011). Besides, the analysis

samples of these studies are inclusive of the urban mobile money users with less focus on

the rural communities which tend to be more financially excluded. Moreover, recent studies

on mobile money in Uganda are centered on analyzing adoption and use patterns (Susan

and Zarazua, 2011; Ndiwalana, 2010) while other studies rely anecdotal evidence.

4

Following the rapid adoption of mobile money services in Uganda, there is need to assess

whether there is any direct welfare improvement that accrues to its users.

This paper seeks to fill the literature gap by investigating the impact of mobile

money access on the welfare of rural households in Uganda. This study is unique in a way

that it targets particularly households in rural locations which often tend to have less access

to formal banking services coupled with relatively high poverty rates. We use a two-year

panel of 907 households from 94 Local Council 1s in Uganda2, collected in 2009 and 2012.

In less than four years since its inception in March 2009, the number of active mobile

money subscribers has expanded to over nine million users.3 Between December 2011 and

December 2012, the number of mobile money users increased from 2.9 million users to 9

million users. This is expected to facilitate inter-household transfer of funds especially and

thereby increase household welfare. The number of LC1s with at least one mobile money

booth increased from 26 to 90 out of 94 LC1s in our sample between the two survey rounds.

At the same time, household adoption of mobile money services expanded from less than

one percent to 38 percent.

From our preferred specification, results indicate that using mobile money is

associated with a 69 percent increase in household per capita consumption. This is made

possible through the facilitation of remittances among family members and friends. In

particular, we find that households with at least one mobile money subscriber are 20

percentage points more likely to receive remittances from their members in towns and that

the total annual value of remittances received is 33 percent higher compared with their non-

user counterparts.

The rest of the paper is organized as follows. In section II, we provide background

information about mobile money in Uganda before proceeding to the conceptual framework

of the expected welfare impact of mobile money and the mechanisms underlying this

impact in Section III. Section IV discusses the data and summary statistics, followed by

2 An LC1 is the second smallest unit of administration in Uganda.

3Bank of Uganda estimate as of December 2012.

5

empirical strategy in section V. Empirical results are discussed in section VI while section

VII concludes.

II. Background on mobile money in Uganda.

In March 2009, Mobile Telephone Network (MTN) -Uganda established MTN

Mobile Money, the first of its kind in the country, following the massive success of

Safaricom’s M-PESA in Kenya. Airtel Uganda, formerly known as Zain, joined the

service when it rolled out its Airtel Money in June the same year. This new financial

innovation proved to be an efficient way for telecom companies to increase their

market shares by widening the range of services available to their clients. This

attracted Uganda Telecom’s M-Sente in March 2010, followed by Warid Pesa from

Warid Telecom in December 2011 and Orange Money from Orange Telecom in the

first half of 2012 (Uganda Communications Commission-UCC 2012).

Since mobile money was established in Uganda, the number of subscribers

has been steadily increasing. By the end of 2012, Uganda had over 9 million mobile

money users all over the country. This represents a three-fold expansion from 3

million users in 2011. The number of mobile money transactions increased from 180

million to 242 million between 2011 and 2012 while the total value exchanged

through the platform increased from $1.5 billion to $4.5 billion in the same period

(BoU, 2012).MTN Mobile Money alone has over 15,000 agents as compared with

455commercial bank branches with 660 Automated Teller Machines (ATMs). This

rapid expansion partly owes to the high rates of both the roll-out of mobile phone

network and adoption of mobile phones. In our sample, the proportion of households

owning a mobile phone increased from 52 percent to 73 percent between the two

survey rounds while all LC1s were covered by mobile phone network in both surveys.

Mobile money allows users to deposit money as e-float on a SIM card-based account,

called an m-wallet, which can be converted into cash at any mobile money agent

location all over the country. In the initial stages of its establishment, the range of

6

services offered was largely limited to person-to-person transfer but with the growing

interest from stake-holders, coupled with competition among the mobile network

operators (MNOs), this platform has expanded the range of services to include more

complex uses like payment of utility bills, school fees, airtime purchase and direct

purchase of goods and services.

Recent developments in the mobile banking arena have made it possible for

users to access their bank accounts using their mobile phones without having to

physically visit their bank branches, thanks to the partnership between MNOs and

banks.4 This is expected to raise financial inclusion especially at the lower end of the

social spectrum while reducing the cost of access and use of basic financial services.

With the rapid urbanization in Uganda over the past years, the number of people

migrating to towns has been steadily increasing. Those who migrate to cities often

render financial support to their rural households in the form of remittances. The

efficiency of this remittance system heavily relies on the quality of physical

infrastructure as most of these transactions involve physical transfer of cash by the

receiver, sender, and agents like bus and taxi drivers among others informal channels.

Besides, the massive geographical dispersion between senders and receivers implies

high transaction costs in terms of transport fares and travel time involved in sending

and receiving money among household members especially across geographically

distant and remote locations.

III. Conceptual Framework.

This section provides a conceptualized idea of inter-household exchange of resources

as an informal insurance scheme with an ultimate goal of maximizing overall

household welfare. We assume household members are scattered in rural and urban

locations yet they maximize joint welfare. This isolation of members is partially due

to migration of some members into cities in search for better opportunities like jobs

with higher pay and those who work in towns regularly remit money to their

4Major partnerships exist between MTN Mobile Money and Stanbic Bank, M-Sente and Standard

Chartered Bank and WaridPesa and DFCU Bank.

7

members in the country side. The effectiveness of this sharing mechanism relies on

the cost of transferring money between the two ends, which in turn depends on the

variety of transfer mechanisms and the physical distance between the sender and the

receiver.

In our theoretical model, we assume two periods – one prior to and none after

the introduction of mobile banking. Given that the receiving members are located in

remote areas, access to formal financial institutions is highly limited. This implies

that in period one, cash is transferred physically either by the sender or receiver or

through agents like bus and taxi drivers because formal financial institutions are only

available at the sending end but not at the receiving end of the social sharing network.

This poses a high cost of exchanging financial assistance among members of the

social network (household) both in terms of transport fare and opportunity cost of

travel time between the two locations. Limiting the effectiveness of this sharing

scheme can be reflected in reduced per capita household consumption, which reduces

the overall welfare of the household members. Under these conditions, remittance

channels that increase convenience and reduce the cost of remittances may facilitate

funds flow within the network, raising per capita consumption and overall welfare.

In period two, mobile banking is introduced which makes it possible for

members at the receiving end to access funds remitted by their working members.

This financial innovation is cheap and convenient both for the provider and the user.

On the side of the service provider, the benefit is the ability to reach out rural

households without massive investment in financial infrastructure like banks while

users benefit from the reduced time and monetary costs of accessing financial

services. In this period, members who work in towns have a convenient platform to

send money to their rural households without physically travelling back home or

sending physical cash, which in turn increases the safety of remitted funds. This

period is associated with a higher probability that a rural household receives

remittances from its members working in town and a corresponding increase in both

the number of remittances and total value received.

8

If mobile money reduces the economic and opportunity costs of transferring

money across geographically distant areas, then households which have access to this

service should have higher welfare, made possible through the availability of a cheap

and convenient remittance channel. The number of people using mobile money in

Uganda increased from 3 million to 9 million between 2011 and 2012 while in our

sample, the proportion of users increased from less than one percent to 38 percent

between 2009 and 2012. This rapid expansion is expected to increase financial

inclusion and raise welfare among rural communities. We therefore believe that

household per capita consumption increases with mobile money adoption. We also

hypothesize that access to mobile money services increases the probability that a

household receives any remittance, the frequency of remittances as well as the total

value of remittances received from their members who mostly work in urban centers.

IV. Data and Summary Statistics

We mainly use data from household and community surveys collected in Uganda in

2009 and 2012 as a part of the Research on Poverty, Environment and Agricultural

Technology (RePEAT) project. This is part of the four survey rounds administered

jointly by Makerere University, the Foundation for Studies on International

Development (FASID) and the National Graduate Institute for Policy Studies

(GRIPS) in 2003, 2005, 2009 and 2012. In the baseline survey of 2003, 94 LC1s were

sampled and 10 households were randomly selected from each of the LC1s, making a

total of 940 households. The follow-up surveys of 2005, 2009 and 2012 successfully

captured 856, 816 and 866 of the original households, respectively. The high attrition

rate in the third round was partially offset by the inclusion of neighboring households

to replace those that could not be traced

The major household-level information that was captured in the surveys

included demography, income and consumption expenditure, wealth indicators, use

of telecommunication and financial services like mobile phones and mobile banking

and farming practices. Community characteristics like distance and travel time to the

9

market and district towns, availability of mobile phone network and quality of roads

were captured in the community-level surveys.

Analysis in this paper is based on a balanced panel of 838 households

generated from the third and fourth rounds in 2009 and 2012. Table 1 shows the

summary statistics of this sample. The proportion of households which report having

at least one member using mobile money services increased from less than 1 percent

in 2009 to 38 percent in 2012. The rapid adoption of mobile money services owes to

the availability of mobile phone network and the expansion in mobile phone adoption.

While all LC1s were covered by mobile phone network in both surveys, mobile

phone adoption at the household level increased from 52 percent to 73 percent.

[Insert table 1 here]

Although bank account information was not captured in 2009, we do not expect a

substantial change between the two rounds. It is not surprising that only 22 percent of

households reported having at least one bank account in 2012 because our sample

households are predominantly from rural-based. This throws light on the relative

exclusion of majority of rural households and individuals from the formal financial

sector services.

The proportion of households that reported having received remittance in the

past 12 months prior to the survey month increased by 20 percent from 50 percent in

2009 to 70 percent in 2012 and on average, the annual frequency of remittances

increased from 2 to 4 times. There was a two-fold increase in the total value of

remittances received from UGX 274,682 (USD 135)5 in 2009 to UGX 582,546 (USD

227) in 2012. Among households that received remittances, the total value received

represents approximately 10% of total annual household consumption expenditure.

5 According to the annual Bank of Uganda Report 2012, the Uganda shilling was equivalent to USD 2028 and

2557 in financial years 2008/2009 and 2011/2012, respectively.

10

During the same period, monthly per capita consumption rose from UGX

58,234(USD 29) to UGX 96,184 (USD 47). Asset value was, on average UGX

826,284 (USD 323) and the land holding per household was five acres (two hectares).

There was significant change in the household size6 at 8 persons on average, while

the proportion of female-headed households increased by 3 percent, from 13 percent

in 2009 to 16 percent in 2012. An average household head was 51 years old with six

years of schooling while the level of education attainment by the most educated

member of the household was averagely 9 years.

We stratify our sample by mobile money adoption status in Table 2 to give a

rough understanding of the difference in the characteristics between the users and

non-users. Seventy eight percent of mobile money users reported that they had ever

received at least one remittance in the past twelve months prior to the survey month,

with an annual remittance frequency of 5 times and total value of UGX 859,394

(USD 336). This represents a significant difference from 56 percent of non-user

households that received remittances slightly less than 3 times a year amounting to a

total value of UGX 324,182 (USD 126) in the same period. Monthly per capita

consumption was 43 percent higher relative to non-users. In general, mobile money

user households are relatively wealthier, are headed by more educated members and

have a larger number of members. Concerning gender and age of the household head,

there is no significant difference between users and non-users.

[Insert table 2 here]

V. Empirical Strategy

In this section, we estimate three major equations; (i) the determinants of mobile

money adoption at the household level, (ii) the effect of mobile money adoption on

household per capita consumption and (iii) the impact of mobile money use on

6 A household member is one who had lived in the household for at least one month in the 12 months prior

to the respective survey month.

11

measures of household remittances; probability of receiving remittances, frequency

and total value of remittances received.

A. Baseline Specification

1) Determinants of mobile money adoption.

Since the outcome variable we want to estimate in this section is a binary variable, we

estimate the probability of adopting mobile money services as a latent variable model. The

decision to adopt mobile money services cannot be observed but depends on observed

characteristics of the household and village in the form

��������∗ = λ + βXijt + ɳjt + ɛijt (1)

where Mmoney*ijt is a latent variable that captures the decision to adopt mobile money

services by household i in village j at time period t; ɳjt captures location and time effects

and Xijt is a vector of household characteristics which include household size, log of

asset and land endowments, age, gender and education level of the household head and a

dummy for household mobile phone possession. Our variable of interest, Mmoney, is

then observed only if the latent variable ��������∗ is positive. Specifically,

Mmoney= � 1��������∗ > 0

0��ℎ������ (2)

We also estimate mobile money adoption using a more robust linear probability fixed

effects estimation to rule out the effect of unobservable time-invariant household and

village characteristics that might simultaneously influence welfare and the decision of

the household to adopt mobile money services. As we shall show in the results section,

the change of estimation method does not qualitatively change our results.

2) Mobile money and household per capita consumption

12

We first examine the effect of mobile money adoption on household welfare using a

simple difference-in-differences strategy that compares the monthly per capita

consumption of mobile money users against that of non-users.

cijt = γ + αi + µMmoneyijt + ψXijt + ɳjt + εijt (3)

where cijt is the monthly per capita consumption of household i in village j in period t, αi

is a household fixed effect, Mmoneyijt is a dummy variable equal to one if there is at least

one mobile money subscriber in the household. The parameter µ thus represents the

coefficient of our interest which is expected to be positive. We use household per capita

consumption as a proxy for household welfare. As an alternative, we could use total

household income as it is also directly linked to the ability of a household to improve the

wellbeing of its members. However, this measure is more vulnerable to short-term

economic effects compared to the consumption measure (Gilligan and Hoddinott, 2009).

3. Mechanisms: Mobile Money and Remittances

To assess whether remittance patterns differ across users and non-users of mobile money,

we estimate the following equation, which is a slight modification of equation (3).

rijt = γ + αi +π Mmoneyijt + ψXijt + σjt + εijt (4)

where rijt is a measure of remittances received by household i in village j in period t. This

measure takes three variants; the probability that a household receives a remittance, the

number of remittances received in the past 12 months of the respective survey round and

the total value received within the same period. In order to more concretely account for

family dynamics in the remittances structure, we include a dummy variable equal to one

if the household reported having at least one member who moved out to search for a job

outside the home village, hereafter used interchangeably as job-seeking behavior and

having a migrant worker. In equation (4) we include a full set of controls as in (3) above.

B. Falsification Test

13

In order to confirm that the observed difference in consumption and remittances between

users and non-users of mobile money is genuinely due to this financial platform, we

replicate the estimation strategy as described above, using RePEAT data for the period

prior to mobile money. We thus estimate equations (3) and (4) using 2003 and 2005 data.

This constitutes the first and second rounds of the RePEAT series, as described in the

Data section of this paper. Using this data, we examine whether there existed differences

in consumption and remittance patterns between households that later adopted mobile

money against non-adopters. Since mobile money was not available in this period, we

use a placebo binary treatment variable equal to one for households that adopted mobile

money in/after 2009. We also examine having a migrant worker in a household had an

influence over remittance patterns. This strategy enables us to assess whether the

differences in outcome variables (consumption and remittance measures) between users

and non-users are indeed a result of mobile money adoption status. We expect no

significant difference between households that later adopted mobile money services and

those that did not. If this is true, then the emergence of a significant relationship between

mobile money and the outcome variables could be attributed to mobile money.

C. Instrumental Variable and Tobit Regressions

So far, we have assumed that mobile money adoption by the household is conditionally

mean-independent, given the other control variables included in the regressions. This

implies that the estimated coefficients are only valid if mobile money adoption is not

correlated with the error term conditional on the other controls. Although we are able to

rule out the effect of unobserved time-invariant household heterogeneity using fixed effects

estimation, the decision to adopt mobile money services may be highly correlated with

time-variant un-observables that also affect household consumption expenditure. Also,

being a remittance recipient in the past might induce the household to adopt mobile money

as a cheaper and convenient platform to receive remittances from their members in towns.

This endogeneity resulting from simultaneous effects might confound our OLS and fixed

effects estimates. To address the issue, we resort to instrumental variable estimation of

14

consumption using log of the distance to the nearest mobile money agent as an instrument

for mobile money adoption at the household level. We employ a Tobit model in

combination with a control function method to deal with two critical challenges associated

with our remittance variables. The first challenge concerns the corner solution nature of the

remittance measures, owing to the fact that the number and total value of remittances

received are only available for households which received positive remittances. This

implies that these variables have a skewed distribution given the many zeroes for non-

recipients. The control function approach deals with the second challenge - potential

endogeneity resulting from the correlation between remittance variables and time-variant

unobserved household characteristics (Vella, 1993). In both variants of our Tobit models,

we include time averages of household characteristics to rule out the effect of time-

invariant household characteristics that could confound our results (Mason, 2013). Like in

the standard IV method described above, we include the log of distance to the nearest

mobile money agent in estimating the number and total value of remittances received.

D. Reduced form analysis

The effectiveness of mobile money services heavily relies on the availability and ease of

access to mobile money agents as these facilitate cash-in and cash-out transactions. In this

section, we examine whether access to a mobile money agent influences household welfare,

supposedly through mobile money-based remittances. In the spirit of Jack and Suri (2011),

we use the log of distance to the nearest mobile money booth as a measure of access to

mobile money services and use the specification below to assess this relation.7

cijt = γ + αi + π logdistjt + ψXijt + σjt + εijt (5)

where logdistjt is the log of distance in kilometers from village j to the nearest mobile

money booth. Xijt is a vector of household controls as discussed above. We expect π to

have a negative sign because the further the mobile money agent, the harder it may be

for a household to access mobile banking services and this might translate into

7 Distance to the nearest mobile money location is captured at the community level.

15

reduced ability of a household to receive financial assistance in form of remittances

from its members. This would, in turn, reduce the power of a household to smooth

consumption as described in earlier sections.

VI. Results

A. Basic Results

1) Determinants of household mobile money adoption.

Table 3 presents the determinants of household mobile money adoption. The probit

results in Column 1 reveal that households with mobile phones are nine percentage

points more likely to use mobile money services. This is not surprising because

mobile money services are offered through a cell phone handset. Education of the

household head has a positive and significant impact on the decision to adopt mobile

money services; an additional year of education of the household head leads to one

percentage point increase in the probability of adopting mobile banking. This could

partly capture the literacy effect of educated household heads who could be more able

to operate mobile handsets. Alternatively, it could be true that educated household

heads are more able to send their children to school who, upon graduation, find jobs

in towns and extend financial assistance in form of remittances through mobile

money platforms. This claim is partly supported by the significantly positive impact

of the job-seeking dummy on mobile money use by the household.

As expected, the coefficient on the 2012 year dummy is positive and significant,

consistent with the fact that mobile money adoption increased tremendously between

2009 and 2012 both in our sample and at the national level. These results remain

qualitatively unchanged with the fixed effects estimation in Column 2. The

significantly negative coefficient on the distance to the nearest mobile money agent

implies that households choose to subscribe to mobile money services if the distance

from the nearest booth is relatively shorter. This further supports the notion that the

relative urban concentration of banks is partially responsible for the slow adoption of

16

formal financial services. It should be noted that mobile money booths and agents are

instrumental in facilitating mobile money transactions in a way that they act as cash-

in and cash-out agents.

[Insert Table 3 here]

2) Mobile money and household per capita consumption

Table 4A reports the results from the estimation of (3) as OLS and fixed effects

models with a full set of household and community characteristics. In column 1 we

include location-by-time controls among the covariates in our OLS model. The

results suggest a 13 percent increase in household per capita consumption given the

adoption of mobile money services. To address the possibility of bias in our OLS

results that could potentially result from unobserved and time-invariant household

heterogeneity, we estimate a fixed effects model with and without location-by-time

effects in columns 2 and 3, respectively. Across all specifications, the estimates

remain qualitatively similar, suggesting a significantly higher level of per capita

consumption for mobile money users. The location-by-time effects in Column 3

capture district-level trends that might be correlated with both mobile money

adoption and per capita consumption.

[Insert Table 4A here]

We further disaggregate our consumption expenditure measure into three categories –

expenditure on food items, non-food household basics and social contributions.8

Table 4B gives a report of these three measures using both OLS and fixed effects

estimations. Column 1 shows that mobile money adoption has a positive impact on

per capita food expenditure, although the relationship disappears after controlling for

unobserved time-invariant household characteristics in Column 2. The average

8Expenditure on household basics includes expenditure on school, medical, transport, clothing,

cooking and lighting materials. Social contributions cover expenses on ROSCAs, mutual support organizations – both funeral and non-funeral, churches and mosques, other local organizations and credit repayments.

17

impact for basic expenditure ranges between 15% and 20% for OLS and fixed effects

models, respectively (Columns 3 and 4). Columns 5 and 6 reveal that a household

that uses mobile money services experiences between 47 and 56 percent higher value

of social contributions. These results should, however, be interpreted carefully, as

they are likely to be capturing reverse causality effects.9 Nonetheless, they suggest

that social contributions and basic expenditures respond more strongly to mobile

money adoption as compared to food expenditure. This result is not rather surprising,

owing to the rural nature of households in our sample which implies that a large

fraction of consumed food comes from own farms. Chetty and Looney (2006) argue

that when consumption is close to subsistence level, any shocks to income might not

necessarily translate into reduced household consumption because its level is already

too low such that it cannot be reduced any further.

3) Mechanisms

Based upon our theoretical prediction, the impact of mobile money on household

welfare is achieved through the facilitation of remittances. We explore into this claim

by examining whether households that have access to mobile money services have

differential access to remittances. These results are reported in table 5A. Being a

mobile money user is associated with a significantly higher probability of receiving

remittances and the remittances received are larger in number and total value

compared with non-users. In estimating the probability of a household receiving

remittance, we estimate equation (4) as a Probit model, since the dependent variable

is binary. The results in Column1 show that mobile money adoption increases the

probability of receiving remittances by seven percentage points. These results remain

qualitatively unchanged when using OLS regression in Column 2. In columns 3

through 6, we present the results from the other two measures of remittances –

number of remittances and total value received in the past 12 months. From Columns

9Household that make numerous social contributions may be convinced by members of their social

networks to join mobile money services for easier transmission of contributions.

18

3 and 4, mobile money users receive approximately one more remittance at a given

time, compared to non-users. The OLS estimates of total value of remittances in

Column 5 reveal that adopting mobile money services increases the total value of

remittance received by 36%. This translates into approximately 116,706 Uganda

Shillings (USD 61), as evaluated at the mean value of non-users. The fixed effects

estimation of remittance value in Column 6 yields similar results even after

controlling for unobserved time-invariant heterogeneity between users and non-users.

In all specifications, we include controls for household characteristics (mobile phone

possession, household size, asset value, land size, as well as age, education and

gender of household head). The inclusion of location-by-time effects in our

regressions captures local macro trends that may have differential influence on

household access to remittances.

[Insert Table 5A here]

B. The influence of migration (job-seeking behavior)

We now account for the source of remittances and examine the possibility of

differential remittance structure between households that send their members to find

jobs in towns and those that do not. These results are reported in Table 5B. Column 1

reveals that, conditional on mobile money status and other covariates, households that

send their members to find town jobs are 11 percentage points more likely to receive

remittances. Columns 2 and 3 report results for the number and total value of

remittances received, respectively. Having a member working outside the village

increases the number and total value of remittances by 1.4 times and 42%,

respectively. We believe that the introduction of mobile money reduced the monetary

and opportunity costs that hitherto hindered these workers from transferring money to

villages. Our presumption is that, even when members were working in towns prior

to the introduction of mobile money, the idiosyncratic lack of a cheap and convenient

money transfer mechanism rendered it hard for the members to remit financial

assistance back to their rural households. To check this claim, we perform similar

19

analysis on a sub-sample covering the period before mobile money inception in 2009

– survey rounds of 2003 and 2005. The results in appendix Table A1 suggest no

significant relationship between working outside the village and all measures of

remittances and consumption. The fact that this relationship emerged after mobile

money establishment provides partial evidence in support of impact of mobile money

on remittances.

[Insert Table 5B here]

C. Results from Reduced Form Analysis

Table 6 reports the results from our reduced form analysis using log of distance to the

nearest mobile money booth as a measure of access to mobile money services at the

community level. The dependent variable in column 1 is the log of monthly

household per capita consumption. As earlier predicted, being located away from the

mobile money booth is associated with a significant reduction in household per capita

consumption. The probability, number and total value of remittances received, as

measures of remittances, are reported in columns 2, 3 and 4, respectively. Results are

consistent with those reported in our previous estimations. Households in located one

kilometer away from the mobile money booth have two percentage point lower

probability of receiving remittances (Column 2). Similarly, the frequency and total

value of remittances received reduces significantly with an increase in the distance to

the mobile money agent. Note that the treatment variable in this case is a community-

level variable and the inclusion of district and time dummies implies that our estimate

is a conservative estimate of the true effect of mobile money access as these controls

absorb much of the variations in mobile money access. Most importantly, controlling

for district and time effects rules out the potentially confounding effect of local

access to services that tend to be concentrated in district towns.

[Insert Table 6 here]

20

D. IV and Tobit Results

Results reported so far rely on the assumption that mobile money is not correlated

with the error term conditional on the other controls included in the regressions.

However, where this assumption does not hold, both OLS and fixed effects estimates

may be biased. As earlier noted, mobile money is potentially endogenous given

reverse causality concerns – households may adopt mobile money when they expect

to receive remittances. In this section, we account for this endogeneity using standard

fixed effects IV method for consumption and Tobit models with a control function

approach for remittances. Apart from capturing potential endogeneity, the latter

technique takes into account the corner solution problem resulting from the censored

nature of our remittance variables, that is, households that never received remittances

have no observations for the number and total value of remittances. In the control

function version of our Tobit model, we include residuals from the first stage

estimation of the determinants of mobile money in the main model. In both methods,

we use log of distance to the nearest mobile money agent as an excluded instrument

for the potentially endogenous mobile money variable.

The results of these estimation methods are reported in Table 7. Column 1 reports

results of the consumption measure using standard fixed effects IV method. Columns

2 through 5 report the Tobit estimates of the number and total value of remittances

received. In columns 3 and 5, we combine Tobit with control function methods to

control for corner solution and endogeneity problems. Estimates in Column 1 reveal

that per capita consumption increases by 69 percent upon adoption of mobile money.

Columns 2 and 3 show that mobile money adoption approximately doubles the total

value of remittances received while Columns 4 and 5 show that users receive more

than one additional remittance relative to non-users. The number of remittances is

positively associated with mobile money usage, although the coefficient is not

statistically distinguishable from zero at conventional levels of significance. In line

21

with Mason 2013, the significance of the residual in Columns 3 and 5 not only

implies potential endogeneity of the treatment variable but also deals with the

problem. We therefore focus on the results in Columns 3 and 5 for our measures of

remittances.

[Insert Table 7 here]

E. Alternative Explanations

One might argue that the changes in remittance patterns could have resulted from mobile

phone possession which could have enabled rural households to contact their members in

towns in times of hardship. If this were the case, then mobile phone possession would be

expected to have a positive and significant effect on the flow of remittances among the

household members even in the absence of mobile money. In order to explore into this

possibility and thus disentangle any impact of mobile phone from that of mobile money, we

examined the relationship between mobile phone possession and household per capita

consumption and remittances prior to the introduction of mobile money. We therefore run

regressions of the outcome variables on a dummy variable of mobile phone possession

using 2003 and 2005 data, including a full set of controls as in previous sections. As

reflected in Table 8, there is no significant relationship between mobile phone possession

on one hand and consumption (Column 1) and remittances on the other (Columns 2 through

5). At best, the remittance impact of mobile phone possession is positive and statistically

indistinguishable from zero. This partially rules out the possibility that the observed

consumption and remittance changes resulted majorly from mobile phone possession.

VI. Conclusion

Lack of access to financial services is a typical challenge to rural livelihood in many

developing countries. Apart from the direct hindrance on the ability to borrow and

save, the associated high costs of remitting funds to financially inaccessible areas

impose a limit on the effectiveness of informal sharing mechanisms among friends

22

and relatives. Mobile money - a new financial service that allows direct transaction

via a mobile phone –serves to bridge this gap given its relatively lower cost and

convenience. In Uganda, mobile money adoption has expanded tremendously over

the past three years since its inception in 2009. In this paper, we examine the welfare

impact associated with this service by estimating its impact on monthly household per

capita consumption. Specifically, we provide evidence that households using this

financial innovation experience a significant increase in per capita consumption. The

result is robust to sensitivity checks, mainly the change in empirical specification.

Disaggregating consumption into food, basic and social expenditures, we find

stronger impacts of mobile money for the social expenditure measure, partially

suggesting investment in informal social and insurance networks and saving

mechanisms.

There are a number of potential pathways through which this result might be

realized as cited in the literature including the facilitation of savings (Jack and Suri,

2011) and self-insurance through remittances. We provide evidence that the estimated

impact is achieved through the facilitation of remittances; households with access to

mobile money services are more likely to receive remittances, receive remittances

more frequently and receive higher value of remittances relative to non-users.

Although we do not explicitly demonstrate due to data limitations, we are convinced,

based on anecdotal evidence that the average cost of remitting funds across

households reduced greatly with the event of mobile money technology. We further

venture into the role of family dynamics by comparing remittance patterns across

households with and without members working outside the village. We provide a

falsification test that the relationship between this migration measure and remittances

did not exist prior to mobile money, suggesting that its emergency after 2009 partially

reflects reduction in transaction costs that made it possible for workers to remit funds

to their rural households.

23

The results presented in this paper suggest significant welfare benefits of

access to financial services which might go afield in reducing rural poverty through

reduction in vulnerability by the rural poor. Dercon (2006) suggests stronger welfare

benefits of informal insurance mechanisms if random reductions in consumption

affect poverty dynamics through persistent income reduction in incomes. One

concern however is that, although we plausibly assume reduction in remittance cost

as the major pathway of the welfare and remittance impact of mobile money, we do

not test this premise within the limitation of the data. This and the analysis of risk-

sharing behavior will form the foundation for further research.

References

Adger, W. N., Winkels, A. (2007). Vulnerability, poverty and sustaining well-being.

Handbook of sustainable development, 189.

Aker, J. C. (2010). Information from markets near and far: Mobile phones and agricultural

markets in Niger. American Economic Journal: Applied Economics, 46-59.

Aker, J. C., Boumnijel, R., McClelland, A., Tierney, N. (2011). Zap It to Me: The Short-

Term Impacts of a Mobile Cash Transfer Program. Center for Global Development

Working Paper, (268).

Becker, S., Hvide, H. (2013). Do entrepreneurs matter?. Unpublished Manuscript.

24

Buys, P., Dasgupta, S., Thomas, T. S., Wheeler, D. (2009). Determinants of a digital divide

in Sub-Saharan Africa: A spatial econometric analysis of cell phone coverage. World

Development, 37(9), 1494-1505.

Chavan, A. L., Arora, S., Kumar, A., Koppula, P. (2009). How mobile money can drive

financial inclusion for women at the Bottom of the Pyramid (BOP) in Indian urban

centers. In Internationalization, Design and Global Development (pp. 475-484).

Springer Berlin Heidelberg.

Chetty, R., Looney, A. (2006). Consumption smoothing and the welfare consequences of

social insurance in developing economies. Journal of Public Economics, 90(12), 2351-

2356.

Claessens, S. (2006). Access to financial services: A review of the issues and public policy

objectives. The World Bank Research Observer, 21(2), 207-240.

Combes, J. L., Ebeke, C. H., Etoundi, S. M. N., & Yogo, T. U. (2014). Are Remittances

and Foreign Aid a Hedge Against Food Price Shocks in Developing Countries?. World

Development, 54, 81-98.

Combes, J. L., Ebeke, C. H., Etoundi, S. M. N., & Yogo, T. U. (2014). Are Remittances

and Foreign Aid a Hedge Against Food Price Shocks in Developing Countries?. World

Development, 54, 81-98.

Deaton, A., Zaidi, S. (2002). Guidelines for constructing consumption aggregates for

welfare analysis (No. 135). World Bank Publications.

Demirguc-Kunt, A., Klapper, L. (2012). Measuring Financial Inclusion: The Global

Findex.World Bank Policy Research Wp, 6025.

25

Dercon, S. (2006). Vulnerability: a micro perspective. Securing Development in an

Unstable World, 117-146.

Dercon, S., Gilligan, D. O., Hoddinott, J., Woldehanna, T. (2009). The impact of

agricultural extension and roads on poverty and consumption growth in fifteen

Ethiopian villages. American Journal of Agricultural Economics, 91(4), 1007-1021.

Donovan, K. (2012). Mobile money for financial inclusion. Information and

communication for development, 61-74.

Elul, R. (1995). Welfare effects of financial innovation in incomplete markets economies

with several consumption goods. Journal of Economic Theory, 65(1), 43-78.

Gilligan, D. O., Hoddinott, J., Taffesse, A. S. (2009). The impact of Ethiopia's Productive

Safety Net Programme and its linkages. The journal of development studies, 45(10),

1684-1706.

Gu, J. C., Lee, S. C., Suh, Y. H. (2009). Determinants of behavioral intention to mobile

banking. Expert Systems with Applications, 36(9), 11605-11616.

Heckman, J. J. (1979). Sample selection as a specification error. Econometrica: Journal of

the econometric society, 153-161.

Jack, W., Suri, T. (2011). Mobile money: the economics of M-PESA (No. w16721).

National Bureau of Economic Research.

Jack, W., Suri, T. (2011). Risk sharing and transaction costs: Evidence from Kenya’s

mobile money revolution’. Working paper.

26

Jack, W., Ray, A., Suri, T. (2013). Transaction Networks: Evidence from Mobile Money in

Kenya. American Economic Review, 103(3), 356-61.

Klein, M., Mayer, C. (2011). Mobile banking and financial inclusion: The regulatory

lessons (No.166). Working paper series//Frankfurt School of Finance & Management.

Marshall, J. N. (2004). Financial institutions in disadvantaged areas: a comparative analysis

of policies encouraging financial inclusion in Britain and the United

States. Environment and Planning A, 36(2), 241-262.

Mbiti, I., Weil, D. N. (2011). Mobile banking: The impact of M-Pesa in Kenya (No.

w17129). National Bureau of Economic Research.

Mohan, R. (2006). Economic Growth, Financial Deepening, and Financial Inclusion.

Muto, M., Yamano, T. (2009). The impact of mobile phone coverage expansion on market

participation: Panel data evidence from Uganda. World Development, 37(12), 1887-

1896.

Ndiwalana, A., Morawczynski, O., Popov, O. (2010). Mobile money use in Uganda: A

preliminary study. M4D 2010, 121.

Pousttchi, K., Schurig, M. (2004). Assessment of today's mobile banking applications from

the view of customer requirements. In System Sciences, 2004. Proceedings of the 37th

Annual Hawaii International Conference on (pp. 10-pp). IEEE.

27

Table 1: Summary Statistics

Round One (2009) Round Two (2012) Difference (2012-2009) Variable Mean SD Mean SD Mean P-Value ICT use 1 if mobile money used

0.0077

0.0877

0.3848

0.4868

0.3770

0.000

1 if mobile phone owned 0.5216 0.4998 0.7344 0.4419 0.2128 0.000

1 if bank account owned - - 0.2285 0.4201 - -

Wealth Total value of assets (UGX)

604,436

1,728,076

1,048,871

1,904,279

444,434

0.000

Land holding size (acre) 5.3884 7.0656 6.2286 7.8175 0.8402 0.016

Remittances

1 if received remittance 0.5022 0.5003 0.7084 0.4548 0.2061 0.000

No. of remittances 2.3814 4.6654 4.0144 6.2181 1.6330 0.000

Total remittance (UGX) 274,682 723,289 582,546 1,208,951 307,864 0.000

Remittance to Consumption Ratio

0.0948 0.2555 0.1013 0 .2638 0.0065 0.618

Welfare Per capita consumption (UGX) 58,234 63,328 96,184 115,614 37,950 0.000

Total wage 18,207 68,899 33,840 126,636 15,633 0.001

HH characteristics Head age 50.2920 14.2717 52.1454 14.1744 1.8534 0.006

1 if head is female 0.1291 0.3355 0.1632 0.3697 0.0340 0.040

Head education 6.1044 3.9445 5.7975 3.8438 0.3068 0.097

Household size 7.5408 3.3399 7.9901 3.7703 0.2733 0.077

Village characteristics

Distance to district town (km) 14.7326 11.6011 9.9685 8.4677 4.7640 0.000

1 if LC1 has MM agent 0.2393 0.4269 0.9912 0.0936 0.7519 0.000

Source: RePEAT data 2009 and 2012. The observations are from 907 balanced panel households.

28

Table 2: Summary Statistics by Mobile Money Status.

Non-user User Difference (user-nonuser)

VARIABLES Mean SD Mean SD Mean P-Value

ICT Use

1 if mobile phone owned 0.5532 0.4973 0.9375 0.2424 0.3843 0.000

1 if bank account owned 0.1329 0.3397 0.3811 0.4864 0.2482 0.000

Wealth

Total value of assets (Ush) 645,245 1,515,850 1,577,302 2,652,796 932,056 0.000

Land holding size (acre) 5.4873 6.9350 7.1657 9.2227 1.6784 0.000

Remittances

1 if received remittance 0.5612 0.4964 0.7841 0.4120 0.2228 0.000

No. of Remittances 2.6412 4.8584 5.4943 7.3756 2.8531 0.000

Total remittance (Ush) 324,182 819,311 859,394 1,486,692 535,212 0.000

Remittances to Consumption Ratio

0.0859 0 .2227 0.1459 0 .3675 0 .0599 0.000

Welfare

Per capita consumption (Ush) 71,196 97,915 102,274 78,073 31,078 0.000

Total wage 15,855 65,032 67,883 184,116 52,027 0.000

HH Characteristics

Head age 50.9789 14.5262 52.2407 13.0762 1.2617 0.138

1 if head is female 0.1431 0.3502 0.1573 0.3646 .01424 0.495

Head education 5.6580 3.8178 7.1908 3.9773 1.5327 0.000

Household Size 7.5667 3.4626 8.5843 3.8779 0.5233 0.007

Village Characteristics

Distance to district town (km) 13.1415 10.7637 8.9965 8.0193 4.1450 0.000

1 if LC1 has MM agent 0.5285 0.4994 0.9719 0.1655 0.4433 0.000

Source: RePEAT 2012.

29

Table 3: Determinants of Household Mobile Money Adoption

(1) (2) Variable Probit FE 1 if mobile phone owned 0.0806*** 0.117*** (0.0142) (0.0273) 1 if HH has migrant worker 0.0349*** 0.0908*** (0.0131) (0.0268) Log of distance to nearest MM agent in km -0.0137*** -0.0442*** (0.00383) (0.0106) HH head’s years of schooling 0.00543*** 0.0115*** (0.00152) (0.00332) Head age 0.00192 0.00471 (0.00234) (0.00472) Head age squared -1.50e-05 -4.16e-05 (2.17e-05) (4.32e-05) Log of land size in acre 0.00207 0.00132 (0.00710) (0.0185) Household size 0.000151 0.000378 (0.00135) (0.00365) 1 if head is female 0.0289 -0.0141 (0.0185) (0.0357) Log value of total assets (UGX) 0.0195*** 0.0248** (0.00485) (0.0114) 1 if year 2012 0.217*** 0.217*** (0.0208) (0.0301) Constant -0.427** (0.189) Observations 1,745 1,745 R-squared 0.448 Number of households 906

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.

30

Table 4A: Mobile money and Household Consumption

Dependent Variable: Household Per capita Consumption

(1) (2) (3) Variable OLS FE FE 1 if mobile money used 0.135*** 0.110* 0.0947*

(0.0394) (0.0565) (0.0565) Constant 9.144*** 8.611*** 9.359*** (0.288) (0.377) (0.383) Time Effects Location*Time

Y Y

Y

Y Y

Observations

1,753

1,753

1,753

R-squared 0.300 0.272 0.379 Number of hhid 914 914

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Table 4B: Mobile money and disaggregated consumption expenditure Dependent Variable: Components of Household Per capita Consumption

Food Expenditure

Non-food Basics Social Contributions

Variable OLS (1)

FE (2)

OLS (3)

FE (4)

OLS (5)

FE (6)

1 if mobile money used

0.0977** -0.0129 0.154*** 0.207** 0.563*** 0.474**

(0.0483) (0.0683) (0.0594) (0.0832) (0.117) (0.187) Constant 10.82*** 11.75*** 7.255*** 8.193*** 6.854*** 7.213*** (0.231) (0.295) (0.236) (0.358) (0.668) (0.893) Observations 1,725 1,753 1,753 1,753 1,725 1,753 R-squared 0.302 0.354 0.303 0.470 0.380 0.373 Number of hhid 914 914 914

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

31

Table 5A: Mobile money and Household Remittances Dependent Variable: Measures of Remittances

Dependent Variable: 1 if Remittances Received

(1) (2) No. of Remittances

(3) (4) Total Remittances (5) (6)

Variable Probit OLS OLS FE OLS FE

1 if mobile money used

0.0706* 0.0581* 0.843** 0.940* 0.360*** 0.381*

(0.0399) (0.0324) (0.421) (0.525) (0.133) (0.220) Constant 0.0273 -5.028** -1.772 5.066*** 5.080*** (0.190) (2.441) (3.354) (0.872) (1.253) Observations 1,702 1,729 1,736 1,736 1,736 1,736 R-squared 0.228 0.188 0.261 0.278 0.286 Number of households

905 905

Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Table 5B: Mobile Money, Job-seeking and Remittances

Dependent Variable: Measures of Remittances

(1) (2) (3) Variable 1 if Remittances

Received Number of

Remittances Total Remittances

1 if mobile money used 0.0952** 1.385** 0.428***

(0.0456) (0.629) (0.163) 1 if HH has migrant worker 0.114*** 1.384*** 0.415*** (0.0327) (0.482) (0.138) Constant 2.831 9.607*** (2.315) (0.605) Observations 1,709 1,736 1,736 R-squared 0.265 Number of households 905 905

Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

32

Table 6: Reduced Form Results

(1) (2) (3) (4) VARIABLES Consumption 1 if

Remittances Received

No. of Remittances

Total Remittances

Log (distance to booth) -0.0481** -0.0211* -0.517*** -0.259** (0.0238) (0.0127) (0.182) (0.123) Constant 11.48*** 0.622*** 1.733 9.642*** (0.257) (0.141) (1.687) (1.104) Observations 1,762 1,750 1,757 1,757 R-squared 0.345 0.216 Number of hhid 915 914 914

Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Table 7: IV and Tobit Results. Dependent Variables: Measures of Consumption and Remittances

Consumption Total Remittances No. of Remittances

Variable FE-IV (1)

Tobit (2)

Tobit-CF (3)

Tobit (4)

Tobit-CF (5)

1 if mobile money used 0.727* 1.160*** 1.002*** 1.449* 1.253 (0.382) (0.369) (0.372) (0.777) (0.782) Residual 6.357*** 20.84*** (2.244) (7.052) Constant 2.139 6.220** -11.58*** 0.883 (1.902) (2.490) (4.439) (6.281) Observations 1,664 1,746 1,746 1,746 1,746 R-squared 0.194

33

Number of hhid 832 Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1 Table 8: Effect of Mobile Phone Possession on Household Consumption and Remittances (2003-2005) (1) (2) (3) (4)

Variable

Consumption Total Remittances No. of Remittances

Prob(remit>0)

1 if mobile phone owned -0.101 -0.250 0.0150 -0.0201

(0.106) (0.475) (0.117) (0.0561)

Constant 8.113*** 6.932*** -0.0561

(0.528) (1.868) (0.431)

Controls Location*Time

Y Y

Y Y

Y Y

Y Y

Observations 1,748 1,735 1,735 1,735

R-squared 0.258 0.152 0.429

Number of households 934 931 931

Notes: Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 The controls included in all the above regressions are; household size, log of asset and land endowments, age, gender and education level of the household head and a dummy for household mobile phone possession. Sample of 2003-2005 used. Appendix

Table A1: Falsification Test-Consumption, Job-seeking behavior and Remittances (2003-2005 sub-sample)

(1) (2) (3) (4)

Variable Consumption 1 if Remittances Received

No. of Remittances

Total Remittances

1 if mobile money used -0.0650 0.0113 -0.108 -0.0788

(0.0675) (0.0418) (0.238) (0.0708)

1 if HH has migrant worker 0.0245 0.105 -0.0589

(0.0350) (0.243) (0.0697)

Constant

8.222***

6.967***

-0.0446

(0.533) (1.870) (0.431)

34

Observations 1,735 1,735 1,735 1,735

R-squared 0.261 0.153 0.431

Number of households 931 931 931

Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Rural Push, Urban Pull and... Urban Push?New Historical Evidence from Developing Countries∗

Remi Jedwab† and Luc Christiaensen‡ and Marina Gindelsky§

January 2014

Abstract: Standard models explain urbanization by rural-urban migrationin response to an (expected) urban-rural wage gap. The Green Revolutionand rural poverty constitute rural push factors of migration. The Indus-trial Revolution and the urban bias are urban pull factors. This paper offersan additional demographic mechanism, based on internal urban populationgrowth, i.e. an urban push. Using newly compiled historical data on urbanbirth and death rates for 7 countries from Industrial Europe (1800-1910)and 33 developing countries (1960-2010), we show that many cities of to-day’s developing world are “mushroom cities” vs. the “killer cities” of In-dustrial Europe; fertility is high, while mortality is much lower. The highrates of urban natural increase have then accelerated urban growth and ur-banization in developing countries, with urban populations now doublingevery 18 years (15 years in Africa), compared to every 35 years in IndustrialEurope. This is further found to be associated with higher urban congestion,possibly mitigating the benefits from agglomeration and providing furtherinsights into the phenomenon of urbanization without growth. Both migra-tion and urban demographics must be considered in debating urbanization.

Keywords: Urbanization; Demographic Transition; Migration; Poverty; SlumsJEL classification: O1; O18; R11; R23; J11;

∗We would like to thank Paul Carrillo, Carmel Chiswick, Denis Cogneau, Jeremiah Dittmar, Dou-glas Gollin, James Foster, Fabian Lange, William Masters, Jean-Philippe Platteau, Harris Selod,Stephen Smith, David Weil, Anthony Yezer and seminar audiences at EUDN Scientific Conference(Berlin), George Mason-George Washington Economic History Workshop, George Washington (IIEPand SAGE), Harvard Kennedy School (NEUDC), Paris School of Economics, University Paris 1, the Ur-ban Economic Association meetings (Atlanta) and World Bank-George Washington University Con-ference on Urbanization and Poverty Reduction 2013 for very helpful comments. We thank theInstitute for International Economic Policy at George Washington University for financial assistance.

†Corresponding Author: Remi Jedwab, Department of Economics, George Washington University,2115 G Street, NW, Washington, DC 20052, USA (e-mail: jedwab@gwu.edu).

‡Luc Christiaensen, Development Research Group, The World Bank, 1818 H St NW, Washington,DC 20433, USA (e-mail: lchristiaensen@worldbank.org).

§Marina Gindelsky, Department of Economics, George Washington University, 2115 G Street, NW,Washington, DC 20052, USA (e-mail: marina_g@gwu.edu).

1. INTRODUCTION

Developing countries have dramatically urbanized over the past 60 years (WorldBank, 2009). While their urbanization process shares many similarities with theurbanization process of developed countries in the 19th century, the two processesalso differ in several dimensions. First, urban growth has been faster in today’sdeveloping world. The Industrial Revolution led to a dramatic acceleration of ur-banization (see Figure 1): Europe’s urbanization rate increased from about 15% in1800 to 40% in 1910. In 1950, Africa and Asia were made up of predominantlylow-income, rural countries (urbanization rate around 15%). In 2010, their urban-ization rate was around 40%. African and Asian countries have thus experiencedthe same growth in urbanization as Europe, in half the time. Second, while incomegrowth remains the main driver of urbanization, the world is becoming more andmore urbanized at a constant income level. In 1960, the 35 countries whose incomeper capita was less than $2 a day had an average urbanization rate of 15% (WorldBank, 2013). In 2010, the 34 countries with similar incomes had an average rate of30%. The cities of today’s developing world are also much larger. Mumbay, Lagosand Jakarta have the same population as New York, Paris and London respectively,at a much lower income level. Dhaka, Kinshasa and Manila are urban super-giantslocated in very poor countries. This raises several questions. Where do these citiescome from? Did they grow as a result of migration? Did they grow too fast?

In models of urbanization, there is rural-to-urban migration as long as the expectedurban real wage is higher than the rural real wage (Harris & Todaro, 1970). Thiswage gap could be the result of a rural push or an urban pull. There are variousrural push factors. If the country experiences a Green Revolution, the rise in foodproductivity releases labor for the modern sector and people migrate to the cities(Schultz, 1953; Matsuyama, 1992; Caselli & Coleman II, 2001; Gollin, Parente &Rogerson, 2002; Nunn & Qian, 2011; Motamed, Florax & Matsers, 2013). Ruralpoverty due to land pressure or natural disasters causes rural migrants to flock tocities (Barrios, Bertinelli & Strobl, 2006; da Mata et al., 2007; Yuki, 2007; Poel-hekke, 2010; Henderson, Storeygard & Deichmann, 2013).1 Then there are variousurban pull factors. If the country experiences an Industrial Revolution, the urbanwage increases, which attracts workers from the countryside (Lewis, 1954; Hansen& Prescott, 2002; Lucas, 2004; Alvarez-Cuadrado & Poschke, 2011). A country thatexports natural resources also urbanizes if the resource rents are spent on urbangoods and services, causing the urban wage to rise (Gollin, Jedwab & Vollrath,2013; Jedwab, 2013). If the government adopts urban-biased policies, the urbanwage also increases (Lipton, 1977; Bates, 1981; Ades & Glaeser, 1995; Davis & Hen-derson, 2003; Majumdar, Mani & Mukand, 2004; Shifa, 2013). While the GreenRevolution, Industrial Revolution and resource exports theories find that urbaniza-tion is associated with economic development, the rural poverty and urban biastheories imply that urbanization may occur without growth (Fay & Opal, 2000). Allthese theories assume that urbanization comes from migration only.

1Overoptimistic expectations about the incomes migrants can earn at the destination location alsocreate excessive migration pressure (McKenzie, Gibson & Stillman, 2013; Farré & Fasani, 2013).

1

In this paper, we offer an additional mechanism for urbanization based on an urbanpush. Many cities of today’s developing world can be classified as “mushroom cities”vs. the “killer cities” of the developing world of the 19th century; fertility is high,while mortality has fallen to low levels, due to the epidemiological transition of the20th century. This has led to a high rate of natural increase in urban areas. First, weshow that the urban push has accelerated urban growth and urbanization in devel-oping countries, conditional on income. Second, we show that fast urban growthis associated with more congested cities, which has implications for economic de-velopment. We use the expression “urban push” as opposed to the “rural push” and“urban pull”. “Rural push” implies that rural workers are pushed to the cities bychanges in rural economic conditions. “Urban pull” implies that rural workers areattracted to the higher-wage cities. “Urban push” suggests that cities are growinginternally and “pushing” their own boundaries. It is not that urban workers are be-ing pushed to the countryside, but rather, high urban rates of natural increase arecreating an urban population “push”. Our analysis consists of three steps.

First, we provide historical evidence on the rapid growth of cities in today’s devel-oping world. The growth rate of the urban population has been about 4% a year indeveloping countries post-1960, vs. 2.0% a year in Industrial Europe in 1800-1910(see Figure 2). We then use various historical country-level sources to create an ex-tensive new data set on the crude rates of birth and death separately for the urbanand rural areas of 7 European (or Neo-European) countries in the 19th century (ev-ery forty years in 1800-1910) and 33 countries that were still developing countriesin 1960 (every ten years in 1960-2010). We can thus accurately compare the demo-graphic foundations of the urbanization processes of the old and new developingworlds.2 We show that the fast growth of cities in today’s developing world wasmostly driven by natural increase, and not by migration as in Europe. We confirmthat the cities of Industrial Europe were “killer cities”, where mortality was high andfertility was low. On the contrary, the cities of today’s developing world are “mush-room cities”, where fertility is high and mortality is low. The resulting difference inurban rates of natural increase caused the population of cities in today’s developingworld to double every 18 years (15 years in Africa), compared with 35 years inIndustrial Europe. Even if natural increase contributed to urban growth, and raisedthe absolute number of urban residents, it also contributed to rural growth. Theurbanization rate, the relative number of urban residents, may not have risen as aresult. Yet simulations suggest it also increased urbanization rates.

Second, we use our panel data set on 33 countries (1960-2010) to investigateeconometrically the effects of urban natural increase on the speeds of urban growthand urbanization. We show that the stylized facts that have been established by thecomparative analysis hold when including country and decade fixed effects, con-trolling for income growth and the various rural push and urban pull factors that

2Our analysis builds on the previous work of historians and geographers such as Rogers (1978),Keyfitz (1980) and Rogers & Williamson (1982). We complete their preliminary analysis by using his-torical data on 40 “developing” countries, past and present, in two centuries. First, most economistshave focused on the individual cases of England or the U.S. in the 19th century (Williamson, 1990;Haines, 2008). We have been able to collect the same type of data for as many as 7 European coun-tries, which allows us to generalize their results for the old developing world. Second, while thereare individual case studies for a few developing countries for selected periods, we have systemati-cally collected the same type of data for 33 countries every ten years from 1960 to 2010. We couldnot increase the sample size as historical consistent data does not exist for other countries as far backas 1960. The numerous historical sources that we used are described in the Online Data Appendix.

2

are traditionally put forward in the literature, and even adding region fixed effectsinteracted with a time trend (e.g., Western Africa, Eastern Africa, etc.). The iden-tification then comes from the within-country comparison of neighboring countriesof the same region over time. Even if we cannot be sure that our effects are causal(as is often the case with cross-country regressions), we are able to rule out manypotential alternative explanations. Since we study an important macro question,we must use macroevidence, even if the effects will not be as well identified asin the microdevelopment literature (see Cohen & Easterly (2010) for a descriptionof how different methodologies can help address different questions). The resultsalso hold when using cross-sectional data for 97 countries that were still developingcountries in 1960, but for the most recent period only. Urban natural increase has astrong effect on urban growth and urbanization. A 1 standard deviation increase inthe rate of urban natural increase leads to a 0.50 standard deviation increase in theurban growth rate and a 0.30 standard deviation increase in the change in urbaniza-tion. We find that differences in urban natural increase explain why urban growthhas been faster in today’s developing world, and in Africa in particular. These dif-ferences may also contribute to explaining why African and Asian countries haverecently experienced the same growth in urbanization as Industrial Europe, but inhalf the time, and why Africa is relatively urbanized for its income level. The urbanpush has thus accelerated the speed of urban growth and urbanization.

Third, fast urban growth can give rise to urban congestion, which may decreaseurban welfare. If capital (e.g., houses, schools, hospitals and roads) cannot beaccumulated as fast as population grows, cities grow too fast and the stock of urbancapital per capita is reduced. If the urban population of today’s developing worlddoubles every 18 years, the housing stock also needs to double every 18 years.Congestion effects arise if agents are not investing in advance, whether they arecredit-constrained or not forward-looking. Urban labor supply shocks can also leadto a deterioration of urban labor market outcomes. Using a novel data set on urbancongestion for a large set of countries, we show that fast urban growth due tonatural increase is indeed associated with more congested cities today. The urbanpush is correlated with a higher proportion of urban population living in slums,lower investment in urban human capital, more polluted cities, and more workers inthe urban informal sectors. The evidence suggests a world in which slums developnot just because migrants flock to cities, but also as a result of internal growth.We do not find any effect of the speed of urbanization, as what matters for urbancongestion is really the absolute, rather than relative, number of urban residents.Our results are all the more important since fertility remains high in many cities,that will keep growing in the future. There are still 30 countries where the urbanpopulation doubles in less than 18 years, indicating the scope of the problem.

The paper also contributes to the literature on urbanization and growth. There is astrong correlation between development and urbanization, because of the two-wayrelationship between them. On the one hand, countries urbanize when they develop(Overman & Venables, 2005; Henderson, 2010; Henderson, Roberts & Storeygard,2013). On the other hand, agglomeration promotes growth (Rosenthal & Strange,2004; Glaeser & Gottlieb, 2009; Henderson, 2010). Given that urbanization is aform of agglomeration, cities could promote growth in developing countries (Du-ranton, 2008, 2013; World Bank, 2009). Urban natural increase can, however,create a disconnect between urbanization and growth. First, poor cities can ex-pand even without an increase in standards of living. We provide an explanation

3

for over-urbanization, additional to the existing theories of urban bias and ruralpoverty. Second, because natural increase accelerates urban growth, it can give riseto urban congestion effects, which may reduce the benefits from agglomeration.The speed of urban growth is, to our knowledge, a dimension of the urbanizationprocess that has been understudied in the economics literature. All in all, urbannatural increase in poor countries may have thus directly contributed to the “ur-banization of poverty”, the fact that the urban areas’ share of the world’s poor hasbeen rising over time (Ravallion, 2002; Ravallion, Chen & Sangraula, 2007). Third,whether urban growth is driven by migration or natural increase has strong policyimplications. When urban congestion is the result of excessive migration, it maynot be justified to invest in urban infrastructure, as it could further fuel migration.However, if urban growth is due to urban natural increase, the resulting immediateincrease in the urban population necessitates investment in urban infrastructure.If agents do not internalize the negative externalities associated with their fertilitydecisions, another policy option may be to encourage lower urban fertility rates.Lastly, we have created a consistent data set that will allow researchers to system-atically study the urbanization process across space and time. Bandiera, Rasul &Viarengo (2013) provide another example of how collecting historical demographicdata can help us revisit issues that are still extremely relevant today.

Our findings also advance the literature on the effects of demographic growth.Population growth promotes economic growth if high population densities encour-age human capital accumulation or technological progress (Kremer, 1993; Becker,Glaeser & Murphy, 1999; Lagerlöf, 2003). However, population growth has a nega-tive effect on per capita income if capital (e.g., land) is inelastically supplied. Anypositive income shock is then temporary; fertility increases and mortality decreases,so that any increases in the stock of capital (and income) per capita are eventuallynegated. Income is stable and low in the long-run.3 Countries only develop if tech-nology progresses and the demographic transition limits population growth (Galor& Weil, 1999, 2000; Hansen & Prescott, 2002). If the economy is Malthusian, anyincrease (decrease) in population decreases (increases) the capital-labor ratio andper capita income.4 In this paper, we use an increase in population, studying it fromthe perspective of cities. Second, since urban space is constrained, the potential forcongestion effects is high. Third, there are few studies of the effects of populationgrowth in Africa (Young, 2005; Ashraf, Weil & Wilde, 2011; McMillan, Masters &Kazianga, 2011). We show that African cities will keep growing at a fast pace in thefuture, which has implications for the growth process of the continent.

The paper is organized as follows: Section 2 offers a framework to analyze theeffects of urban natural increase. Section 3 presents the historical background andthe data. Sections 4, 5, and 6 show the effects of urban natural increase on urbangrowth, urbanization and urban congestion respectively. Section 7 concludes.

3During the Malthusian growth regime, the most advanced societies have larger populations, butnot significantly higher incomes (Diamond, 1997; Ashraf & Galor, 2011; Vollrath, 2011).

4A few studies have examined the effects of disease eradication on mortality, population growthand economic development (Acemoglu & Johnson, 2007; Bleakley, 2007; Bleakley & Lange, 2009;Bleakley, 2010; Cutler et al., 2010). Other studies have looked at the effects of decreases in popula-tion on development, whether these are caused by disease, war or fertility restrictions (Young, 2005;Voigtländer & Voth, 2009; Ashraf, Weil & Wilde, 2011; Voigtländer & Voth, 2013a,b).

4

2. CONCEPTUAL FRAMEWORK

This section provides a simple framework to analyze the relationships between nat-ural increase, migration, urban growth, urbanization and urban congestion.

2.1 Urban Natural Increase and Urban Growth

Urban growth consists of four components: urban natural increase, rural-to-urbanmigration, international-to-urban migration and urban reclassification. There arerural (international) migrants as long as the urban wage is higher than the ruralwage (wage in the country of origin). We abstract from the issues of expectations,prices and amenities to simplify the analysis. The wage gap could be the result ofan urban pull or a rural push. Lastly, rural land is reclassified as urban when villagesare absorbed by a city, or when a locality becomes urban given the urban definition.In many countries, a locality is considered urban if its population size exceeds acertain population threshold. The equations of urban and rural growth are:

4U popt = Unit ∗ U popt + Rmigt + IUmigt + U rect (1)

4Rpopt = Rnit ∗ Rpopt − Rmigt + IRmigt − U rect (2)

where 4U popt (4Rpopt) is the growth of the urban (rural) population in yeart, Unit (Rnit) is the urban (rural) crude rate of natural increase in year t, U popt

(Rpopt) is the urban (rural) population at the start of year t, Rmigt is the numberof net rural-to-urban migrants in year t, IUmigt (IRmigt) is the number of netinternational-to-urban (rural) migrants in year t, and U rect is the number of ruralresidents reclassified as urban in year t. The urban (rural) crude rate of naturalincrease is the urban (rural) crude birth rate minus the urban (rural) crude deathrate. If urban (rural) fertility is higher than urban (rural) mortality, the urban (ru-ral) rate of natural increase is positive, and the urban (rural) population expands.Equation (1) must be divided by the urban population at the start of year t to beexpressed in percentage form. The number of “residual migrants” (Migt) is definedas the sum of rural migrants, international migrants and rural residents reclassi-fied as urban. The urban growth rate is thus equal to the sum of the rate of urbannatural increase (Unit) and the “residual migration” rate (Migt/U popt):

4U popt/U popt = Unit +Migt/U popt (3)

2.2 Urban Natural Increase and Urbanization

The urbanization rate at the start of year t, Ut , is the ratio of the urban popula-tion U popt to the total population Popt . The change in the urbanization rate inyear t, 4Ut , is positive if urban growth is faster than rural growth. Even if naturalincrease contributes to urban growth, it also contributes to rural growth. For coun-tries that are mainly rural, rural natural increase disproportionately augments thesize of the rural population: RnitRpopt ≥ Unit U popt , even if Rnit ≤ Unit , becauseRpopt ≥ U popt . Therefore, natural increase reduces the urbanization rate for apredominantly rural country. As the country becomes more urbanized, the contri-bution of urban natural increase to urbanization rises. In countries that are alreadyurbanized, this contribution declines, as there is less room to grow. We expect aninverted-U relationship between the change in urbanization and urban growth. Tostudy this relationship, we decompose the change in the urbanization rate using theequations above. Nnit is the national rate of natural increase in year t. The other

5

variables are the same as above. We obtain the following equations:

4Ut =U popt+1

Popt+1−

U popt

Popt=

U popt+1

Popt+1

Rpopt

Popt−

U popt

Popt

Rpopt+1

Popt+1(4)

4Ut = (1− Ut)(1+ Unit)U popt +Migt

(1+ Nnit)Popt− Ut

(1+ Rnit)Rpopt −Migt

(1+ Nnit)Popt(5)

4Ut =Ut

(1+ Nnit)[(1− Ut)(Unit − Rnit) +

Migt

U popt] (6)

The change in urbanization positively depends on the differential between the ur-ban and rural rates of natural increase (Unit vs. Rnit) and the “residual migration”rate (Migt/U popt). It also depends on the initial urbanization rate (Ut) and ag-gregate natural increase (Nnit , which is a function of Rnit , Unit and Ut). To studythe potential effect of urban natural increase, we simulate equation (6) using thefollowing parameters: Rni = 2.5% and Migt/U popt = 1.5% per year. These val-ues have been chosen based on the comparative analysis in section 3.7. We useUni = 0.5% as a benchmark to see how raising the urban rate of natural increasealters urbanization. Figure 3 shows the results of the simulation for five values ofUnit = {1; 1.5;2; 2.5;3}, given an initial urbanization rate Ut . The effects are large.Increasing the urban rate of natural increase from 0.5% to 3% raises the changein the urbanization rate by 0.45 percentage points on average. As aforementioned,the effects are higher for median values of the urbanization rate.

2.3 Urban Natural Increase and Urban Congestion

Cities grow too fast if urban population grows faster than urban capital, and thestock of capital per capita decreases. Various types of capital could be accumu-lated: physical and human capital, the housing stock, or transport infrastructure.Assuming that capital cannot be accumulated as fast as population grows, fast urbangrowth leads to urban congestion. For example, raising the urban rate of naturalincrease from 0.5% to 3%, given a migration rate of 1.5%, causes the urban popu-lation to double every 15 years, instead of 35 years. Then, the urban housing stockalso needs to double every 15 years. This is possible if the urban growth is not un-expected, agents are forward-looking, and have sufficient credit available to makethe investment. If not, congestion effects are likely to arise when urban growth isfast. We expect a lower effect of the change in urbanization, as what matters forurban congestion is the absolute, rather than relative, number of urban residents.Though urban congestion may reduce future migration, migration may still remainhigh as it depends on the difference between rural and urban welfare.

2.4 Empirical Considerations

Urban natural increase may determine the speeds of urban growth and urbaniza-tion. The speed of urban growth is then a factor of urban congestion. We nowdiscuss various issues regarding the empirical analysis of these relationships.

Dynamic model. Equation (3) assumes that the relationships between urban growthand its two components are additive. When estimating this relationship empirically,the coefficient of the rate of urban natural increase could be equal to one. However,we could imagine that urban natural increase and migration influence themselves

6

and each other dynamically, which could bias (downward or upward) the coeffi-cient of the urban rate of natural increase. Four relationships should be considered:(i) Migt = f (Migt−1): High migration rates have a dissuasive effect on future mi-gration, if the migrants crowd out the cities, or if the pool of potential migrants isreduced, (ii) Unit = g(Migt−1): Urban residents adjust their fertility rates if mi-grants crowd out the cities. However, migration may actually have a positive effecton future urban fertility if urban congestion impoverishes everyone, which preventsany adjustment in fertility. Fertility is indeed higher in poorer contexts, becauseof the trade-off between child quantity and child quality. Besides, a high share ofmigrants in the urban population also affects urban fertility and mortality if it al-ters the age structure of the cities. If migrants are of reproductive age, migrationalso increases future urban fertility, (iii) Migt = h(Unit−1): Urban natural increasehas a dissuasive effect on future migration, if the urban newborns crowd out thecities, and (iv) Unit = j(Unit−1): Urban residents adjust their fertility rates if urbannewborns crowd out the cities. However, urban natural increase may actually havea positive effect on future urban fertility if urban congestion impoverishes every-one, which prevents any adjustment in fertility. Lastly, urban natural increase couldalso affect the age structure of the cities. We will control for these four dynamicrelationships in the analysis to test the additivity and causality of the effects.

Urban reclassification. Births and deaths are usually registered depending on themain place of residence. This location is classified either as urban or rural, whichpermits the estimation of urban and rural birth and death rates. This is importantwhen distinguishing the effects of natural increase and migration. For example, achild who is born in an urban family is counted as “urban”, no matter whether thefamily moved to the city twenty years prior or just the year before the census. Thefamily contributes to the urban population, because it lives in a city. However, achild that follows her parents when they migrate to a city is also counted as a ruralmigrant. There could be composition effects as argued above, hence the need tocontrol for past migration.5 Urban reclassification could then be higher in countrieswhere the urban rate of natural increase is high, since the rural rate of naturalincrease could also be high in such countries (U rect = ϕ(Rnit)). Fast rural growthcould increase overall population densities, and the largest villages could becomecities. Or it could increase the pool of potential rural migrants. Another possibilitycould be that, in countries where urban growth is fast due to natural increase, citiesdisproportionately absorb their surrounding rural areas when they expand spatially(U rect = χ(4Ut−1)). These mechanisms could lead to an upward bias, if urbanreclassification is indeed more important in countries where urban natural increaseis high. Therefore, it will be essential to control for the effects of rural naturalincrease and urban growth on future urban growth via urban reclassification.

Causality. Though the previous analysis treats urban natural increase as exogenous,it could be endogenously determined by the economic conditions in the cities (i.e.,the urban wage). We will show in section 3.3 that urban mortality does not varymuch across countries, and that urban fertility is the main determinant of urban nat-ural increase. Many low-income countries have not yet completed their urban fertil-ity transition. Higher returns to education in fast-growing countries have somewhat

5The numbers of urban newborns and residents are estimated using permanent residence. Tem-poral migrants contribute to the rural population, and their newborns are counted as “rural”. In ouranalysis, we focus on permanent residence, since this is what matters for urbanization.

7

modified the trade-off between child quantity and quality in favor of child quality.We could thus expect higher urban fertility rates in poorer and less urbanized coun-tries. In accordance with convergence, less urbanized countries should urbanizefaster than more urbanized countries. This could give rise to multiple equilibria. Incountries that have already achieved their urban fertility transition, urban growthis slower, and urban congestion effects are limited. If urban congestion (e.g., roadcongestion) reduces urban productivity, growth in these areas is only slightly af-fected by congestion. If income remains high, fertility stays low. Countries in whichurban fertility is high experience fast urban growth. If urban growth is too fast,urban congestion effects kick in, which lower urban productivity. If income is low,urban fertility remains high, and urban fertility and urban congestion reinforce eachother. The urbanization rate will not increase if rural growth is also high, as ruralfertility does not adjust. That is why it will be important in our empirical analysisto compare countries with similar initial income and urbanization levels, but whoserates of urban natural increase differ. This will not solve the endogeneity issue, butthis will allow us to show that urban natural increase is associated with the urbanoutcomes, conditional on the feedback mechanism discussed above. In the panelanalysis, we will also include country and decade fixed effects, controls for the ruralpush and urban pull factors of urbanization as well as the relationships discussedabove, and even region fixed effects interacted with a time trend. The effect is notcausal if there are still unobservable factors that explain why urban natural increaseand the urban outcomes are correlated over time within countries, relative to theneighboring countries of the same region, conditional on the numerous controls weinclude. While we cannot be sure that our effects are entirely causal, we are thusable to rule out many potential alternative explanations.

3. DATA AND BACKGROUND

We now discuss the historical background and the data we use in our analysis. TheOnline Data Appendix contains more details on how we construct the data.

3.1 New Data for Developing Countries, 1700-2010

In order to analyze the contribution of urban natural increase to urban growth andurbanization, we need historical data on urbanization, urban fertility and urbanmortality for the developing worlds of the 19th and 20th centuries. First, we com-pile data from various sources to reconstruct the urban growth and urbanizationrates for 19 European and North American countries from 1700-1950 (about everyforty years), and 116 African, Asian and non-North American countries that werestill developing countries in 1960, from 1900-2010 (about every ten years). Thisallows us to compare the urbanization process of five “developing” areas: “Indus-trial Europe” (which includes the United States in our analysis), Africa, Asia, LatinAmerica (LAC) and the Middle-East and North Africa (MENA). Second, we obtainhistorical demographic data for 40 of these countries: 7 European countries forthe 1700-1950 period (about every forty years), and 33 countries in Africa (10),Asia (11), the LAC region (8) and the MENA region (11) for the 1960-2010 period(about every ten years). For each country-period observation, we obtained the na-tional, urban and rural crude rates of birth, crude rates of death and crude ratesof natural increase (per 1,000 people). Since historical demographic data was notreadily available, we recreated the data ourselves using various historical sources,

8

as well as the UN Statistical Yearbooks and various reports of the Population andHousing Census, the Fertility Surveys and the Demographic and Housing Surveys ofthese countries.6 We then collect the same type of data for as many countries aspossible that were still developing countries in 1960 (N = 97 out of the full sampleof 116 countries), but for the most recent period only (for the closest year to theyear 2000). We also have demographic data for the largest city only.

3.2 Patterns of Urbanization in Developing Countries, 1700-2010

The most advanced civilizations before the 18th century had urbanization rates ofaround 10%-15% (Bairoch, 1988). When a few countries industrialized, their ur-banization rates dramatically increased, usually from 10% to 40%. Figure 1 showsthe urbanization rate for Industrial Europe from 1700-1950 (using the full sampleof 19 countries). The urbanization rate was stable (around 12.5%) until 1800 andincreased to 41.3% in 1910. Countries that industrialized earlier also urbanizedearlier. Figure 1 also shows the urbanization rate for four developing areas (usingthe full sample of 116 countries): Africa, Asia, LAC and MENA. The LAC regionhad already surpassed the 40% threshold in 1950, while the MENA region did notsurpass it until 1970. In 1950, Africa and Asia were made up of predominantlylow-income, rural countries (urbanization rate around 10%). In 2010, their urban-ization rate was around 40%. In our analysis, we focus on the 1800-1910 periodfor Europe and the 1960-2010 period for Africa and Asia. During these periods, theurbanization rates of the three areas increased from 10% to 40%.

3.3 Urban Growth Rates in Developing Countries, 1700-2010

Figure 2 shows the urban growth rate for Industrial Europe from 1700-1950 (N= 19). It peaked in the late 19th century and declined in the 20th century. Inthe 1800-1910 period, the overall urban growth rate was 2.0% per year. Figure 2also shows the urban growth rate for the four developing areas from 1900-2010 (N= 116). The urban growth rate has been 3.8% on average in today’s developingworld post-1960, and 4.7% a year in Africa, compared to 3.4%, 3.2% and 4.0% inAsia and the LAC and MENA regions respectively. An urban growth rate of 3.8%(or 4.7% as seen in Africa) implies that cities double every 18 (15) years, while arate of 2.0%, as seen in Europe, means that cities double every 35 years. Theserates peaked in the 1950s or 1960s, with the acceleration of rural migration andthe demographic transition. They have been declining since, although they are stillhigh today. We obtain similar urban growth rates when considering the largestcity only. We now use our data to provide descriptive evidence on the respectivecontributions of natural increase and migration to urban growth and urbanizationfor the 40 countries for which we have historical demographic data.

3.4 The “Killer Cities” of Industrial Europe

We use data for 7 countries from 1700-1950 to explain the concept of “killer cities”(Williamson, 1990). We focus on English cities as a classical example. Demographicpatterns in English cities have been described by Williamson (1990), Clark & Cum-mins (2009) and Voigtländer & Voth (2013b). We add to this literature by collecting

6The list of the 40 developing countries that we use in the main analysis, and the data sourcesfor each country are reported in the Online Data Appendix and Online Appendix Tables 1, 2 and 3.We could not increase the sample size as historical consistent data does not exist for other countries.

9

the same data for 6 other countries, which allows us to generalize the results. Re-sults are shown in Figure 4. Fertility was relatively low in England (about 35 per1,000 people before 1910).7 Mortality was high, especially in the cities. In the19th century, the urban death rate was 10 points higher on average than the ruraldeath rate (about 30 vs 20). High urban densities, industrial smoke, polluted watersources and unhygienic practices all contributed to this urban penalty (Williamson,1990; Voigtländer & Voth, 2013b). As a result, the average rate of urban naturalincrease was low in 1800-1910, at 5 per 1,000 people (or 0.5%). Online AppendixTable 1 shows that these patterns are present for the six other countries. In allcountries, the contribution of urban natural increase to urban growth was less than0.6% a year in 1800-1910: 0.5% in England vs. 0.5% in Belgium, 0.1% in France,0.6% in Germany, 0.4% in the Netherlands, 0.3% in Sweden and 0.4% in the UnitedStates. The average rate was 0.5% a year for Industrial Europe.

3.5 The “Mushroom Cities” of The Developing World

We use data on 33 countries from 1960 to 2010, to explain the concept of “mush-room cities”. Figure 5 plots the urban and rural birth rates for the four developingareas in 1960-2010.8 Initially, urban fertility was high in developing countries, andin Africa in particular (about 50 per 1,000 people). Urban fertility rates decreasedalmost everywhere post-1960, yet they remain high in Africa (about 35). Figure 6then plots the urban and rural death rates from 1960-2010.9 In 1960, urban deathrates were already low in most of the developing world, around 10-20. Acemoglu& Johnson (2007) show that the epidemiological transition of the mid 20th cen-tury (e.g., the discovery and consequent mass production of penicillin in 1945) andmassive vaccination campaigns in the colonies resulted in widespread and signifi-cant declines in mortality. The acceleration of urban growth in the 1950s illustratesthis phenomenon (see Figure 2). The colonizers also invested in health, educationaland transport infrastructure, which led to higher standards of living, as shown byanthropometric and other development outcomes (Moradi, 2008; Huillery, 2009;Jedwab & Moradi, 2013). Cities were centers of diffusion of innovation, explainingwhy urban mortality was low initially. Differences in urban natural increase are thusdriven by differences in urban fertility. While urban mortality does not vary muchacross countries, urban natural increase is highly correlated (correlation coefficientof 0.93) with urban fertility, whose variance is much higher (Online Appendix Fig-ure 2 shows this for 97 countries). Figure 7 then shows the rates of natural increasefrom 1960-2010. These rates were high both for the cities and the countrysideacross all regions in 1960 and have been decreasing since. While urban naturalincrease was high in the LAC and MENA regions in 1960, these areas have almost

7Most European countries were then characterized by the “European Marriage Pattern”, in accor-dance with which women married late and fertility was lower (Hajnal, 1965). What explains thisspecific pattern is unclear, but Voigtländer & Voth (2013a) show how the Black Death in the 14thcentury had a long-term impact on marital and fertility patterns.

8The birth rate is a function of the total fertility rate and the number of women of reproductiveage. The urban fertility rate is the main determinant of urban birth rates. For 97 developing coun-tries for which we have data for the closest year to the year 2000 in the interval 1990-2010, thecorrelation coefficient between the two variables is 0.93 (see Online Appendix Figure 1).

9The death rate is a function of the child mortality rate (0-5 years), the youth mortality rate (5-15years) and the adult mortality rate (15 and above years). At the cross-country level in developingcountries, child mortality is the main factor of aggregate mortality. We focus on the period 1960-2010, while HIV-related adult mortality only became a major concern in the 2000s. For example, inSouthern Africa, the average prevalence rate was about 20% in 2000 and 2010, but 2.5% in 1990.

10

completed their fertility transition. Asia started its transition earlier. Then, urbannatural increase is still more important in Africa in 2010 than it was in Asia in 1960.African cities will keep growing due to natural increase for several decades.

3.6 Urban Natural Increase and Urban Growth

We use equation (3) to decompose urban growth into urban natural increase andresidual migration for the 40 countries. Figure 3 shows the decomposition in Eng-land from 1700-1910. Urban growth was driven by migration, while the contribu-tion of natural increase was small. England could not have urbanized without ruralresidents migrating to unhealthy urban environments. Results from the six othercountries confirm these patterns (see Online Appendix Table 1). During the 1800-1910 period, Industrial Europe’s urban growth was 2.2% per year, while the urbanrate of natural increase was 0.5%. The difference, about 1.7%, was accounted forby residual migration. Figure 8 shows the decompositions for the four develop-ing regions (N = 33), as well as the decompositions for England (1700-1950) andthe developing world (1960-2010) (see Online Appendix Table 2 for each country).Migration rates, which average 1.6%, were not different in developing countries(post-1960) from Industrial Europe. The difference in urban growth (3.8% vs.2.2%) comes from urban natural increase (2.3% vs. 0.5%), which accounted foralmost two thirds of urban growth post-1960. Urban growth was faster in Africa(4.9%) than in the MENA region (3.6%), Asia (3.5%) and the LAC region (3.1%)because the urban rate of natural increase was also higher. While it was 2.9% onaverage in Africa, it was 2.6% in the MENA region, 1.6% in Asia and 2.2% in theLAC region. Therefore, across space and time, the contribution of migration to ur-ban growth was around 1.5% per year. Countries differed in their urban growth asa result of urban natural increase only. For example, using an urban rate of natu-ral increase of 2.9% (1.6%), as in Africa (Asia), a family of four migrants in 1960becomes a family of about fifty (thirty) urban residents in 2010.

3.7 Urban Natural Increase and Urbanization

Europe and the four developing areas widely differed in their urban rates of naturalincrease. On average, their rural rates of natural increase were much more similar:around 2% in Europe and Asia, and 2.5% in other regions. In Figure 3, we simulatedequation (6), using the following parameters: Rnit = 2.5% and Migt/U popt =1.5% per year. We used Uni = 0.5% as a benchmark, and showed the results of thesimulation for five values of Unit = {1;1.5; 2;2.5; 3}, given an initial urbanizationrate Ut . This allows us to compare the potential effects of urban natural increaseceteris paribus for East Asia (Unit ≈ 1%), Asia (1.5%), the LAC region (2%), theMENA region (2.5%), and Africa (3%), relative to Europe (0.5%). The annualeffects are potentially large (e.g. 0.2 points of urbanization for Africa, given aninitial urbanization rate of 10%). The larger the urban rate of natural increase, andthe closer to 50% the initial urbanization rate, the larger the effect on urbanization.In 2010, Africa’s urbanization rate was about 40% and urban natural increase was2.5%. The urbanization rate could increase to 45% in 2020.

4. RESULTS ON URBAN GROWTH

In this section, we use econometric regressions and our panel data for 33 countries(1960-2010) to investigate the effects of urban natural increase on urban growth.

11

4.1 Main Results

We use panel data for 33 countries that were still developing countries in 1960. Werun the following model for t = [1960s, 1970s, 1980s, 1990s, 2000s]:

U grc,t = α+ βUnic,t + γc +δt + uc,t (7)

where U grc,t is the annual urban growth rate (%) of country c in decade t. Ourvariable of interest is the urban rate of natural increase (per 100 people, or %) ofcountry c in decade t (Unic,t). All regressions include country and decade fixedeffects (γc; δt). The country fixed effects control for time-invariant heterogene-ity at the national level. The identification of the effect then comes from decadalvariations in urban natural increase within countries. Table 1 presents the results.Column (1) shows that urban natural increase has a strong effect on urban growth(0.95***). In column (2), we show that this effect is robust to controlling for logGDP per capita and the urbanization rate at the start of the decade, and log GDPper capita at the end of the decade. First, poor countries have a high fertility rate(they have not completed their fertility transition yet) and their cities will growfaster since they are initially smaller. Controlling for initial income and urbaniza-tion adjusts for these convergence effects. Second, since we are controlling forincome at the end of the decade, our effects are estimated conditional on contem-porary income and income growth during the decade. This allows to measure thecontribution of urban natural increase to urbanization without growth.

There are several alternative theories for urbanization in developing countries thatmay make the results in columns (1) and (2) spurious. We include four area fixedeffects (Africa, Asia, LAC and MENA) interacted with a time trend to control fortime-variant heterogeneity at the continental level. We also control for the variousrural push and urban pull factors mentioned in the conceptual framework. Includ-ing income in the regression controls for the Green and Industrial Revolutions, asthe structural change literature has shown how they were highly correlated. Wealso include the following controls at the country level: (i) Green Revolution (ruralpush): average cereal yields (hg per ha) in the same decade; (ii) Industrial andService Revolutions (urban pull): the share of manufacturing and services in GDP(%) 2010 interacted with decade fixed effects (the same share is missing for toomany countries in earlier decades); (iii) natural resource exports (urban pull): theshare of natural resource exports in GDP (%) in the same decade; (iv) rural poverty(rural push): rural density (1000s of rural population per sq km of arable area),the number of droughts (per sq km), and an indicator equal to one if the countryhas experienced a civil or interstate conflict in the same decade to control for landpressure and disasters; and (v) urban bias (urban pull): an indicator equal to oneif the country’s average combined polity score is strictly lower than -5 (the countryis then considered autocratic according to Polity IV), and the primacy rate (%) - analternative measure of urban bias - in the same decade. The urban bias was indeedstronger in more autocratic regimes (Ades & Glaeser, 1995; Shifa, 2013). The in-clusion of area fixed effects (column (3)) and controls (column (4)) does not alterthe positive association of urban growth with urban natural increase.

In column (5), we include ten region fixed effects (Central Africa, Eastern Africa,Southern Africa, Western Africa, East Asia, South-East Asia, South Asia, Oceania,

12

the Caribbean, Central America, South America, Middle-East and North Africa) in-teracted with a time trend, to control for time-variant heterogeneity at the regionallevel. The effect is then identified by comparing neighboring countries of the sameregion over time. The effect is almost equal to one now (1.01***). This suggeststhat the relationship between urban growth and urban natural increase is additive.A 1 standard deviation increase in the urban natural increase rate leads to a 0.51standard deviation increase in the urban growth rate. Then, if the urban rate ofnatural increase of today’s developing world had been the same on average as inthe developing world of the 19th century (2.3 vs 0.5), its average annual urbangrowth rate would have been 2.1% instead of 3.8% ceteris paribus, and thus almostthe same as in Industrial Europe (2.2%). Likewise, if Africa’s urban rate of naturalincrease had been the same on average as in Asia in 1960-2010 (2.9 vs 1.7), itsaverage annual urban growth rate would have been 3.7% instead of 4.9% ceterisparibus, and thus almost the same as in Asia (3.9%). In column (6), we decomposethe urban rate of natural increase into the urban birth rate and the urban deathrate. Both rates have a strong effect on urban growth (0.98*** and -1.12**).

4.2 Robustness

Robustness. The results of various robustness checks are displayed in Table 2.Column (1) replicates the main result from column (5) of Table 1 (the effect was1.01***). In columns (3)-(7), we add variables estimated in decade t-1 and loseone round of data (N = 132 instead of 165). We thus verify in column (2) that thebaseline effect is unchanged when dropping this round (1.05***). In column (3),we show that the effect remains the same (1.02***) when controlling for residualmigration and urban natural increase in the previous decade. As discussed in theconceptual framework, there are four dynamic relationships that should be consid-ered. We do not find a significant effect of lagged natural increase and migrationon urban natural increase (Unic,t) or on residual migration (Migrc,t) (columns (6)and (7)). The relationship between urban growth and urban natural increase is ad-ditive. In column (4), we include the annual urban growth rate in decade t-1 (thesum of the residual migration and urban natural increase rates in t-1). The maineffect remains the same (1.02***). To control for countries in which urban growthis fast and cities expand spatially leading agglomerations to absorb surrounding ru-ral areas in the next census year, the lag of urban growth rate is added. However, itis insignificant.10 Including more lags give similar results (not shown, but availableupon request), though their inclusion can lead to overfitting given the small numberof observations. In column (5), we control for rural natural increase in decades tand t-1, as urban and rural natural increase could be correlated and influence eachother. Besides, if rural growth is fast where urban growth is fast, because of ru-ral natural increase, urban growth will be disproportionately associated with urbanreclassification. The effect is almost unchanged (1.09***).

External validity. One limitation of the panel analysis is that we only employ datafor 33 countries. For 64 other countries, we found the urban rate of natural increasefor the closest year to 2000. We can run the following cross-sectional regression for(33 + 64 =) 97 countries that were still developing countries in 1960:

10Since we include country fixed effects, we control for the fact that countries use different urbandefinitions, which affect urban reclassification and urban growth. Urban reclassification is only an is-sue if it is correlated with changes in urban natural increase within countries, relative to neighboringcountries of the same region (as we include region fixed effects interacted with a time trend).

13

U grc,1960−2010 = α′+ β ′Unic,2000+ u′c,1960−2010 (8)

where U grc,1960−2010 is the annual urban growth rate (%) of country c from 1960-2010 (i.e., the long difference). Our variable of interest is the urban rate of naturalincrease (per 100 people, or %) of country c in 2000 (Unic,2000). Urban demo-graphic data does not exist for many countries before the 1990s. For the 33 coun-tries for which we have historical data, the coefficient of correlation between theurban rate of natural increase in 2000 and the average of the same rate in 1960-2010 is 0.80. The rate in 2000 can thus be used as a proxy for the rate post-1960.The results are presented in Table 3. The unconditional regression shows a strongeffect of urban natural increase on urban growth (column (1)). This effect is ro-bust to: (i) controlling for income and urbanization in 1960, and income in 2010(column (2)); (ii) adding area fixed effects (column (3)); (iii) including varioustime-invariant controls at the country level (column (4))11; and (iv) adding regionfixed effects (column (5)). The effect in column (5) is lower than 1 (0.76***). Thecross-sectional estimates are less reliable than the panel estimates, as the urban rateof natural increase in 2000 is simply a proxy for the same rate in 1960-2010. Weshould expect the relationship between urban natural increase and urban growth tobe less well-measured as a result, which should lead to a downward bias.

We also focus on the largest city of these countries. We use the same cross-sectionalmodel as in column (5), except the dependent variable is the annual growth rate(%) of the largest city of each country from 1960-2010, and the variable of interestis the birth rate of this city in 2000 (which we use as a proxy for its rate of natu-ral increase in 1960-2010). We could not find data on the death rate. The largestcity’s birth rate has a strong effect on the growth of that city (1.19***, column (6)).The effect is different from 1, but we cannot control for death rates here. There-fore, urban natural increase has accelerated urban growth in developing countries,whether we consider large agglomerations or small and medium-sized cities.

5. RESULTS ON URBANIZATION

In this section, we use econometric regressions and our constructed panel data setfor 33 countries (1960-2010), as well as cross-sectional data for 97 countries (1960-2010), to investigate the effects of urban growth, and urban natural increase andresidual migration in particular, on the change in the urbanization rate.

5.1 Main Results

We use panel data for 33 countries that were still developing countries in 1960. Werun the following model for t = [1960s, 1970s, 1980s, 1990s, 2000s]:

11The controls are the same as in Table 1, except we consider the year 2010 or the period 1960-2010 to estimate the variables, instead of the current decade. The controls are described in thefootnote below Table 3. As we cannot include country fixed effects, we also include various time-invariant controls at the country level. First, if countries with high urban fertility rates systematicallyuse different methods for measuring urbanization, the correlations may reflect measurement error.We get around this issue by adding controls for the different possible definitions of cities in differentcountries: four indicators for each type of definition used by the countries of our sample (admin-istrative, threshold, threshold and administrative, and threshold plus condition) and the value of thepopulation threshold to define a locality as urban when this type of definition is used. Second,we also control for country area (sq km), country population (1000s), a dummy equal to one if thecountry is a small island (< 50,000 sq km) and an indicator equal to one if the country is landlocked,as larger, non-island and landlocked countries could be less urbanized for various reasons.

14

∆U r bratec,t = a+κU grc,t + θc +λt + vc,t (9)

where ∆U r bratec,t is the change in the urbanization rate (in percentage points) ofcountry c in decade t. Our variable of interest is the annual urban growth rate (%).Our hypothesis is that fast urban growth has raised urbanization rates in developingcountries. All regressions include country and decade fixed effects (θc; λt). The re-gressions are the same as when urban growth was the dependent variable. Column(1) of Table 4 shows that fast urban growth is associated with higher urbanizationrates (2.02***). This effect is robust to: (i) controlling for log GDP per capita atthe beginning and the end of the decade (column (2)), which captures the effectsof initial income and income growth on the change in urbanization, (ii) addingcontinent fixed effects interacted with a time trend (column (3)), (iii) including thetime-varying controls at the country level (column (4)), and (iv) adding region fixedeffects interacted with a time trend (column (5)). In the last specification, the effectis identified by comparing neighboring countries of the same region over time. Theeffect shows than a 1 percentage point increase in urban growth leads to a 1.91 per-centage point increase in the urbanization rate every ten years. This effect is large.A 1 standard deviation increase in the urban growth rate is associated with a 0.90standard deviation increase in the urbanization rate. As shown in the conceptualframework, there cannot be urbanization without fast urban growth.

Urban growth comes from residual migration or natural increase. When using thefull specification, we find that the effect of migration is larger than the effect of nat-ural increase (2.02*** vs. 1.21**, column (6))). A 1 standard deviation increase inresidual migration (urban natural increase) is associated with a 0.77 (0.30) stan-dard deviation increase in the change in urbanization. While urban natural increaseis the main factor of urban growth, migration is the main determinant of urbaniza-tion. Recall that a rural migrant has a large effect on urbanization, removing oneresident from the countryside (decreasing the rural population by one) and addingthis resident to the cities (increasing the urban population by one). Therefore, whilemigration (i.e., the rural push and urban pull factors) remains the main driver of ur-banization, urban natural increase has become a component of urbanization.

Since this increase in urbanization is disconnected from income growth, it alsoproduces urbanization without growth. For example, Europe’s urbanization rateincreased from 15% in 1800 to 40% in 1910. Africa and Asia realized the sameperformance in half the time, between 1960 and 2010. Europe’s urbanization ratehas risen by about 2.5 percentage points every ten years during the 1800-1910period. The decadal change was 4.5 percentage points in Africa and Asia post-1960. On average, the urban rate of natural increase was 1.7 percentage pointshigher in Africa and Asia than in Europe. Given an effect of 1.21, this gives adifference of about (1.7 x 1.21 =) 2.1 percentage points of urbanization every tenyears. Urban natural increase thus contributes to explaining why today’s developingworld has urbanized at a much faster pace than the old developing world. It mayalso contribute to explaining why Africa is relatively urbanized for its income level,since it is the region with the highest urban rate of natural increase.

5.2 Robustness

Robustness. The results of various robustness checks are displayed in Table 5. Col-umn (1) replicates the main results from column (6) of Table 4. In columns (3)-(6),we add variables estimated in decade t-1 and lose one round of data. The effect of

15

urban natural increase slightly increases when dropping this round (1.58** ratherthan 1.21**, column (2)). In column (3), we confirm that the relationship is ad-ditive by showing that the natural increase effect remains the same (1.53**) whencontrolling for residual migration and urban natural increase in decade t-1. In col-umn (4), we control for the urban growth rate in decade t-1, which is the sum ofthe residual migration and urban natural increase rates in decade t-1. The maineffect is almost unchanged (1.41**). The effect is high if we control for the changein urbanization in decade t-1 instead (1.24**, column (5)). The lag of the changein urbanization has a small effect (0.24**). In column (6), we control for ruralnatural increase in decades t and t-1. The effect of urban natural increase is higher(1.94**). The results of columns (4)-(6) show that urban reclassification is not amajor issue here, as the results hold when controlling for past urbanization or ruralnatural increase. The effect of migration is high and significant across all specifica-tions (2.02-2.29). The effects are robust to controlling for the initial urbanizationrate in 1960 interacted with decade fixed effects, to control for convergence effectsin urbanization (not shown, but available upon request). Lastly, we test the effectof urban natural increase on the change in urbanization is higher for urbanizationrates close to 50%, as seen in the simulation graph (Figure 3). We interact the ur-ban rate of natural increase with a dummy variable equal to one if the urbanizationrate at the start of the decade was between 30 and 70%. We find that the naturalincrease effect is higher for the observations in this interval (not shown).

External validity. We also run the following cross-sectional regression model for97 countries c that were still developing countries in 1960:

∆U r bratec,1960−2010 = a′+κ′U grc,1960−2010+ v′c,1960−2010 (10)

where ∆U r bratec,1960−2010 is the change in the urbanization rate (in percentagepoints) between 1960 and 2010, and U grc,1960−2010 is the annual urban growth rate(%) of country c from 1960-2010. We use the sample of 97 countries for whichwe know the urban rate of natural increase in 2000. The results are presented inTable 6. The unconditional regression shows a strong effect of urban growth on thechange in urbanization (2.29**, column (1)). This effect increases as we: (i) controlfor income in 1960 and 2010 (column (2)); (ii) add area fixed effects (column (3));(iii) include various controls at the country level (column (4)); and (iv) add regionfixed effects (column (5)). The point estimates are higher in the full specification(5.57***, column (5)), because we correctly control for the other factors of urban-ization. The cross-sectional estimates are more sensitive to the specification thanthe panel estimates, possibly because the panel regressions allowed us to includecountry fixed effects that already captured these factors well. A 1 percentage pointincrease in urban growth leads to a 5.57 percentage point increase in urbanizationover 50 years, or a 1.11 percentage point increase every ten years. By comparison,the panel regressions showed a 1.91 percentage point increase every ten years. Thecross-sectional effect is lower, likely because we estimate the relationship over 50years rather than over 10 years, which should lead to a downward bias if thereare swift changes within countries over time. In column (6), we find that the twosubcomponents of urban growth indeed have a positive effect on the change inurbanization. A 1 percentage point increase in urban natural increase (residual mi-gration) leads to a 3.64 (5.97) percentage point increase in urbanization over 50years, or a 0.73 (1.19) percentage point increase every ten years.

16

6. RESULTS ON URBAN CONGESTIONUrban natural increase has thus accelerated urban growth and urbanization in de-veloping countries, conditional on income. If urban growth is too fast, urban naturalincrease may result in urban congestion. Congestion effects arise from the fact thatthe urban population grows faster than available urban capital. Population growthmay be unexpected, which reduces the stock of capital per capita. Or populationgrowth is expected, but capital cannot be accumulated as fast as the populationgrows. Urban congestion reduces urban welfare, unless rising population densitiesproduce large agglomeration effects, so that the net effects of this fast urban growthare positive. Panel data on the evolution of urban income over time does not exist,so we cannot test this hypothesis. But we can use cross-sectional data on variousmeasures of urban congestion for the most recent period.

6.1 Fast Urban Growth and Slum Expansion

Our main measure of urban congestion is the share of the urban population livingin slums (%) in 2005. We have data for 113 countries that were still developingcountries in 1960. Slum data was recreated using UN-Habitat (2003) and UnitedNations (2013) data. We focus our analysis on 95 countries for which we alsohave data on urban natural increase in 2000. We run the following cross-sectionalregression:

Slumc,2005 = b+φU grc,1960−2010+π4U r bratec,1960−2010+wc,2005 (11)

where Slumc,2005 is the slum variable (%), U grc,1960−2010 is the annual urban growthrate (%) between 1960 and 2010, and 4U r bratec,1960−2010 is the change in theurbanization rate (%) between 1960 and 2010.12 The hypothesis is that countriesin which the urban population grew faster in the past have larger slums today. Moreprecisely, if the urban population doubles every 18 years, the housing stock mustbe doubled every 18 years as well. This implies that agents invest now in orderfor the required housing stock to be available in 18 years. Otherwise, there willbe congestion effects in housing markets. Slum expansion results from fast urbangrowth, whether because migrants flock to the cities, or because urban naturalincrease accelerates urban growth. The change in the urbanization rate should havea lower effect, since what matters for urban congestion is the absolute, rather thanrelative, number of urban residents. There are three caveats to our analysis.

First, we rely on cross-sectional estimates, as data is not available for a sufficientnumber of countries before 2005. Though data collection on slums began in 1990,2005 is the first year in which it was systematic across countries.13 Second, weassume that slum expansion is a good measure of housing congestion. If urbangrowth has been fast in the developing world, urban land expansion has also beenfast (Angel et al., 2010; Seto et al., 2011). In many countries, urban areas grewfaster than urban population, and urban densities decreased. Does that imply thathousing supply increased faster than urban population? On the contrary, the fall inurban densities is a symptom of urban housing shortages. Wealthier cities are char-acterized by high densities, because people work and live in multi-storey buildings.

12The summary statistics of Slumc,2005 are: mean: 49.3; std. dev. 32.3; min: 0; max: 99.4.13Congestion effects should be larger for large agglomerations, as their growth is higher in abso-

lute numbers, for a more constrained urban space. However, we do not have data on congestion forspecific cities, and must use data for all cities instead.

17

In poor countries, the scarcity of multi-storey buildings forces people to move tothe outskirts of their cities. There, people build one-storey shacks, thus producinga continuous decline in urban densities. Slum expansion is the right measure ofper capita housing congestion. Third, we cannot be sure that the effects are causal.The correlation is spurious if urban fertility is higher in poorer countries that havenot completed their fertility transition yet, and if cities in poorer countries havelarger slums. Thus it is important to control for income in all regressions. Even ifwe control for many observable factors such as income, we cannot control for un-observable factors. Congested cities are less functional, which could then preventany adjustment in urban fertility rates for reasons other than low urban incomes.If these reasons are not captured by the controls and the region fixed effects, theeffects will not be causal. Our objective is more modest, in that we want to char-acterize an equilibrium (or trap) where fast urban growth is associated with urbancongestion, no matter whether they reinforce each other.

Main Results. The results are displayed in Table 7. Column (1) shows the uncon-ditional results, while we control for income in 1960 and 2010 in column (2). Incolumns (3) and (4), we also add area fixed effects and the time-invariant controlsat the country level. In column (5), we include region fixed effects. The identifica-tion comes from comparing neighboring countries within a region over time. Thecorrelation between urban growth and slums holds when using the most demand-ing specification (6.43**, column (5)). A 1 standard deviation increase in urbangrowth is associated with a 0.32 standard deviation increase in the share of theurban population living in slums. The change in the urbanization rate has no ef-fect. A 1 standard deviation decrease in the income variables (whose coefficientsare not shown) is then associated with a 0.40 standard deviation increase in theslum share. Thus, while low income explains slum expansion, fast urban growthmay have also contributed to this expansion. Another way to assess the magnitudeof these results is to compare across continents. If the urban growth rate had beenthe same in Africa as in Asia (3.5 instead of 4.9), the slum share would have been10 percentage points lower (given a mean of 49.3% in the sample).

Additional Results. If countries are unable to cope when urban growth is veryfast, we could expect non-linearities in the relationship between slums and urbangrowth. What really matters for slum expansion is the number of years in which anurban population doubles (i.e., the “true” speed of urban growth). An urban pop-ulation doubles in t years if (1+ U gr/100)t = 2. The number of years in which itdoubles is then equal to log(2)/log(1+U gr/100). There is thus a convex, decreas-ing relationship between the true speed of urban growth and the urban growthrate. In column (6), we use the full specification to show that the number of yearsin which an urban population doubles reduces the slum share (-0.5***). For ex-ample, the urban population of today’s developing world doubled every 18 years,compared to every 35 years in Industrial Europe, implying a potential 8.5 percent-age point increase in slum share. In column (6), we investigate whether the effect islarger for countries whose average number of years in which the urban populationdoubles is below the sample mean (about 20 years). The effect for the group ofcountries experiencing fast urban growth (whose population doubles in less than20 years) is twice higher now (-0.6 + -0.7 = -1.3***). The slum share is 6 percent-age points higher in countries where the urban population doubles every 20 yearsrather than every 30 years, and 13 percentage points higher in countries where theurban population doubles every 10 years rather than every 20 years.

18

The two components of urban growth – urban natural increase and residual mi-gration – are then correlated with slum expansion (14.44*** vs. 4.58*, column(8)). The coefficient is higher, and more precisely estimated, for the former thanfor the latter. When standardizing the variables, we find that a 1 standard devi-ation increase in urban natural increase (residual migration) is associated with a0.30 (0.20) standard deviation increase in the slum share. The standardized ef-fect is also lower for migration. One interpretation could be that the type of urbangrowth matters for slum expansion. Natural increase raises the number of childrenand the dependency rate, which lowers income per capita. Migration increases thenumber of adults and reduces the dependency rate, and migrants may be highlymotivated, which increases income per capita. Higher incomes allow householdsand governments to invest more in the quality of the housing stock.14

6.2 Alternative Measures of Urban Congestion

We now focus on alternative measures of urban congestion, for the most recentperiod. This type of urban data does not also exist for earlier decades, and we haveto rely on cross-sectional regressions. We use the full specification, as in column(5) of Table 7. We control for income in 1960 and 2010, and we include the othercontrols and the region fixed effects. The results are displayed in Table 8. For thesake of space, we do not report the coefficient of the change in the urbanizationrate. The effects of a 1 one standard deviation increase in each variable of intereston one standard deviation in the dependent variable are reported in brackets.

Other housing measures: A slum household is defined as a group of individualsliving under the same roof lacking one or more of the following conditions (UN-Habitat, 2003): (i) sufficient-living area, (ii) structural quality, (iii) access to im-proved water source, and (iv) access to improved sanitation facilities. We study thevarious subcomponents of the slum variable. Data is available for a lower numberof countries for some subcomponents, which may reduce the significance of the ef-fects. First, we obtain a positive correlation between urban natural increase and theshare of urban inhabitants who lack sufficient-living area, i.e. who live in dwellingunits with more than 3 persons per room (8.6*, column (1)). The effect is smallerand not significant for migration. Second, there is a negative (but not significant)correlation between urban natural increase and the share of urban inhabitants wholive in a residence with a finished floor, a measure of structural quality (-6.5, column(2)). Third, there is a negative correlation between urban natural increase and theshare of urban inhabitants who have access to an improved water source (-3.5**,column (3)). Migration also has a positive effect (-2.0*). Fourth, the effects aresmall when the dependent variable is the share of urban residents with improvedaccess to sanitation facilities (column (4)). Sanitation facilities are more importantthan other dimensions of housing. A household is considered to have access to im-proved sanitation if an excreta disposal system is available to the household. Givena constrained budget, households and local governments prioritize this dimension

14Another interpretation could be that urban newborns live in slums located in the cities, whilemigrants reside in slums in the periphery. If peripheral slums are not always classified as urban, thisreduces the association between slums and migration. However, this is only an issue if there areseparate slums for newborns and migrants, and if migrants decide to stop exactly at the peripheryof these cities, which may not be credible. Additionally, we examine the correlation between slumstoday and the demographic rates in 2000, which proxy for rates in 1960-2010. If there may havebeen distinct slums when cities were still small in 1960, current agglomerations will likely haveincorporated the periphery-slums now (2010), minimizing these concerns.

19

over other dimensions, which could explain the non-effect. The lack of sufficient-living area and an easy access to improved drinking water may be less essential.Fast urban growth would then be a constraint, as there are too many non-essentialdimensions in which agents must and can act.15

Educational infrastructure: As the population of some cities grew very fast, thenumber of health and educational facilities had to increase rapidly to match the de-mand for human capital. However, the health and education sectors are often highlyregulated in the cities. Governments may have been unable to keep up with thepopulation growth. They needed to invest in new facilities and train and hire newspecialized workers (e.g., physicians and teachers). Rather unfortunately, cross-country data on urban health infrastructure per capita does not exist. Then, sincewe do not have cross-country data on the overcrowding of urban schools, we use asa dependent variable the urban share of 6-15 year-old children that attended schoolin the last year. We use as our main sources of data IPUMS census microdata and theDemographic and Health Surveys that are available for many countries. One issuewith this measure is that it captures both the supply and demand for educational in-frastructure per capita. As we control for income in the regressions, it may capturethe factors driving the demand for education, but we cannot be sure. Urban nat-ural increase is strongly associated with lower attendance rates (-11.8***, column(5)). The effect is lower and not significant for migration. This is logical if naturalincrease disproportionately increases the population share of children.

Transport infrastructure: Unfortunately, we do not have data on road congestionin cities of developing countries today. This type of data is not collected by inter-national organizations, and population censuses and household surveys do not askquestions about how much time people spend commuting on average. We knowthat traffic jams have become a major issue in these cities though (Kutzbach, 2009).For example, UN-Habitat (2008) describes how the outward spreading of Africancities, the lack of efficient public transport and an increase in car ownership ratesall contribute to rising road congestion. Zenou (2011) explains that improving thetransport infrastructure in the cities can increase urban employment. We use par-ticulate matter (PM) concentrations in residential areas of cities with more than100,000 residents in 2000 as a proxy for car pollution and road congestion (WorldBank, 2013). Urban natural increase is indeed positively associated with car pollu-tion (17.18*, column (6)). Urban natural increase is not the only driver of pollution.However, it has contributed to it; a 1 standard deviation increase in urban naturalincrease is associated with a 0.27 standard deviation increase in car pollution. Mi-gration has no effect, possibly because cities that attract migrants are wealthier, andtheir local governments are able to invest in transport infrastructure.

Labor market outcomes: Urban natural increase also results in urban labor supplyshocks. If urban demand does not rise as fast as urban labor supply, the newcomerswill be unemployed, or employed by the urban refugee sectors - low productivitysectors that mostly employ unskilled workers such as “personal and other services”.

15It is interesting to note that urban congestion does not necessarily increase urban mortality indeveloping countries today (see Figure 6). Sewage systems were often inadequate in the cities ofIndustrial Europe. They were a major source of water-borne diseases and urban mortality (Cutler& Miller, 2004; Voigtländer & Voth, 2013a). Sewage systems may be of better quality in today’sdeveloping world, thanks to advances in public health in the last century. The fact that fast urbangrowth does not lead to urban congestion in sanitation in our sample is in line with this hypothesis.

20

We do not have consistent data on urban unemployment or informal employmentas countries often use different definitions, which lead to large variations in urbanunemployment and informality rates. However, as described in the Online DataAppendix, we use IPUMS census microdata, and labor force survey and householdsurvey data to recreate the sectoral composition of urban areas for as many coun-tries as possible around 2000. For each country, we know the urban employmentshares of 11 sectors.16 In column (7), we regress the urban employment share of“personal and other services” on the urban rate of natural increase. It is for ex-ample the least productive non-agricultural sector in the sample of 40 countries ofMcMillan & Rodrik (2011). Its employment share is a good proxy for the absorptivecapacity of labor markets in developing countries. Column (7) shows that urbannatural increase is associated with a higher urban employment share of personalservices (4.00**). A 1 standard deviation increase in the rate of urban natural in-crease is then associated with a 0.49 standard deviation increase in the employmentshare of this refugee sector. The migration effect is high, but not significant.

Overall, it is interesting to note that migration is significantly less associated withurban congestion than urban natural increase ceteris paribus. There are probablyvarious reasons for that, although our analysis can only be speculative without bet-ter data. First, many rural workers migrate to the cities because productivity andincome are rising there. The strong correlation between income and urbanizationin cross-country data suggests that urban income growth must be a strong driver ofmigration (even if rural poverty may also contribute to migration). Second, whilemigration lowers the dependency ratio, urban natural increase increases it. Thislowers incomes in the short run (the time for the urban newborns to enter the labormarket). Third, rising incomes imply that urban residents and governments havethe resources to minimize these urban congestion effects. These channels may ex-plain why urban congestion was less of a problem in Industrial Europe. Boston,London, Manchester and New York were also growing fast in the 19th century, andthese cities were also affected by slum proliferation. However, economic growthwas high, as a result of technological progress that led to industrialization. It isbecause urban incomes were rising that migrants kept moving to these unhealthyurban environments. Lastly, congestion effects were not large enough to offset thegains from agglomeration. Technological progress may be a less important factorin many cities of today’s developing world, as many countries are urbanizing with-out industrializing (Barrios, Bertinelli & Strobl, 2006; Yuki, 2007; Poelhekke, 2010;Gollin, Jedwab & Vollrath, 2013). These countries must cope with the rapid growthof their cities, without capturing the full benefits of agglomeration.

6.3 Policy Implications

The urban developing world grew at 3.8% per year between 1960 and 2010. More-over, growth rates were greater than 5% in many developing cities. If cities oftoday’s developing world grew too fast due to high urban rates of natural increase,and urban congestion reduced urban welfare, what can be done about it? There aretwo possibilities: a reduction in urban fertility or improved urban planning.

16We use data for the closest year to the year 2000, in the 1990-2010 interval. Similarly to Gollin,Jedwab & Vollrath (2013), the 11 sectors are: “agriculture”, “mining”, “public utilities”, “manufac-turing”, “construction”, “wholesale and retail trade, hotels and restaurants”, “transportation, storageand communications”, “finance, insurance, real estate and business services”, “government services”,“education and health” and “personal and other services”.

21

First, any urban population growth slowdown could contribute to increasing theurban capital-labor ratio, and prevent congestion effects from kicking in. This couldbe achieved through a reduction in urban fertility. It is still very high in manydeveloping regions. For example, Africa’s urban birth rate was 36.2 per 1,000 inthe 2000s. This is still higher than the birth rates in Europe during the 19th century(around 30), Asia in 1960 (35), and other developing countries today (10-20).Given an urban death rate of 11.3, reducing the urban birth rate from 36.2 to 20would lead to a natural increase rate of 8.7 (instead of 24.9 now). With a migrationrate of 1.7%, there will be an urban growth rate of 2.6%, similar to industrializingEurope and present-day Asia. Our analysis stresses the role of urban family planningpolicies, as rapid demographic growth also happens in the cities. The urbanizationof the developing world’s population is mechanically driving the “urbanization ofglobal poverty” (Ravallion, 2002; Ravallion, Chen & Sangraula, 2007). The fact thatthe developing world is urbanizing also implies that rapid demographic growth isbecoming increasingly an urban problem.

Second, better urban planning could permit an internalization of negative urban ex-ternalities. The objective for local and national governments would be to minimizeurban congestion, given their minimal fiscal resources. There are several possibleapproaches. First, the remodeling of Paris by Baron Haussmann in the 1850s is aperfect example of the authoritarian approach. He cleared the narrow medievalstreets of the capital in favor of broad boulevards. This transformation increasedthe standard of living of the Parisians in the later period. Though this approach wasundoubtedly beneficial in the long-run, it is highly controversial as a policy model,due to its high societal costs. China may nonetheless be moving in this direction.Second, many cities were planned as a result of (unplanned) creative destruction.For example, many American cities were rebuilt in a better way after a Great Fire(e.g. New York in 1776, Chicago 1871, Boston 1872 and San Francisco 1906). Cityfires in developing countries today are much less destructive, for various reasons.Houses are built with cement and shacks are built with metal sheets, rather thanwood. Fire departments are also more efficient. Third, local urban renewal projectsare examples of a more decentralized approach, whether they are implemented bylocal governments or private promoters. These urban renewal projects may havenet positive effects when well-implemented (Kaufmann & Quigley, 1987; Collins &Shester, 2013). However, in developing countries, the absence of strong privatemarkets (and rent-seeking) may decrease the economic returns to such programs.Thus, without any improvement in urban (and not just national) institutions, ur-ban congestion will remain a major issue. Lastly, congestion effects were probablymore important in large agglomerations. This could explain why migration fromlarge agglomerations to small and medium-sized cities has been observed in Africa(Potts, 2009). One policy could be to remove the constraints on the growth of thenon-primate cities that are often prevalent in developing countries (Christiaensen,Weerdt & Todo, 2013; Christiaensen & Todo, 2013). More generally, it could beworthwhile to invest in the cities of today’s developing world. While urban-biasedpolicies in the past have imposed an unfair burden over the rural residents of thesecountries, many of their cities will keep growing at a fast pace in the future. Whileinvesting in these cities could further fuel migration, not investing in them couldmake things even worse, especially for the next cohorts of urban residents that areborn every year. Alternatively, one may invest in the rural areas of these countriesto slow down excessive migration and relieve the already overcrowded cities.

22

7. CONCLUSION

This paper documents several new facts regarding the processes of urbanization,internal migration, natural increase, and economic development. Using an exten-sive new historical dataset on urbanization and the urban demographic transition,we show that: (i) urban growth has been faster in the developing world of the 20thcentury than in the developing world of the 19th century; (ii) this fast urban growthwas mostly driven by natural increase, and not by migration as in Europe. Manycities of today’s developing world can be classified as “mushroom cities” vs. the“killer cities” of Industrial Europe; fertility remains high, while mortality has fallento low levels, which has led to high urban rates of natural increase; (iii) urbannatural increase has accelerated urbanization in today’s developing world, and thisconditional on income, thus producing urbanization without growth; and (iv) fasturban growth, and urban natural increase in particular, are associated with morecongested cities, which has strong implications for economic development.

Our results make the following contributions. First, our paper adds to the literatureon rural push and urban pull factors by offering an additional mechanism for urbangrowth and urbanization based on a urban push. Urbanization does not come frommigration only, as internal growth also matters. We also hope that the consistentdata set that we have created will help researchers study the urbanization processacross space and time. Second, our paper contributes to the literature on the re-lationship between urbanization and economic development. Our results suggestthat economic development is not the only driver of urban growth and urbaniza-tion. Besides, the resulting urbanization per se may not necessarily be conducive tofurther economic growth and increased welfare, as congestion effects may limit thebenefits from agglomeration. The “origin” of urbanization may thus impact its rela-tionship with development. Third, our findings advance the literature on the effectsof population growth on economic growth. We study an increase in population andcongestion effects from the perspective of cities, not countries.

This paper leaves several open questions. The first is why many countries and citiesdid not complete their fertility transition earlier. Urban fertility remains high in var-ious parts of the world, and their cities will keep growing at a fast pace in the future.A second question that we leave unanswered is why some countries were better ableto reap the benefits from urban agglomeration and solve urban congestion. Whilewe believe that answering these questions is essential for understanding and po-tentially “improving” on the urbanization process of developing countries, they arebeyond the scope of this paper, and we leave them for future research.

23

REFERENCESAcemoglu, Daron, and Simon Johnson. 2007. “Disease and Development: The Effect of

Life Expectancy on Economic Growth.” Journal of Political Economy, 115(6): 925–985.Ades, Alberto F, and Edward L Glaeser. 1995. “Trade and Circuses: Explaining Urban

Giants.” The Quarterly Journal of Economics, 110(1): 195–227.Alvarez-Cuadrado, Francisco, and Markus Poschke. 2011. “Structural Change Out of

Agriculture: Labor Push versus Labor Pull.” American Economic Journal: Macroeconomics,3(3): 127–58.

Angel, Shlomo, Jason Parent, Daniel Civco, Alexander Blei, and David Potere. 2010.“A Planet of Cities: Urban Land Cover Estimates and Projections for All Countries, 2000-2050.” Lincoln Institute of Land Policy Working Paper.

Ashraf, Quamrul, and Oded Galor. 2011. “Dynamics and Stagnation in the MalthusianEpoch.” American Economic Review, 101(5): 2003–41.

Ashraf, Quamrul H., David N. Weil, and Joshua Wilde. 2011. “The Effect of Interventionsto Reduce Fertility on Economic Growth.” National Bureau of Economic Research WorkingPaper 17377.

Bairoch, Paul. 1988. Cities and Economic Development: Frow the Dawn of History to thePresent. Chicago: The University of Chicago Press.

Bandiera, Oriana, Imran Rasul, and Martina Viarengo. 2013. “The Making of ModernAmerica: Migratory Flows in the Age of Mass Migration.” Journal of Development Eco-nomics, 102(C): 23–47.

Barrios, Salvador, Luisito Bertinelli, and Eric Strobl. 2006. “Climatic Change andRural-Urban Migration: The Case of Sub-Saharan Africa.” Journal of Urban Economics,60(3): 357–371.

Bates, Robert. 1981. Markets and States in Tropical Africa: The Political Basis of AgriculturalPolicies. Berkeley: University of California Press.

Becker, Gary S., Edward L. Glaeser, and Kevin M. Murphy. 1999. “Population and Eco-nomic Growth.” American Economic Review, 89(2): 145–149.

Bleakley, Hoyt. 2007. “Disease and Development: Evidence from Hookworm Eradicationin the American South.” The Quarterly Journal of Economics, 122(1): 73–117.

Bleakley, Hoyt. 2010. “Malaria Eradication in the Americas: A Retrospective Analysis ofChildhood Exposure.” American Economic Journal: Applied Economics, 2(2): 1–45.

Bleakley, Hoyt, and Fabian Lange. 2009. “Chronic Disease Burden and the Interaction ofEducation, Fertility, and Growth.” The Review of Economics and Statistics, 91(1): 52–65.

Caselli, Francesco, and Wilbur John Coleman II. 2001. “The U.S. Structural Transfor-mation and Regional Convergence: A Reinterpretation.” Journal of Political Economy,109(3): 584–616.

Christiaensen, Luc, and Yasuyuki Todo. 2013. “Poverty reduction during the rural-urbantransformation : the role of the missing middle.” The World Bank Policy Research Work-ing Paper Series 6445.

Christiaensen, Luc, Joachim Weerdt, and Yasuyuki Todo. 2013. “Urbanization andPoverty Reduction: the Role of Rural Diversification and Secondary Towns.” AgriculturalEconomics, 44(4-5): 435–447.

Clark, Gregory, and Neil Cummins. 2009. “Urbanization, Mortality, and Fertility in Malthu-sian England.” American Economic Review, 99(2): 242–47.

Cohen, Jessica, and William Easterly. 2010. What Works in Development?: Thinking Bigand Thinking Small. Washington D.C.: Brookings Institution Press.

Collins, William J., and Katharine L. Shester. 2013. “Slum Clearance and Urban Renewalin the United States.” American Economic Journal: Applied Economics, 5(1): 239–73.

Cutler, David M., and Grant Miller. 2004. “The Role of Public Health Improvements inHealth Advances: The 20th Century United States.” National Bureau of Economic Re-search, Inc NBER Working Papers 10511.

Cutler, David, Winnie Fung, Michael Kremer, Monica Singhal, and Tom Vogl. 2010.“Early-Life Malaria Exposure and Adult Outcomes: Evidence from Malaria Eradication inIndia.” American Economic Journal: Applied Economics, 2(2): 72–94.

da Mata, D., U. Deichmann, J.V. Henderson, S.V. Lall, and H.G. Wang. 2007. “Determi-nants of city growth in Brazil.” Journal of Urban Economics, 62(2): 252–272.

Davis, James C., and J. Vernon Henderson. 2003. “Evidence on the Political Economy ofthe Urbanization Process.” Journal of Urban Economics, 53(1): 98–125.

Diamond, Jared. 1997. Guns, Germs, and Steel: The Fates of Human Societies. New York: W.W. Norton & Company.

Duranton, Gilles. 2008. “Viewpoint: From Cities to Productivity and Growth in DevelopingCountries.” Canadian Journal of Economics, 41(3): 689–736.

24

Duranton, Gilles. 2013. “Growing through cities in developing countries.” Unpublishedmanuscript, Department of Economics, University of Pennsylvania.

Farré, Lídia, and Francesco Fasani. 2013. “Media Exposure and Internal Migration - Evi-dence from Indonesia.” Journal of Development Economics, 102(C): 48–61.

Fay, Marianne, and Charlotte Opal. 2000. “Urbanization Without Growth: a Not-so-Uncommon Phenomenon.” The World Bank Policy Research Working Paper Series 2412.

Galor, Oded, and David Weil. 1999. “From Malthusian Stagnation to Modern Growth.”American Economic Review, 89(2): 150–154.

Galor, Oded, and David Weil. 2000. “Population, Technology, and Growth: From Malthu-sian Stagnation to the Demographic Transition and Beyond.” American Economic Review,90(4): 806–828.

Glaeser, Edward L., and Joshua D. Gottlieb. 2009. “The Wealth of Cities: AgglomerationEconomies and Spatial Equilibrium in the United States.” Journal of Economic Literature,47(4): 983–1028.

Gollin, Douglas, Remi Jedwab, and Dietrich Vollrath. 2013. “Urbanization with and with-out Industrialization.” Unpublished manuscript, Oxford University Department of Inter-national Development.

Gollin, Douglas, Stephen L. Parente, and Richard Rogerson. 2002. “The Role of Agricul-ture in Development.” American Economic Review, 92(2): 160–164.

Haines, Michael R. 2008. The Population of the United States, 1790-1920. Cambridge: Cam-bridge University Press.

Hajnal, John. 1965. “European Marriage Pattern in Historical Perspective.” In Population inHistory. , ed. D.V. Glass and D.E.C. Eversley. London: Arnold.

Hansen, Gary D., and Edward C. Prescott. 2002. “Malthus to Solow.” American EconomicReview, 92(4): 1205–1217.

Harris, John R., and Michael P. Todaro. 1970. “Migration, Unemployment & Development:A Two-Sector Analysis.” American Economic Review, 60(1): 126–42.

Henderson, J. Vernon. 2010. “Cities and Development.” Journal of Regional Science,50(1): 515–540.

Henderson, Vernon, Adam Storeygard, and Uwe Deichmann. 2013. “Has ClimateChange Promoted Urbanization in Sub-Saharan Africa?” Unpublished manuscript, De-partment of Economics, Brown University.

Henderson, Vernon, Mark Roberts, and Adam Storeygard. 2013. “Is Urbanization inSub-Saharan Africa Different?” Unpublished manuscript, Department of Geography, LSE.

Huillery, Elise. 2009. “History Matters: The Long-Term Impact of Colonial Public Invest-ments in French West Africa.” American Economic Journal: Applied Economics, 1(2): 176–215.

Jedwab, Remi. 2013. “Urbanization without Industrialization: Evidence from ConsumptionCities in Africa.” Unpublished manuscript, Department of Economics, George WashingtonUniversity.

Jedwab, Remi, and Alexander Moradi. 2013. “Transportation Technology and EconomicChange: Evidence from Colonial Railroads and City Growth in Africa.” Unpublishedmanuscript, Department of Economics, George Washington University.

Kaufmann, Daniel, and John M. Quigley. 1987. “The Consumption Benefits of Invest-ment in Infrastructure: The Evaluation of Sites-and-Services Programs in Underdevel-oped Countries.” Journal of Development Economics, 25(2): 263–284.

Keyfitz, Nathan. 1980. “Do Cities Grow by Natural Increase or by Migration?” GeographicalAnalysis, 12: 142–156.

Kremer, Michael. 1993. “Population Growth and Technological Change: One Million B.C.to 1990.” The Quarterly Journal of Economics, 108(3): pp. 681–716.

Kutzbach, Mark J. 2009. “Motorization in Developing Countries: Causes, Consequences,and Effectiveness of Policy Options.” Journal of Urban Economics, 65(2): 154–166.

Lagerlöf, Nils-Petter. 2003. “Mortality and Early Growth in England, France and Sweden.”Scandinavian Journal of Economics, 105(3): 419–440.

Lewis, Arthur. 1954. “Economic Development with Unlimited Supplies of Labour.” TheManchester School, 22(2): 139–191.

Lipton, Michael. 1977. Why Poor People Stay Poor: Urban Bias in World Development. Cam-bridge: Harvard University Press.

Lucas, Robert E. 2004. “Life Earnings and Rural-Urban Migration.” Journal of Political Econ-omy, 112(S1): S29–S59.

Majumdar, Sumon, Anandi Mani, and Sharun W. Mukand. 2004. “Politics, Informationand the Urban Bias.” Journal of Development Economics, 75(1): 137–165.

Matsuyama, Kiminori. 1992. “Agricultural Productivity, Comparative Advantage, and Eco-nomic Growth.” Journal of Economic Theory, 58(2): 317–334.

McKenzie, David, John Gibson, and Steven Stillman. 2013. “A Land of Milk and Honeywith Streets Paved with Gold: Do Emigrants Have Over-Optimistic Expectations about

25

Incomes Abroad?” Journal of Development Economics, 102(C): 116–127.McMillan, Margaret S., and Dani Rodrik. 2011. “Globalization, Structural Change and

Productivity Growth.” NBER Working Papers 17143.McMillan, Margaret S., William A. Masters, and Harounan Kazianga. 2011. “Rural De-

mography, Public Services and Land Rights in Africa: A Village-Level Analysis in BurkinaFaso.” National Bureau of Economic Research, Inc NBER Working Papers 17718.

Moradi, Alexander. 2008. “Confronting Colonial Legacies-Lessons from Human De-velopment in Ghana and Kenya, 1880-2000.” Journal of International Development,20(8): 1107–1121.

Motamed, Mesbah, Raymond Florax, and William Matsers. 2013. “Agriculture, Trans-portation and the Timing of Urbanization: Global Analysis at the Grid Cell Level.” Un-published manuscript, Tufts University, Department of Economics.

Nunn, Nathan, and Nancy Qian. 2011. “The Potato’s Contribution to Population and Ur-banization: Evidence from an Historical Experiment.” The Quaterly Journal of Economics,126(2).

Overman, Henry, and Anthony J. Venables. 2005. “Cities in the Developing World.” Centrefor Economic Performance, LSE CEP Discussion Papers dp0695.

Poelhekke, Steven. 2010. “Urban Growth and Uninsured Rural Risk: Booming Towns inBust Times.” Journal of Development Economics, In Press.

Potts, Deborah. 2009. “The Slowing of sub-Saharan Africa’s Urbanization: Evidence andImplications for Urban Livelihoods.” Enviroment and Urbanization, 21(1): pp. 253–258.

Ravallion, Martin. 2002. “On the urbanization of poverty.” Journal of Development Eco-nomics, 68(2): 435–442.

Ravallion, Martin, Shaohua Chen, and Prem Sangraula. 2007. “New Evidence on theUrbanization of Global Poverty.” Population and Development Review, 33(4): 667–701.

Rogers, Andrei. 1978. “Migration, Urbanization, Resources & Development, and discus-sions.” In Alternatives for Growth: The Engineering and Economics of Natural ResourcesDevelopment. , ed. H. McMains and L. Wilcox, Chapter 9, 147–238. Elsevier.

Rogers, Andrei, and Jeffrey Gale Williamson. 1982. “Migration, Urbanization, andThird World Development: An Overview.” Economic Development and Cultural Change,30(3): 463–82.

Rosenthal, Stuart S., and William C. Strange. 2004. “Evidence on The Nature and Sourcesof Agglomeration Economies.” In Handbook of Regional and Urban Economics. Vol. 4, , ed.J. V. Henderson and J. F. Thisse, Chapter 49, 2119–2171. Elsevier.

Schultz, Theodore W. 1953. The Economic Organization of Agriculture. New York: McGraw-Hill.

Seto, Karen, Michail Fragkias, Burt Güneralp, and Michael Reilly. 2011. “A Meta-Analysisof Global Urban Land Expansion.” PLOS One, 6(8): e23777.

Shifa, Abdulaziz B. 2013. “The Dual Policy in the Dual Economy: The Political Economy ofUrban Bias in Dictatorial Regimes.” Journal of Development Economics, 105(0): 77 – 85.

UN-Habitat. 2003. Slums of the World: The face of urban poverty in the new millennium?New York: United Nations.

UN-Habitat. 2008. The State of the African Cities. Nairobi: United Nations Human Settle-ments Programme.

United Nations. 2013. Milledium Development Goals Indicators: The Official United NationsSite for the MDG Indicators. New York: United Nations.

Voigtländer, Nico, and Hans-Joachim Voth. 2009. “Malthusian Dynamism and the Rise ofEurope: Make War, Not Love.” American Economic Review, 99(2): 248–54.

Voigtländer, Nico, and Hans-Joachim Voth. 2013a. “How the West ’Invented’ FertilityRestriction.” Forthcoming in the American Economic Review.

Voigtländer, Nico, and Hans-Joachim Voth. 2013b. “The Three Horsemen of Riches:Plague, War and Urbanization in Early Modern Europe.” Forthcoming in the Review ofEconomic Studies.

Vollrath, Dietrich. 2011. “The Agricultural Basis of Comparative Development.” Journal ofEconomic Growth, 16(4): 343–370.

Williamson, Jeffrey. 1990. Coping with City Growth During the British Industrial Revolution.Cambridge: Cambridge University Press.

World Bank. 2009. World Development Report 2009: Reshaping Economic Geography. Wash-ington DC: World Bank Publications.

World Bank. 2013. World Development Indicators. Washington D.C.: World Bank.Young, Alwyn. 2005. “The Gift of the Dying: The Tragedy of Aids and the Welfare of Future

African Generations.” The Quarterly Journal of Economics, 120(2): 423–466.Yuki, Kazuhiro. 2007. “Urbanization, Informal Sector, and Development.” Journal of Devel-

opment Economics, 84(1): 76–103.Zenou, Yves. 2011. “Search, Migration, and Urban Land Use: The Case of Transportation

Policies.” Journal of Development Economics, 96(2): 174–187.

26

TABLE 1: URBAN NATURAL INCREASE AND URBAN GROWTH,MULTIVARIATE PANEL ANALYSIS (1960-2010)

Dependent Variable: Annual Urban Growth Rate (%, Decade t)

(1) (2) (3) (4) (5) (6)

Urban Natural Increase Rate 0.95*** 0.84*** 0.91*** 0.97*** 1.01***(Per 100 People, Decade t) (0.28) (0.28) (0.27) (0.30) (0.32)Urban Birth Rate 0.98***(Per 100 People, Decade t) (0.32)Urban Death Rate -1.12**(Per 100 People, Decade t) (0.49)Country FE & Decade FE (33; 5) Y Y Y Y Y YControls for Income & Urb. Rate N Y Y Y Y YArea FE (4) x Time Trend N N Y Y Y YTime-Varying Controls N N N Y Y YRegion FE (10) x Time Trend N N N N Y YObservations (33 x 5) 165 165 165 165 165 165Adj. R-squared 0.70 0.74 0.75 0.76 0.80 0.79Notes: Robust standard errors clustered at the country level are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. The sampleconsists of 33 countries that were still developing countries in 1960, for the following decades: [1960s, 1970s, 1980s, 1990s, 2000s].All regressions include country and decade fixed effects. In column (2), we also control for log GDP per capita (PPP, cst 2005$) and theurbanization rate (%) at the start of the decade, and log GDP per capita at the end of the decade. In column (3), we also include area fixedeffects (Africa, Asia, LAC, MENA) interacted with a time trend. In column (4), we also include the following controls: (i) Rural push factors:average cereal yields (hg per ha), rural density (1000s of rural pop. per sq km of arable area), the number of droughts (per sq km), and adummy equal to one if the country has experienced a conflict, in decade t; (ii) Urban pull factors: the share of manufacturing and services inGDP (%) in 2010 interacted with decade fixed effects, the share of natural resource exports in GDP (%), a dummy equal to one if the countrywas autocratic, and the primacy rate (%), in decade t; and (iii) Population (1000s) in decade t. In columns (5)-(6), we also include regionfixed effects (Western Africa, etc.) interacted with a time trend. See the Online Data Appendix for data sources and construction of variables.

TABLE 2: URBAN NATURAL INCREASE AND URBAN GROWTH,MULTIVARIATE PANEL ANALYSIS (1960-2010), ROBUSTNESS

Dependent Variable: Annual Urban Growth Rate (%, Decade t) Unic,t Migrc,t

(1) (2) (3) (4) (5) (6) (7)

Urban Natural Increase Rate 1.01*** 1.05*** 1.02*** 1.02*** 1.09***(Per 100 People, Decade t) (0.32) (0.37) (0.37) (0.36) (0.37)Residual Migration Rate 0.06(Per 100 People, Decade t-1) (0.08)Urban Natural Increase Rate 0.06(Per 100 People, Decade t-1) (0.31)Annual Urban Growth Rate 0.06 0.06 0.06(Per 100 People, Decade t-1) (0.07) (0.08) (0.06)Annual Urban Growth Rate 0.06 0.17(Per 100 People, Decade t-1) (0.31) (0.10)Rural Natural Increase Rate -0.06(Per 100 People, Decade t) (0.27)Rural Natural Increase Rate -0.00(Per 100 People, Decade t-1) (0.24)Specification Col. (5) Table 1 Y Y Y Y Y Y YObservations (33 x {5; 4; 3}) 165 132 132 132 132 132 132Adj. R-squared 0.78 0.81 0.80 0.81 0.80 0.67 0.84Notes: Robust standard errors clustered at the country level are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. The sampleconsists of 33 countries that were still developing countries in 1960, for the following decades: [1960s, 1970s, 1980s, 1990s, 2000s].In columns (1)-(5), the dependent variable is the annual urban growth rate (%) in decade t. In columns (6) and (7), the dependentvariables are the respective contributions of urban natural increase and “residual migration” to urban growth (%) in decade t. Thebaseline regression (col. (1)) is the same as in column (5) of Table 1. When we add variables estimated in decade t-1, we lose one roundof data. In column (2), we test that the main effect is the same without this round of data. The specification is the same as in column(5) of Table 1. All regressions include country and decade fixed effects, controls for income and urbanization, time-varying controls, andregion fixed effects interacted with a time trend. See the Online Data Appendix for data sources and construction of variables.

27

TABLE 3: URBAN NATURAL INCREASE AND URBAN GROWTH,MULTIVARIATE CROSS-SECTIONAL ANALYSIS (1960-2010)

Dependent Variable: Annual Urban Growth Rate Largest(1960-2010, %) City

(1) (2) (3) (4) (5) (6)

Urban Natural Increase Rate 1.31*** 0.92*** 0.76*** 0.78*** 0.76***(Per 100 People, 2000) (0.19) (0.25) (0.25) (0.25) (0.27)Largest City’s Birth Rate 1.19***(Per 100 People, 2000) (0.37)

Controls for Inc. & Urb. Rate N Y Y Y Y YArea FE (4) N N Y Y Y YTime-Invariant Controls N N N Y Y YRegion FE (13) N N N N Y YObservations 97 97 97 97 97 94R-squared 0.29 0.46 0.54 0.66 0.66 0.64Notes: Robust standard errors are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. The sample consists of 97 countries thatwere still developing countries in 1960. In columns (1)-(5), the dependent variable is the annual urban growth rate (%) between 1960and 2010. In column (6), it is the growth rate of the largest city (%) between 1960 and 2010. The urban natural increase rate in 2000is used as a proxy for urban natural increase in 1960-2010. In column (2), we control for income in 1960 and 2010, and urbanizationin 2010. In column (3), we also include area fixed effects. In column (4), we also add the following controls: (i) Urban definition:four dummies for each type of definition (administrative, threshold, threshold and administrative, and threshold plus condition) and thevalue of the population threshold to define a locality as urban when this definition is used; (ii) Rural push factors: cereal yields in 2010(hg per ha), rural density (1000s of rural pop. per sq km of arable area) in 2010, the number of droughts (per sq km) since 1960,and a dummy equal to one if the country has experienced a conflict since 1960; (iii) Urban pull factors: the share of manufacturingand services in GDP (%) in 2010, the share of natural resource exports in 1960-2010 (%), a dummy equal to one if the country wasmostly autocratic since 1960 and the primacy rate in 2010 (%); and (iv) Other controls: area (sq km), population (1000s) in 2010,and two dummies equal to one if the country is landlocked or a small island (< 50,000 sq km). Columns (5)-(6) also include regionfixed effects (Western Africa, etc.). See the Online Data Appendix for data sources and construction of variables.

TABLE 4: URBAN NATURAL INCREASE AND URBANIZATION,MULTIVARIATE PANEL ANALYSIS (1960-2010)

Dependent Variable: Change in the Urbanization Rate (%, Decade t)

(1) (2) (3) (4) (5) (6)

Annual Urban Growth Rate 2.02*** 1.98*** 2.00*** 2.01*** 1.91***(Per 100 People, Decade t) (0.32) (0.30) (0.30) (0.29) (0.30)Urban Natural Increase Rate 1.21**(Per 100 People, Decade t) (0.60)Residual Migration Rate 2.02***(Per 100 People, Decade t) (0.32)

Country FE & Decade FE (33; 5) Y Y Y Y Y YControls for Income N Y Y Y Y YArea FE (4) x Time Trend N N Y Y Y YTime-Varying Controls N N N Y Y YRegion FE (10) x Time Trend N N N N Y YObservations (33 x 5) 165 165 165 165 165 165Adj. R-squared 0.65 0.66 0.66 0.66 0.68 0.69Notes: Robust standard errors clustered at the country level are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. Thesample consists of 33 countries that were still developing countries in 1960, for the following decades: [1960s, 1970s, 1980s, 1990s,2000s]. All regressions include country and decade fixed effects. In column (2), we also control for log GDP per capita (PPP, cst2005$) at the start and the end of the decade. In column (3), we also include area FE (Africa, Asia, LAC, MENA) interacted with atime trend. In column (4), we also include the same controls as in Table 1 (see the footnote below the table). In column (5)-(6),we also include region FE (Western Africa, etc.) interacted with a time trend. See the Online Data Appendix for data sources andconstruction of variables. 28

TABLE 5: URBAN NATURAL INCREASE AND URBANIZATION,MULTIVARIATE PANEL ANALYSIS (1960-2010), ROBUSTNESS

Dependent Variable: Change in the Urbanization Rate (%, Decade t)

(1) (2) (3) (4) (5) (6)

Urban Natural Increase Rate 1.21** 1.58** 1.53** 1.41** 1.24** 1.94**(Per 100 People, Decade t) (0.60) (0.62) (0.56) (0.62) (0.55) (0.77)Residual Migration Rate 2.02*** 2.22*** 2.20*** 2.22*** 2.29*** 2.24***(Per 100 People, Decade t) (0.32) (0.48) (0.46) (0.47) (0.44) (0.49)Urban Natural Increase Rate -0.79(Per 100 People, Decade t-1) (0.70)Residual Migration Rate 0.43(Per 100 People, Decade t-1) (0.26)Annual Urban Growth Rate 0.31(Per 100 People, Decade t-1) (0.28)Change in the Urbanization Rate 0.24**(Per 100 People, Decade t-1) (0.12)Rural Natural Increase Rate -0.65(Per 100 People, Decade t) (0.60)Rural Natural Increase Rate 0.34(Per 100 People, Decade t-1) (0.74)

Specification Column (5) Table 4 Y Y Y Y Y YObservations (33 x 5) 165 132 132 132 132 132Adj. R-squared 0.69 0.70 0.71 0.70 0.72 0.69Notes: Robust standard errors clustered at the country level are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. The sampleconsists of 33 countries that were still developing countries in 1960, for the following decades: [1960s, 1970s, 1980s, 1990s, 2000s].The baseline regression (col. (1)) is the same as in column (6) of Table 4. When we add variables estimated in decade t-1, we lose oneround of data. In column (2), we test that the main effect is the same without this round of data. The specification is the same as incolumn (6) of Table 4. All regressions include country and decade fixed effects, controls for income, time-varying controls, and regionfixed effects interacted with a time trend. See the Online Data Appendix for data sources and construction of variables.

TABLE 6: URBAN NATURAL INCREASE AND URBANIZATION,MULTIVARIATE CROSS-SECTIONAL ANALYSIS (1960-2010)

Dependent Variable: Change in the Urbanization Rate (%, 1960-2010)

(1) (2) (3) (4) (5) (6)

Annual Urban Growth Rate 2.29** 3.16*** 3.57*** 4.93*** 5.57***(Per 100 People, 1960-2010) (1.02) (1.06) (1.16) (1.15) (1.04)Urban Natural Increase Rate 3.64*(Per 100 People, 2000) (2.02)Residual Migration Rate 5.97***(Per 100 People, 2000) (1.09)Controls for Income N Y Y Y Y YArea FE (4) N N Y Y Y YTime-Invariant Controls N N N Y Y YRegion FE (13) N N N N Y YObservations 97 97 97 97 97 97Adj. R-squared 0.07 0.16 0.20 0.33 0.51 0.50Notes: Robust standard errors are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. The sample consists of 97 countries thatwere still developing countries in 1960. The rate of urban natural increase in 2000 is used as a proxy for urban natural increase in1960-2010. In column (2), we control for log GDP per capita (PPP, cst 2005$) in 1960 and 2010. In column (3), we also include areafixed effects. In column (4), we also add the same country-level controls as in Table 3 (see the footnote below the table). Columns (5)-(6)also include region fixed effects (Western Africa, etc.). See the Online Data Appendix for data sources and construction of variables.

29

TABLE 7: URBAN NATURAL INCREASE, URBAN GROWTH AND SLUMS (2000)MULTIVARIATE CROSS-SECTIONAL ANALYSIS (1960-2010)

Dependent Variable: Urban Population Living in Slums (%, 2005)

(1) (2) (3) (4) (5) (6) (7) (8)

Annual Urban Growth Rate 8.93***4.39***3.83** 6.52***6.43**(%, 1960-2010) (3.12) (1.26) (1.56) (2.27) (2.79)Change in Urbanization Rate -0.18 0.35** 0.35** 0.17 -0.00 0.05 0.06 0.10(%, 1960-2010) (0.29) (0.17) (0.17) (0.21) (0.28) (0.26) (0.22) (0.20)Number of Years in which the Urban Population Doubles -0.5*** -0.6***(Average, 1960-2010) (0.2) (0.2)Number of Years in which the Urban Population Doubles -0.7**(Average 1960-2010) * Dummy “Number of Years Below Mean” (0.3)Urban Natural Increase 14.44***(%, 2000) (5.01)Residual Migration 4.58*(%, 2000) (2.61)

Controls for Income N Y Y Y Y Y Y YArea FE (5) N N Y Y Y Y Y YTime-Invariant Controls N N Y Y Y Y Y YRegion FE (13) N N N N Y Y Y YObservations 95 95 95 95 95 95 95 95Adj. R-squared 0.17 0.61 0.62 0.67 0.68 0.70 0.69 0.70Notes: Robust standard errors are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. The sample consists of 95 countries that werestill developing countries in 1960. The rate of urban natural increase in 2000 is used as a proxy for urban natural increase in 1960-2010. Theresidual migration rate is estimated as the difference between the annual urban growth rate in 1960-2010 and the rate of urban natural increasein 2000. The number of years in which the urban population doubles on average in 1960-2010 is estimated using the annual urban growth rate.We create a dummy variable equal to one if this number is below the mean in the sample (19.4). In column (2), we control for log GDP per capita(PPP, cst 2005$) in 1960 and 2010. In column (3), we also include area fixed effects. In column (4), we also add the same country-level controlsas in Table 3. Columns (5)-(8) also include region fixed effects. See the Online Data Appendix for data sources and construction of variables.

TABLE 8: URBAN NATURAL INCREASE AND MEASURES OF URBAN CONGESTIONMULTIVARIATE CROSS-SECTIONAL ANALYSIS (2000)

Dependent Variable: Lack Finished Access Access School PM10 EmplSufficient Floor Improved Improved Attend. (mg ShareLiving Water Sanitation (6-15 per Perso.Area Source Facilities y.o.) cubic Serv.(%) (%) (%) (%) (%) m) (%)

(1) (2) (3) (4) (5) (6) (7)

Urban Natural Increase 8.6* -6.5 -3.5** -1.2 −11.8*** 17.8* 4.0**(%, 2000) (4.6) (5.6) (1.6) (2.7) (3.9) (10.0) (2.0)

[0.46] [0.20] [0.21] [0.03] [0.49] [0.27] [0.49]Residual Migration 2.9 -1.3 -2.0* -2.0 -3.4 -0.0 1.2(%, 2000) (2.8) (3.6) (1.1) (1.9) (3.0) (5.7) (1.0)

[0.24] [0.07] [0.25] [0.11] [0.22] [0.00] [0.31]Specification Col. (8) Table 8 Y Y Y Y Y Y YSample Mean 18.8 77.9 89.5 65.1 80.2 71.3 5.5Observations 57 66 93 93 64 93 72Notes: Robust standard errors are reported in parentheses; * p<0.10, ** p<0.05, *** p<0.01. The effects of 1 standard deviation increase in thevariable of interest on 1 standard deviation in the dependent variable are reported in brackets. We regress various measures of urban congestionin 2000 on the rates of urban natural increase in 2000, which we use as a proxy for urban natural increase in 1960-2010. The residual migrationrate is estimated as the difference between the annual urban growth rate in 1960-2010 and the rate of urban natural increase in 2000. In column(1), the dependent variable is the share of urban inhabitants who lack sufficient-living area (%), i.e. who live in dwelling units with more than 3persons per room. In column (2), it is the share of urban inhabitants who live in a residence with a finished floor (%). In columns (3) and (4), itis the share of urban inhabitants who have access to an improved water source and improved sanitation facilities respectively (%). In column (5),it is the urban share of 6-15 year-old children that attend school (%). In column (6), it is a measure of particulate matter (PM) concentrations inresidential areas of cities with more than 100,000 residents. In column (7), it is the urban employment share of personal and other services (%),an informal refugee sector. The specification is the same as in column (8) of Table 7. All regressions include controls for income, time-invariantcontrols, and region fixed effects. See the Online Data Appendix for data sources and construction of variables.

Figure 1: Urbanization Rates (%) for Europe (1700-1950)and The Developing World (1900-2010)

Notes: This figure plots the urbanization rate (%) for Europe (1700-1950) and four developingregions (1900-2010): Africa, Asia, Latin America and the Caribbean (LAC) and Middle-East andNorth Africa (MENA). Europe includes 18 Western European countries and the United States, asone example of a Neo-European country. We then use data for 116 African, Asian and non-NorthAmerican countries that were still developing countries in 1960. Averages are estimated using thepopulation weights for the same year. See the Online Data Appendix for data sources.

Figure 2: Annual Urban Growth Rates (%) for Europe (1700-1950)and The Developing World (1900-2010)

Notes: This figure plots the annual urban growth rate (%) for Europe (1700-1950) and four devel-oping regions (1900-2010): Africa, Asia, Latin America and the Caribbean (LAC) and Middle-Eastand North Africa (MENA). Europe includes 18 Western European countries and the United States,as one example of a Neo-European country. We then use data for 116 African, Asian and non-NorthAmerican countries that were still developing countries in 1960. Averages are estimated using thepopulation weights for the same year. See the Online Data Appendix for data sources.

31

Figure 3: Natural Increase and Change in Urbanization, Simulation

Notes: This figure shows the relationship between the change in the urbanization rate in year t(∆Ut , in percentage points) and the urban crude rate of natural increase in year t (Unit , per 100people), given the initial urbanization rate at the start of year t (Ut). We assume that the rural cruderate of natural increase (Rnit) = 2.5% and the residual migration rate (Migt) = 1.5% per year. Weuse Uni = 0.5% as a benchmark. This allows us to compare the “relative” effects of urban naturalincrease on the change in the urbanization rate for various values of Uni = {1; 1.5; 2; 2.5; 3}.

Figure 4: Natural Increase and Urban Growth in England (1700-1950)

Notes: This figure plots the crude birth rate, the crude death rate and the crude rate of naturalincrease (per 1,000 people) for rural England and urban England (1700-1950). This figure alsoplots the decomposition of annual urban growth (%) into annual natural increase (%) and annual“residual migration” (%). See the Online Data Appendix for data sources.

32

Figure 5: Crude Birth Rates for The Developing World (1960-2010)

Notes: This figure plots the crude birth rate (per 1,000 people) for the rural and urban areas of fourdeveloping regions (1960-2010): Africa, Asia, Latin America and the Caribbean (LAC) and Middle-East and North Africa (MENA). We use demographic data that we have collected for 33 countriesthat were still developing countries in 1960. See the Online Data Appendix for data sources.

Figure 6: Crude Death Rates for The Developing World (1960-2010)

Notes: This figure plots the crude death rate (per 1,000 people) for the rural and urban areas of fourdeveloping regions (1960-2010): Africa, Asia, Latin America and the Caribbean (LAC) and Middle-East and North Africa (MENA). We use demographic data that we have collected for 33 countriesthat were still developing countries in 1960. See the Online Data Appendix for data sources.

33

Figure 7: Crude Rates of Natural Increase for The Developing World (1960-2010)

Notes: This figure plots the crude rate of natural increase (per 1,000 people) for the rural and urbanareas of the four developing areas (1960-2010). We use historical demographic data for 33 countriesthat were still developing countries in 1960. See the Online Data Appendix for data sources.

Figure 8: Natural Increase and Urban Growthfor The Two Developing Worlds (1700-1950 and 1960-2010)

Notes: This figure plots the decomposition of annual urban growth (%) into annual natural increase(%) and annual “residual migration” (%) for the four developing areas, the developing world as awhole in 1960-2010 and England in 1700-1950. We use historical demographic data for 33 countriesthat were still developing countries in 1960. See the Online Data Appendix for data sources.

34

FOR ONLINE PUBLICATION: DATA SOURCES

This appendix describes in details the data we use in our analysis.

Spatial Units for Industrial Europe and Today’s Developing World:We use three different samples in our analysis. First, we obtain historical urban data for19 European and North American countries from 1700-1950, and 116 Africa, Asian or non-North American countries that were still developing countries in 1960, from 1960-2010. Weexclude from our analysis the European countries for which we could not find historicallyconsistent urban data, as well as the former CIS countries. We use these countries to de-scribe urban patterns in “Industrial Europe” (which also includes a Neo-European country,the United States) and four developing areas: Sub-Saharan Africa (which we call “Africa”),Asia, Latin America and the Caribbean (LAC), and Middle-East and North Africa (MENA).Second, our main sample consists of 40 of these countries from 1700 to 2010. These arethe countries for which we found historical demographic data. Historical consistent datawas not found for other countries. The list of countries and years (or periods) for whichwe have data is reported in Appendix Tables 1, 2 and 3. These countries belong to the fivedeveloping areas: Industrial Europe (N = 7, about every 40 years in 1700-2010), Africa (N= 10, every ten years in 1960-2010), Asia (N = 11, ditto), LAC (N = 8, ditto) and MENA(N = 4, ditto). Third, we also collect cross-sectional data for 97 out of the 116 countriesfor which we were to able to find demographic data, for the most recent period. The coun-tries of Africa, Asia, the LAC and MENA regions are then classified into 13 regions: CentralAfrica, Eastern Africa and Western Africa for Africa; East Asia, Pacific Islands, South Asiaand South-East Asia for Asia; Caribbean, Central America and South America for the LACregion; and Middle-East and North Africa for the MENA region.

Urban Growth and Urbanization in Industrial Europe:The annual urban growth rate is the average growth rate of the urban population betweentwo years (%). The urbanization rate is defined as the share of the urban population in totalpopulation (%). We use Bairoch (1988) and Malanima and Volckart (2007) to reconstructconsistent urban growth and urbanization rates for 18 Western European countries and theUnited States for the following periods: 1700-1750, 1750-1800, 1800-1850, 1850-1910and 1910-1950. Averages are estimated using the population weights for the same period.We then consider 7 countries in our main analysis (listed in Appendix Table 1). We alsouse Bairoch (1988), Batou and Chevre (1988) and Wikipedia (2013) to obtain the annualgrowth rate of the largest city for the 19 countries for each period.

Urban Growth and Urbanization in Today’s Developing World:We reference Bairoch (1988), Sluglett (2008) and WUP (2011) to reconstruct the urbangrowth and urbanization rates for Africa, Asia and the LAC and MENA regions for the fol-lowing periods: 1900-1920, 1920-1930, 1930-1950, 1950-1960, 1960-1970, 1970-1980,1980-1990, 1990-2000 and 2000-2010. For the last six decades, we use data for 116African, Asian and non-North American countries from 1950-2010. Averages are estimatedusing the population weights for the same period. We consider 33 countries in the panelanalysis from 1960-2010 (listed in Appendix Table 2). We then consider 97 out of the 116countries for the cross-sectional analysis from 1960-2010. We also use WUP (2011) andWB (2013) to estimate the growth rate of the largest city for each country, for the followingperiods: 1960-1970, 1970-1980, 1980-1990, 1990-2000 and 2000-2010.

Urban Demographic Transition in Industrial Europe:For each of the 7 countries of Industrial Europe, we use various historical sources to obtainthe national, urban and rural crude rates of birth, crude rates of death and crude ratesof natural increase (per 1,000 people) for several decades during the 1800-1910 period(sources listed in Panel A, Appendix Table 3). For England, our main European country of

A. 1

analysis, we have data from 1700 to 1950. For the six other countries, demographic dataonly exists for shorter periods.

Urban Demographic Transition in Today’s Developing World:For each of the 33 countries of today’s developing world, we use reports from the Populationand Housing Censuses, CICRED Monographs, Fertility Surveys, and Demographic and HealthSurveys (DHS) as well as the Statistical Yearbooks of the United Nations, to obtain the na-tional, urban and rural crude rates of birth, crude rates of death and crude rates of naturalincrease (per 1,000 people) for each decade during the 1960-2010 period (sources listed inPanel B, Appendix Table 3). We could not find consistent historical data for other countries.Indeed, demographic data does not always exist for countries as far back as the 1960s. For64 other countries of today’s developing world, we use reports from the Population andHousing Censuses and Demographic and Health Surveys to obtain an estimate of the urbanand rural crude rates of birth and death for the closest year to 2000, in the 1990-2010interval. For the 33 + 64 = 97 countries, we also used the same sources to retrieve theurban fertility rate for the closest year to 2000, in the 1990-2010 interval. For 94 countriesof today’s developing world, we also use the sources mentioned above to obtain the birthrate of the largest city for the closest year to 2000, in the 1990-2010 interval. Data on thecrude death rate of the largest city does not exist.

Measures of Urban Congestion:Data on the share of the urban population living in slums (%) comes from UN-Habitat(2003), UN (2013) and WB (2013). A slum household is usually defined as a group ofindividuals living under the same roof lacking one or more of the following conditions (UN-Habitat 2003): (i) sufficient-living area, (ii) structural quality, (iii) access to improved watersource, and (iv) access to improved sanitation facilities. We have data for 113 countries,but we focus on 95 countries for which we also have data on urban natural increase in2000. Data is available for a lower number of countries for some subcomponents of theslum variable. UN-Habitat (2003) reports the share of urban residents that lack “sufficient-living area”, i.e. who live in dwelling units with more than 3 persons per room. We useas a measure of “structural quality” the share of urban inhabitants who live in a residencewith a finished floor. We reconstruct this variable using the International Public-Use Micro-data Series (IPUMS, 2013) and the stat compiler of the Demographic and Health Surveys(DHS, 2013). Data on the share of urban inhabitants who have access to an improved wa-ter source and improved sanitation facilities (%) comes from WB (2013). A household isconsidered to have access to an improved water source if it has sufficient amount of wa-ter for family use, at an affordable price, available to household members without beingsubject to extreme effort, especially to women and children. A household is considered tohave access to improved sanitation, if an excreta disposal system is available to householdmembers. Data on the urban share of 6-15 year-old children that attend school (%) comesfrom the DHS (2013) and IPUMS (2013). Data on our measure of particulate matter (PM)concentrations in residential areas of cities with more than 100,000 residents comes fromWB (2013).

Urban Employment:Data on the urban employment structure in selected countries for 2000-2010 was recreatedusing various sources, as described for each country in Gollin, Jedwab and Vollrath (2013).We use five different sources of data. Our two main data sources are IPUMS (2013), theInternational Public-Use Microdata Series, and ILO (2013), the International Organizationof Labor. We complement these datasets with data from the published reports of Popula-tion and Housing Censuses, Labor Force Surveys and Household Surveys. For each country forwhich data is available, we estimate the employment shares of all urban areas for the follow-ing 11 sectors: “agriculture”, “mining”, “public utilities”, “manufacturing”, “construction”,“wholesale and retail trade, hotels and restaurants”, “transportation, storage and commu-

A. 2

nications”, “finance, insurance, real estate and business services”, “government services”,“education and health” and “personal and other services”.

Income and Other Controls:We have GDP per capita every ten years for 1960-2010. The main variable used in ouranalysis is average log GDP per capita for each decade (constant 2005 international $). Weuse various sources to reconstruct a range of time-invariant or time-varying controls at thecountry-level. In the panel regressions, we include the time-varying controls (estimated inthe same or previous decade). In the cross-sectional regressions, we also include the time-invariant controls (the time-varying controls are estimated for 1960-2010 instead of for thesame or previous decade).

First, we consider various rural push factors: (i) FAO (2013) reports the cereal yields (hg perha) for each country-year observation. We then estimate the average yields for each decade;(ii) Rural density is defined as the ratio of rural population (1000s) to arable area (sq km).The arable area of each country is reported by FAO (2013); (iii) CRED (2013) reports thenumber of droughts experienced by each country every year. We use two variables: thenumber of droughts (per sq km) since 1960, and the number of droughts (per sq km) foreach decade (e.g., 1960-1969 for the 1960s); and (iv) The Polity IV data series includes ameasure of political violence for each country (1964-present). We create an indicator whosevalue is one if the country experienced an interstate or civil conflict in each decade (PolityIV 2013a).

Second, we consider various urban pull factors: (i) The share of manufacturing and servicesin GDP (%) in 2010 is obtained from WB (2013). The data is missing for many country-year observations before the recent period; (ii) We use the data set of Gollin, Jedwab andVollrath (2013) to obtain the average share of natural resource exports in GDP (%) for eachdecade; (iii) We use the Polity IV data series to calculate the average combined polity scorefor each country for each decade (Polity IV 2013b). We create an indicator whose valueis one if the average polity score is lower than -5, the threshold for not being consideredautocratic; and (iv) From WB (2013), we know the share of the largest city in the urbanpopulation, the primacy rate, for all years in 1960-2010.

Third, we use the other following controls: (i) The 97 countries use four different types ofurban definition in their most recent censuses: (a) “administrative cities” are administra-tive centers of territorial units (e.g., provinces, districts, “communes”, etc.), (b) “thresholdcities” are localities whose population is greater than a population threshold of X inhabitants(e.g.,5,000 or 2,500), (c) “administrative or threshold cities” are either administrative cen-ters or localities whose population is greater than a population threshold, and (d) “thresholdwith condition cities” are localities whose population is greater than a population thresholdand who have a large share of the labor force is engaged in non-agricultural activities. Wecreate indicator variables for each definition. For each country using a population threshold,we know the threshold and use it as a control in our regression analysis; (ii) WUP (2011)reports total population for each country every year for 1950-2010; (iii) Country area (sqkm) is obtained from WB (2013); and (iv) We create two indicators whose value is one ifthe country is a small island or if the country is landlocked. We consider an island country“small” if its area is smaller than 50,000 sq km.

REFERENCES

Bairoch, Paul. 1988. Cities and Economic Development: From the Dawn of History to thePresent. Chicago: The University of Chicago Press.

Bairoch, Paul, Batou, Jean, and Pierre Chevre. 1988 Population des villes européennes de 800à 1850 : banque de données et analyse sommaire des résultats. Geneva: Librairie Doz.

A. 3

CRED. 2013. EM-DAT: International Disaster Database Louvain: Centre for Research on theEpidemiology of Disasters (CRED), Université Catholique de Louvain.

DHS. 2013. Demographic and Health Surveys. Washington, DC: USAID.

FAO. 2013. FAOSTAT. Roma: Food and Agriculture Organization.

Gollin, Jedwab and Vollrath (2013). Urbanization with and without Industrialization. Un-published manuscript, Department of International Development, University of Oxford.

ILO. 2013. ILOSTAT Database. Roma: International Labor Organization.

IPUMS. 2013. Integrated Public-Use Microdata Series, International. Minneapolis, MN: Min-nesota Population Center.

Malanima, Paolo, and Oliver Volckart. 2007. Urbanisation 1700-1780. Chapter preparedfor the Third RTN/CEPR Summer Symposium in London: An Economic History of ModernEurope.

Polity IV. 2013a. Polity IV Project: Political Regime Characteristics and Transitions, 1800-2011: Polity IV Annual Time-Series 1800-2011. Vienna: Center for Systemic Peace.

Polity IV. 2013b. Polity IV Project: Political Regime Characteristics and Transitions, 1800-2011: Major Episodes of Political Violence, 1946-2011. Vienna: Center for Systemic Peace.

Sluglett, Peter. 2008. The Urban Social History of the Middle East: 1750-1950. Syracuse:Syracuse University Press.

UN. 2013. Millenium Development Goals Indicators. New York: United Nations.

WB. 2013. World Development Indicators. Washington, DC: World Bank.

Wikipedia. 2013. City of X, in the United States, Economy of Country X, Landlocked country,Largest Cities of Country X and List of Island Countries. San Francisco, CA: Wikipedia.

WUP. 2011. World Urbanization Prospects, the 2011 Revision. New York: United Nations.

A. 4

FOR ONLINE PUBLICATION: APPENDIX FIGURES

Appendix Figure 1: Urban Crude Rates of Birth and Urban Total Fertility Ratesfor 97 Developing Countries (2000-10)

Notes: This figure plots the relationships between the urban crude birth rate (per 1,000 people)and the urban total fertility rate (the average number of children born to an urban woman over herlifetime) for 97 countries that were still developing countries in 1960 and for which we have datafor the period 2000-2010. The linear fit is plotted for the relationship between the urban crude birthrate and the urban total fertility rate. See the Online Data Appendix for data sources.

Appendix Figure 2: Urban Crude Rates of Birth, Death and Natural Increasefor 97 Developing Countries (2000-10)

Notes: This figure plots the relationships between the urban crude rate of natural increase (per1,000 people), the urban crude birth rate (per 1,000 people) and the urban crude death rate (per1,000 people) for the 97 developing countries that were still developing countries in 1960 and forwhich we have data for the period 2000-2010. See the Online Data Appendix for data sources.

A. 5

FOR ONLINE PUBLICATION: APPENDIX TABLES

Appendix Table 1: Decomposition of Annual Urban Growthfor 7 European Countries, 1800-1910

Country Period: 1800-1850 1850-1870 1870-1910 1800-1910

EnglandUrban Growth (%) 2.7 1.8 2.2 2.3Natural Increase (%) 0.0 0.5 1.1 0.5Residual Migration (%) 2.7 1.3 1.1 1.9

BelgiumUrban Growth (%) 1.9 0.3 2.5 1.8Natural Increase (%) _ 0.4 0.6 0.5Residual Migration (%) _ -0.1 1.9 1.3

FranceUrban Growth (%) 1.3 1.0 1.5 1.3Natural Increase (%) _ 0.2 0.1 0.1Residual Migration (%) _ 0.7 1.4 1.2

GermanyUrban Growth (%) 1.8 3.0 3.0 2.5Natural Increase (%) 0.1 0.2 1.0 0.6Residual Migration (%) 1.7 2.8 2.0 1.9

NetherlandsUrban Growth (%) 0.9 0.7 2.1 1.3Natural Increase (%) 0.0 0.5 1.2 0.4Residual Migration (%) 0.9 0.2 0.9 0.9

SwedenUrban Growth (%) 0.8 2.0 3.2 1.9Natural Increase (%) -0.5 0.5 1.0 0.3Residual Migration (%) 1.3 1.5 2.2 1.6

United StatesUrban Growth (%) 5.2 5.7 3.5 4.6Natural Increase (%) 0.3 0.4 0.4 0.4Residual Migration (%) 4.8 5.3 3.1 4.3

AverageUrban Growth (%) 2.1 2.1 2.6 2.2Natural Increase (%) 0.0 0.4 0.7 0.5Residual Migration (%) 2.1 1.7 1.8 1.7

Notes: This table shows the decomposition of annual urban growth into annual natural increase andannual residual migration (%) for 6 European countries and one Neo-European country, the UnitedStates (1800-1910). Averages are not weighted by population. See the Online Data Appendix fordata sources.

A. 6

Appendix Table 2: Decomposition of Annual Urban Growthfor 33 Developing Countries, 1960-2010

Period: 1960-2010 2000-2010

Subregion Country UrbanGrowth

NaturalIncr.

ResidualMigr.

UrbanGrowth

NaturalIncr.

ResidualMigr.

ASIA 3.5 1.7 1.8 2.3 1.1 1.2

East Asia (N = 3): 2.9 1.1 1.8 2.0 0.4 1.6East Asia China 3.7 1.0 2.7 3.8 0.8 3.0East Asia Japan 1.4 0.7 0.7 1.5 0.0 1.5East Asia South Korea 3.6 1.5 2.1 0.9 0.5 0.3South Asia (N = 4): 3.5 1.9 1.6 2.3 1.4 0.9South Asia Bangladesh 5.8 2.1 3.7 3.1 1.1 2.0South Asia India 3.2 1.8 1.4 2.6 1.3 1.3South Asia Pakistan 3.7 2.2 1.5 2.7 1.9 0.8South Asia Sri Lanka 1.3 1.5 -0.2 0.6 1.1 -0.5Southeast Asia (N = 4): 3.9 1.9 2.0 2.5 1.4 1.2Southeast Asia Indonesia 4.5 1.8 2.7 2.9 1.6 1.3Southeast Asia Malaysia 4.6 2.1 2.5 3.5 1.4 2.1Southeast Asia Philippines 3.6 2.4 1.2 2.0 2.1 -0.1Southeast Asia Thailand 3.0 1.4 1.6 1.7 0.4 1.3

LAC 3.1 2.2 0.9 2.1 1.5 0.6

Central America (N = 4): 3.2 2.5 0.7 2.4 1.7 0.7Central America El Salvador 2.7 2.5 0.2 1.3 1.2 0.1Central America Guatemala 3.5 2.8 0.7 3.4 2.8 0.6Central America Mexico 3.1 2.5 0.6 1.7 1.2 0.5Central America Panama 3.5 2.3 1.2 3.0 1.5 1.5South America (N = 4): 3.1 2.0 1.1 1.9 1.3 0.6South America Chile 2.2 1.7 0.5 1.4 1.0 0.4South America Colombia 3.2 1.9 1.3 1.9 1.7 0.3South America Ecuador 3.8 1.9 1.9 2.7 1.1 1.6South America Peru 3.2 2.4 0.8 1.7 1.5 0.2

MENA 3.6 2.6 1.0 2.1 1.6 0.5

Middle-East (N = 2): 4.5 2.8 1.6 2.4 1.8 0.6Middle-East Iran 3.9 2.6 1.3 2.0 1.3 0.7Middle-East Jordan 5.0 3.0 1.9 2.9 2.4 0.4Northern Africa (N = 2): 2.7 2.3 0.4 1.7 1.3 0.4Northern Africa Egypt 2.4 2.2 0.2 2.0 1.7 0.3Northern Africa Tunisia 3.0 2.4 0.6 1.5 0.9 0.5

AFRICA 4.9 2.9 2.1 4.1 2.4 1.7

Eastern Africa (N = 5): 4.9 2.8 2.1 3.7 2.2 1.4Eastern Africa Central Afr. Rep.* 3.5 2.4 1.1 2.1 2.0 0.1Eastern Africa Ethiopia 4.6 2.7 1.9 3.8 2.0 1.7Eastern Africa Kenya 5.7 2.8 2.9 4.4 2.4 2.0Eastern Africa Madagascar 5.1 2.5 2.6 4.7 2.3 2.5Eastern Africa Malawi 5.6 3.7 1.9 3.5 2.6 0.9Western Africa (N = 5): 4.9 2.9 2.0 4.5 2.6 1.9Western Africa Burkina-Faso 6.0 3.0 3.0 6.8 3.1 3.7Western Africa Ghana 4.2 2.5 1.8 4.0 1.8 2.2Western Africa Ivory Coast 5.7 2.7 3.0 3.3 2.4 1.0Western Africa Mali 4.5 3.4 1.1 5.2 3.5 1.7Western Africa Senegal 4.1 2.8 1.4 3.2 2.5 0.8

All Countries 3.8 2.3 1.6 2.8 1.7 1.1

Notes: This table shows the decomposition of annual urban growth into annual natural increase andannual residual migration (%) for 33 developing countries (1960-2010). * The Central African Republicbelongs to Central Africa, but data is missing for other countries of the region. We have included it inEastern Africa. Averages are not weighted by population. See the Online Data Appendix for data sources.

A. 7

Appendix Table 3: Natural Increase Source Information by Country

Panel A: Historical Data for Industrial Europe (1800-1910)

Country Region Years Main Sources

Belgium Europe 1866-1905 Annuaires Statistiques de la Belgique. Belgium. Ministere del’Interieur. Various volumes.

England Europe 1700-1950 Newsholme, A. (1911), The Declining Birth Rate, Its Nationaland International Significance. London: Cassell & CompanyLimited.Friedlander, D. (1969). Demographic Responses and PopulationChange, Demography 6 (4): 359-381.Williamson, J. (1990). Coping with City Growth During theBritish Industrial Revolution. Cambridge: Cambridge UniversityPress.

France Europe 1852-1910 Statistique Annuelle du Mouvement de la Population. France.Statistique Generale. Various volumes.

Germany Europe 1851-1912 Weber, A. (1899). The Growth of Cities in the 19th Century. NewYork: The MacMillan Company.Stedman, T. (1904). Medical Record. New York: William Woodand Company.Pollock, H., and W. Morgan (1913). Modern Cities: Progress ofthe Awakening for Their Betterment Here and in Europe. NewYork: Funk & Wagnalls Company.Holmes, S. (1921). A Study of Present Tendencies in the BiologicalDevelopment of Civilized Mankind. New York: Harcourt, Braceand Company.Vogele, J. (2000). Urbanization and the urban mortality changein Imperial Germany. Health & Place 6: 41-55.

Netherlands Europe 1815-1909 Margaret Sanger (1917). The Case for Birth Control. Modern ArtPrinting Company.Wintle, M. (2004). An Economic and social History of theNetherlands, 1800-1920: Demographic, Economic and SocialTransition. Cambridge: Cambridge University Press.

Sweden Europe 1800-1910 Dyson, T. (2011), The Role of the Demographic Transition in theProcess of Urbanization. Population and Development Review, 37:34-54.

United States Europe 1825-1910 Various Census Reports.Duffy J. (1968). A History of Public Health in New York City,1625-1866. New York: Russell Sage.Rosenwaike, I. (1972). Population History of New York City.Syracuse: Syracuse University Press.Haines, M. (2001). The Urban Mortality Transition in the UnitedStates, 1800-1940. Annales de Demographie Historique 101:33-64.Michael R. Haines, The Population of the United States,1790-1920. Cambridge: Cambridge University Press, 2008.Ferrie, J.P., and W. Troesken (2008). Death and The City:Chicago’s Mortality Transition, 1850-1925. Explorations inEconomic History, 45, 1: 1-16.

A. 8

Appendix Table 3: Natural Increase Source Information by Country

Panel B: Historical Data for Developing Countries (1960-2010)

Country Region Years Main Sources

Bangladesh Asia 1965, 1974, 1985,1991, 2004

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Burkina Faso Africa 1960, 1975, 1985,1996, 2006

Population and Housing Census (Report), Demographicand Health Survey (Report)

Central Afr. Rep. Africa 1960, 1975, 1988,1994-1995, 2003

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report)

Chile LAC 1960, 1970, 1983,1995, 2006

UN Statistical Yearbook, Population and Housing Census(Report), CICRED Monograph

China Asia 1965, 1975, 1985,1995, 2000

UN Statistical Yearbook, Population and Housing Census(Report), CICRED Monograph

Colombia LAC 1965, 1973, 1985,1990, 2000

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Côte d’Ivoire Africa 1965, 1975, 1988,1994, 1999

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report)

Ecuador LAC 1968, 1974, 1985,1993, 2005

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report)

Egypt MENA 1962, 1975, 1985,1996, 2006

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

El Salvador LAC 1965, 1975, 1985,1996, 2006

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report)

Ethiopia Africa 1967, 1974, 1984,1994, 2000

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report)

Ghana Africa 1960, 1970, 1984,1992, 2000

Population and Housing Census (Report), Demographicand Health Survey (Report), CICRED Monograph

Guatemala LAC 1965, 1975, 1980,1992, 1999

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report)

India Asia 1961, 1970, 1985,1989, 2005

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Indonesia Asia 1961, 1975, 1985,1993, 2003

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Iran MENA 1968, 1975, 1986,1990, 2005

UN Statistical Yearbook, Population and Housing Census(Report), CICRED Monograph

Japan Asia 1965, 1975, 1985,1995, 2005

UN Statistical Yearbook, Population and Housing Census(Report), CICRED Monograph

Jordan MENA 1965, 1973, 1990,1997, 2002

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report)

Kenya Africa 1962, 1969, 1979,1989, 1999

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report),CICRED Monograph

A. 9

Country Region Years Main Sources

Madagascar Africa 1965, 1975, 1985,1993, 2000

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report)

Malawi Africa 1970, 1977, 1987,1998, 2008

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report)

Malaysia Asia 1960, 1970, 1980,1990, 2006

UN Statistical Yearbook, Population and Housing Census(Report), CICRED Monograph

Mali Africa 1960, 1976, 1987,1998, 2006

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report)

Mexico LAC 1965, 1974, 1980,1990, 2006

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report)

Pakistan Asia 1968, 1971, 1984,1988, 2000

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Panama LAC 1965, 1969, 1985,1995, 2006

UN Statistical Yearbook, Population and Housing Census(Report), CICRED Monograph

Peru LAC 1960, 1970, 1986,1990, 2000

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report),CICRED Monograph

Philippines Asia 1968, 1978, 1988,1998, 2003

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Senegal Africa 1960, 1976, 1988,1993, 2002

Population and Housing Census (Report), Demographicand Health Survey (Report), Fertility Survey (Report)

South Korea Asia 1960, 1966, 1970,1989, 2006

UN Statistical Yearbook, Population and Housing Census(Report), CICRED Monograph

Sri Lanka Asia 1961, 1971, 1983,1987, 2001

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Thailand Asia 1965, 1975, 1985,1995, 2005

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

Tunisia MENA 1966, 1972, 1980,1989, 2005

UN Statistical Yearbook, Population and Housing Census(Report), Demographic and Health Survey (Report),CICRED Monograph

A. 10

Gender Ratios and Female Labor Market

Outcomes: Evidence from large-scale Mexican

Migration ∗

Emily Conover, Melanie Khamis and Sarah Pearlman

December 6, 2013

1 Long Abstract

The natural ratio of men to women is estimated to be approximately one to

one. However, sex-selective abortion, infanticide, diseases, famines, violence,

wars, incarceration and migration can alter this ratio and lead to missing

men or missing women. The relative scarcity of men or women with respect

to the other gender has important consequences on economic and social out-

comes, such as the marriage and labor markets, and in decisions regarding

investments in human capital and fertility.

The relative scarcity of men, due to wars, violence and incarceration has

effects on female marriage prospects and labor force participation. Women

are less likely to marry and have children, while at the same time out-of-

wedlock child bearing increases (Abramitzky et al. 2011; Charles and Luoh

2010). The evidence on female labor force participation and labor market

outcomes is less clear: Acemoglu et al. (2004) find increases in labor force

participation of women in the US after WWII while Goldin (2001) finds a

lower effect on female employment.

∗Conover: Hamilton College, econover@hamilton.edu; Khamis: Wesleyan University and IZA,mkhamis@wesleyan.edu; Pearlman (corresponding author): Vassar College, sapearlman@vassar.edu .

Changes in gender ratios, henceforth a lower ratio of men to women, and

the subsequent changes in female labor market outcomes such as labor force

participation, types of employment and occupations need to be documented

further, in particular in the context of developing countries.

In this paper, using large-scale migration of Mexican men to the US

as a shock to the gender ratio, we investigate the effects of a lower male-

female gender ratio in the female labor market.1 This migration affected

states and cohorts differently, where working age men were more likely to

migrate abroad. Following Raphael (2013) we exploit the variation over

time and across Mexican states in the gender ratio to identify the effect on

female labor force participation and labor market outcomes. In particular, we

explore whether women were able to break into occupations that traditionally

have had a higher male participation, as fewer men are around to work in

these occupations. Recent literature on gender in the labor market in Mexico

by Juhn et al. (2013, 2014) document changes in occupations for Mexican

females due to trade liberalization, where females increased employment and

wage shares in blue-collar jobs but not in white-collar jobs.

Using data from the Mexican census from 1960 to 2000,2 we first doc-

ument variation in international migration and gender ratios across states.

In a regression of gender ratios on a measure of international migration we

find evidence that the percentage of households that have an international

migrant is negative and significant at the 1 percent level. Thus changes in

the gender ratio are related to the large-scale migration of Mexican males.

Moreover, in the labor market we find a significant and negative relation-

ship between changes in the gender ratio and female self-employment and

white-collar jobs. This implies that as the proportion of men declines, the

proportion of women in self-employment and white-collar jobs increases. Fi-

nally, we create an occupation segregation index, following Beller (1985), and

find some indication of a positive relationship between the changes in gender

ratio and this segregation index. This indicates that as the male-female gen-

1For direct effects of migration on the Mexican labor market see Chiquiar and Hanson2005 and Mishra 2007.

2We have not included 2010 due to the economic crisis, which resulted in return mi-gration.

2

der ratios decline, occupations become less segmented by gender, suggesting

that women may be moving into occupations previously dominated by men

as men become more scarce.

JEL Codes: J16, F22,J21, J24

Keywords: gender ratios, migration, labor force participation, occupations

2 References

Abramitzky, Ran, Delavande, Adeline and Luis Vasconcelos. 2011. ”Marrying

up: The role of sex ratio in assortative mating.” American Economic

Journal: Applied Economics, 3:124-157.

Acemoglu, Daron, David H. Author and David Lyle. 2004. ”Women, War, and

Wages: The Effect of Female Labor Supply on the Wage Structure at

Midcentury.” Journal of Political Economy, 112(3): 497-551.

Amuedo-Dorantes, Catalina and Shoshana Grossbard. 2007. ”Cohort-level sex

ratio effects on women’s labor force participation.”Review of Economics

of the Household, 5(3): 249-278.

Angrist, Joshua. 2002. ”How Do Sex Ratios Affect Marriage and Labor Mar-

kets? Evidence from America’s Second Generation.”The Quarterly

Journal of Economics, 117 (3): 997-1038.

Beller, Andrea. H. 1985. ”Changes in the Sex Composition of U. S. Occupa-

tions, 1960-1981.”The Journal of Human Resources, 20 (2): 235-250.

Charles, Kerwin Kofi and Ming Ching Luoh. 2010. ”Male Incarceration, the

Marriage Market and Female Outcomes.”Review of Economics and

Statistics, 92:614-627.

Chiquiar, David and Gordon H. Hanson. 2005. ”International Migration, Self-

Selection, and the Distribution of Wages: Evidence from Mexico and

the United States.”Journal of Political Economy, 113(2): 239-281.

3

Goldin, Claudia. 1991. ”The Role of World War II in the Rise of Women’s

Employment.” American Economic Review, 81 (4): 741-56.

Juhn, Chinhui, Gergely Ujhelyi and Carolina Villegas-Sanchez. 2013. ”Trade

Liberalization and Gender Inequality.” American Economic Review:

Papers & Proceedings, 103(3): 269-273.

Juhn, Chinhui, Gergely Ujhelyi and Carolina Villegas-Sanchez. 2014. ”Men,

Women, and Machines: how trade impacts gender inequality.” Journal

of Development Economics, 106: 179-193.

Mishra, Prachi. 2007. ”Emigation and wages in source countries: Evidence

from Mexico.” Journal of Development Economics, 82: 180-199.

4

U.S. Border Enforcement and Mexican Immigrant Location Choice∗

Sarah BohnPublic Policy Institute of California

Todd PugatchOregon State University and IZA

February 25, 2014

Abstract

We provide the first evidence on the causal effect of border enforcement on the full spa-tial distribution of Mexican immigrants to the United States. We address the endogeneity ofborder enforcement with an instrumental variables strategy based on administrative delays inbudgetary allocations for border security. We find that 1,000 additional border patrol officersassigned to prevent unauthorized migrants from entering a state decreases that state’s share ofMexican immigrants by 21.9%. Our estimates imply that if border enforcement had not changedfrom 1994-2011, the shares of Mexican immigrants locating in California and Texas would eachbe 8 percentage points greater, with all other states’ shares lower or unchanged.

JEL classification: J15, J61.Keywords: unauthorized immigration, border enforcement, Mexico, residential location choice

∗Author contacts: bohn@ppic.org, todd.pugatch@oregonstate.edu. We thank Scott Borger, Laura Kawano, BryanRoberts, Victor Tremblay, and various seminar participants for helpful input, and Andrew Spaeth for research assis-tance. Pugatch acknowledges support from an Oregon State University Faculty Release Time grant.

1

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

1 Introduction

Since the early 1990s, Mexican immigrants to the United States have increasingly chosen non-

traditional locations, i.e., locations other than those, such as California and Texas, with historically

high Mexican density (Card and Lewis 2007). The reasons for this diffusion of the Mexican migrant

population are complex and varied, but not yet well quantified. A hypothesis advanced by Massey,

Durand and Malone (2002) is that increased border enforcement in traditional migrant crossing

areas has led migrants to choose alternative border crossing routes, and in turn to choose non-

traditional destinations. According to this view, an unintended consequence of strengthened border

enforcement is a change in traditional settlement patterns among Mexican immigrants. In fact,

between 1980 and 2010, the share of Mexican immigrants in California and Texas—the two states

where border enforcement increases were most concentrated—fell from 80 percent to 58 percent.

Of course, enforcement is not the only potential driver of location choice. Economic opportunities,

interior enforcement policies, and social factors are also hypothesized to play a role.

To our knowledge, however, no causal analysis of the effect of border enforcement on the diffusion

of Mexican migrants has been conducted. Indeed, the hypothesis is difficult to evaluate because of

data limitations (crossing locations of Mexican immigrants to the U.S. are not available) and the

endogeneity of border enforcement (the level of enforcement is likely responsive to illegal crossing

behavior). This paper quantifies the causal effect of border enforcement on immigrant location

choice. We overcome the measurement problem by constructing an index that combines data

on enforcement intensity across sectors of the southern border and over time with the historical

destination choice of immigrants, drawing on methods developed in the literature (Pugatch and

Yang 2011, Borger, Hanson and Roberts 2012). We address the endogeneity of the enforcement

index to contemporaneous migration flows by relying on administrative delay in enforcement budget

allocations. Because of this institutional structure, lagged values of our enforcement index provide

identifying variation for the effect of enforcement on immigrant location choice.

We find that increases in border enforcement decreased the share of Mexican immigrants across

1

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

U.S. destinations. Specifically, we find that every 1,000 additional border patrol officers assigned

to prevent unauthorized migrants from entering a state decreases that state’s national share of

Mexican immigrants by 21.9%. These results are stable across subgroups, with slightly stronger

effects for likely or estimated unauthorized population shares, and null effects for immigrants less

likely to be border crossers. Our estimates imply that if border enforcement had not changed from

1994-2011, the shares of Mexican immigrants locating in California and Texas would each be 8

percentage points greater, with all other states’ shares lower or unchanged.

This study is motivated by the change in immigrant settlement patterns depicted in Figure 1 and

the coincident increase in border enforcement displayed in Figure 2. The concentration of Mexican

immigrants in a handful of traditional destinations began to decline in the 1990s, with states in the

Southeast, Great Plains, and Midwest experiencing the fastest growth in Mexican immigration over

the last two decades.1 At the same time, control of the southern U.S. border increased substantially.

As shown in Figure 2, border enforcement increased in intensity concurrently with the falling share

of Mexican immigrants in traditional destinations, prompting Massey et al. (2002) to hypothesize

a causal relationship between them. “The massive buildup of enforcement resources in southern

California, El Paso, and around other ports of entry,” they wrote, “diverted the migratory flows

away from traditional points of destination” (Massey et al. 2002, p. 127).

Gaining a better understanding of the effect of border enforcement on Mexican immigrant

settlement patterns should be of major interest to policymakers. Immigrants play an important

role in equilibrating local labor markets (Borjas 2001, Cadena and Kovak 2013), and their large

share of the workforce has prompted renewed calls for national immigration reform in recent years.

State legislatures have entered the immigration policymaking arena in the absence of federal reform,

and evidence suggests that state policies themselves are driven by rapid inflows of new immigrant

populations (Boushey and Luedtke 2011, Hopkins 2010). Because Mexicans constitute the largest

immigrant group in the United States and have a high propensity to enter the United States without

1These patterns are also amply documented in Card and Lewis (2007) and Singer (2004).

2

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

authorization, their location decisions hold particular importance. Moreover, attempts to thwart

unauthorized immigration come at considerable expense, with the U.S. Customs and Border Patrol

budget for 2012 at nearly $12 billion (Department of Homeland Security 2013). The role that

border enforcement plays in Mexican immigrant locations is thus important at both the national

and local levels.

Before proceeding with exposition of our methodological approach, we briefly place this paper

in the context of two broad literatures—one on impacts of border enforcement and the other re-

lated to immigrant location choice. The influence of border enforcement on aggregate migration

flows is the subject of considerable previous research (see among others Kossoudji 1992, Hanson

and Spilimbergo 1999, Cornelius 2001, Reyes, Johnson and Van Swearingen 2002, Orrenius 2004,

Gathmann 2008, Angelucci 2012). Increases in border enforcement alter migrant crossing locations

(Cornelius 2001, Massey et al. 2002, Sorensen and Carrion-Flores 2007) and increase migration

costs (Orrenius 2004, Roberts, Hanson, Cornwell and Borger 2010). While apprehensions at the

border are apparently correlated with increases in enforcement (Orrenius 2004), it is unclear that

illegal immigration is correlated with enforcement, in part because it is difficult to measure at-

tempted crossing. However, research has indicated that one unintended effect of increased border

enforcement may be to increase the length of stays in the U.S. by discouraging immigrants currently

located in the U.S. from engaging in return and circular migration (Reyes et al. 2002).

While these papers, and many others, have studied migration decisions to and within the U.S,

none (to our knowledge) has evaluated the causal role of border enforcement on the full spatial dis-

tribution of immigrants. The closest antecedents to this study are Pena (2009) and Lessem (2012),

both of which develop models in which border enforcement may influence Mexican immigrant res-

idential locations, rather than just aggregate flows.2 However, Pena’s (2009) analysis is limited

to agricultural workers in four U.S. states, while Lessem’s (2012) sample is limited to returned

migrants from rural Mexican communities that are not nationally representative. Crucially, neither

2Lessem (2012) does not explicitly address the role of border enforcement in immigrant location decisions withinthe U.S., but the structural model she develops could be used for this purpose.

3

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

study accounts for the endogenous response of border policy to migration flows. In contrast, our

paper uses large-scale, nationally representative data on all Mexican immigrants to the U.S., and

isolates plausibly exogenous variation in border enforcement.

Why might border enforcement influence the location of immigrants within the U.S. in addition

to altering the magnitude of overall migration flows? As arguably the most mobile demographic

group in the U.S., immigrants consider several factors when choosing where to reside, including

the presence of others from their home communities (Bartel 1989, Munshi 2003), local employment

opportunities (Cadena 2013a, Cadena 2013b), state immigration policies (Bohn, Lofstrom and

Raphael 2011), and migration costs (Orrenius 1999, Chiquiar and Hanson 2005). The link between

enforcement and location choice is most closely related to the latter. For illustration, suppose

that there is a unique mapping between border crossing locations and U.S. destinations, so that

crossing successfully in a particular location constrains migrants to locate in the associated U.S.

destination. Then changes in enforcement at particular crossing points will alter the relative costs

associated with U.S. destinations, leading marginal migrants to change both their border crossing

and destination.

Of course, in reality migrants may reach any U.S. destination from any border crossing. Nonethe-

less, crossing locations vary in enforcement intensity, direct travel costs to reach a destination,

foregone earnings during travel, and the availability of pre-existing networks to assist with ar-

rival and employment at the destination. If enforcement intensity rises at the crossing location

closest to a migrant’s intended destination, alternate crossing locations become relatively more

attractive. If migration costs to the originally intended destination become sufficiently large, the

migrant’s preference may change to an alternate destination. Alternately, a migrant intended for

the original destination may remain in Mexico, with a migrant willing to reside in a different

destination taking his place, consistent with previous studies that have documented changes in

migrant composition in response to border enforcement (Orrenius and Zavodny 2005, Ibarrraran

and Lubotsky 2007, Angelucci 2012, Lozano and Lopez 2013).

4

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

The propensity for return migration may also change differentially across destinations due to

border enforcement, as circular migrants who anticipate more difficult round trips between the U.S.

and Mexico choose instead to remain in the U.S. (Reyes et al. 2002). The effects on prospective

immigrants and return migrants work in opposite directions, making the role of border enforcement

in immigrant location choice theoretically ambiguous. This multiplicity of channels underscores the

need for a theoretical framework and rigorous empirical analysis. We present a migration choice

model that formalizes this argument and connects it to our empirical analysis in the following

section. Our focus is on consistent estimation of the total effect of border enforcement on the

distribution of the immigrant population across destinations. We leave the question of the spatial

dimension of selection in response to border enforcement to future work.

2 Model and Methodology

Suppose, as in Sjaastad (1962) and Borjas (1987), that a migrant chooses to reside in the location

that offers the highest utility net of migration costs. We adapt their models to a random utility

framework, following closely the exposition of Scanlon, Chernew, McLaughlin and Solon (2002) and

Cadena (2013a) while placing emphasis on the role of border enforcement in the migrant’s location

decision. Conditional on migrating,3 the value function for immigrant i locating in U.S. destination

k in period t is:

Vikt = θekt +Xktβ + εikt (1)

where e is the enforcement intensity associated with locating at the destination in that period, X

is a vector of controls capturing the economic opportunities and other observable characteristics of

a destination relevant to location choice, and ε is the error term. (The controls X do not carry an i

3The model could easily be extended to include the migration decision by specifying the choice to remain in thesource country as the outside option. However, this choice will be unobserved when using U.S. data, so we focus onthe case where a migrant is choosing among locations in the destination country.

5

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

subscript because we will conduct the analysis using destination-level aggregates.) The immigrant

chooses destination k if Vikt ≥ Vijt for all j 6= k. Enforcement affects immigrant location choice by

altering the costs of residing in a destination, as described in the introduction. We formalize this

argument and provide more detail in Appendix A. In addition to altering costs for unauthorized

immigrants, enforcement can also affect the destination choices of authorized immigrants if migrants

from the same source country prefer to live in geographic proximity (Bartel 1989, Munshi 2003).

Although straightforward, several challenges arise immediately in this formulation. First, it is

not obvious how to measure the level of enforcement e faced by potential migrants to a destination

k, particularly for destinations in the interior. Second, even if enforcement can be measured for a

destination, such enforcement is likely endogenous to immigrant location decisions. For instance, if

the government responds to a rapid influx on unauthorized immigrants at a destination by increasing

enforcement, then enforcement intensity e will be correlated with the error term, preventing us from

consistently estimating θ. We address the first of these challenges before returning to a discussion

of how we use (1) as the basis of our empirical specification. We close the section with a description

of an instrumental variables strategy that addresses the second concern.

Consider the problem of measuring enforcement faced by a prospective migrant to destination

k. No large-scale, nationally representative dataset exists that provides information on the cur-

rent U.S. locations of Mexican immigrants and their point of entry. Even if such a dataset were

available, it is not clear that enforcement at the migrant’s point of entry is the proper measure of

enforcement that he or she faced. Migrants have a choice among crossing locations, and could be

influenced by enforcement at alternative locations as well. To address this issue, we build on meth-

ods developed by Pugatch and Yang (2011) and Borger et al. (2012) to construct a new measure of

border enforcement intensity. We combine data on the historical border crossing and destination

patterns of Mexican immigrants to the U.S. with current measures of border policy to assign a

border enforcement index to U.S. locations.

The U.S. Customs and Border Protection (CBP) splits the southern border with Mexico into

6

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

9 sectors, with each sector responsible for preventing unauthorized crossings of people and goods

in its territory. CBP adjusts enforcement intensity across sectors to meet perceived security needs,

leading to variation in enforcement across sectors and over time. This variation will not affect the

desirability of locating in all U.S. destinations equally. Suppose we observed, for example, that

prior to our sample period all migrants to Missouri came from one of two sectors, the Rio Grande

Valley (eastern Texas) and Laredo sectors. Suppose further that 10% of migrants crossing in the

Rio Grande Valley sector located in Missouri, while Missouri’s share of the Laredo sector was 5%.

Then a natural measure of border enforcement intensity for Missouri would be to assign 10% of Rio

Grande Valley and 5% of Laredo’s enforcement to Missouri, with all other sectors contributing zero.

This sector-weighted average of enforcement intensity leads to the following U.S. location-specific

enforcement index:

ekt =9∑

s=1

ωksest (2)

where ωks is the share of immigrants who cross at border sector s who locate in destination k,

and est is enforcement intensity at sector s at time t. We use the number of border patrol agents

(in thousands) as our enforcement measure, so that the index ekt may be interpreted as border

patrol agents assigned to prevent unauthorized immigration to location k at time t. Importantly,

the weights used to construct the index are predetermined with respect to enforcement levels, so

that enforcement patterns do not cause the observed immigrant destination choices. Identifying

variation for the effect of border enforcement on immigrant location choice therefore comes from

three sources: spatial variation in border enforcement across sectors; time series variation in border

enforcement within sectors; and cross-sectional variation in the propensity of immigrants to follow

particular routes from border crossings to U.S. destinations.

Return now to (1), the migrant’s value function for locating in a particular destination. Let

εikt = ηkt +uikt, so that the error may be decomposed into a destination- and time-specific compo-

nent η and an idiosyncratic component u that we assume to be i.i.d. Type I Extreme Value. Then

7

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

the share of immigrants choosing destination k at time t, denoted πkt, may be expressed as:

πkt =exp (θekt +Xktβ + ηkt)∑j exp (θejt +Xjtβ + ηjt)

(3)

Note that this is just the familiar multinomial logit formula with an unobserved destination- and

time-specific component η included. Letting the sample share of immigrants S differ from the

population share π by a multiplicative error ν (assumed uncorrelated with π) and taking logs

yields:

log(Skt) = θekt +Xktβ + ηkt − log(Dt) + νkt (4)

where Dt =∑

j exp (θejt +Xjtβ + ηjt), with the subscript acknowledging that this term is identical

across all destinations at time t. Assume that ηkt may be further decomposed into time-invariant

and time-varying components as ηkt = ζk + φkt. Taking first differences of S yields:

∆ log(Skt) = θ∆ekt + ∆Xktβ −∆ log(Dt) + ∆φkt + ∆νkt (5)

An empirical specification based on this first-differenced equation offers several benefits relative

to multinomial choice estimation. First, it allows for linear estimation with easily interpretable

coefficients; the coefficient of interest θ is the ceteris paribus effect of a one-unit change in enforce-

ment intensity on the percent change in the share of immigrants choosing a destination. Second, the

specification allows for straightforward incorporation of factors common to all destinations within

a time period through the inclusion of period fixed effects, which estimate ∆ log(Dt). Third, the

specification also controls for permanent attributes of a location, such as climate, amenities, and

the role of durable immigrant networks through the term ζ, which differences out of the equation.

A remaining concern, however, is correlation between the destination- and time-specific innova-

tion ∆φkt and changes in enforcement intensity. If border officials respond to shocks that increase

the share of immigrants choosing a location by increasing enforcement intensity, then our estimates

8

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

of θ will be upward biased. We address this issue by instrumenting for ∆ekt with enforcement

lagged two periods. As Borger et al. (2012) note, administrative delays in CBP budget approval

lead to 2-year lags between initial requests and realized outlays. To set its budget, CBP imple-

ments a process known as the “Operational Requirements-Based Budget Program” (ORBBP), in

which border patrol sectors request resources to enforce immigration and customs laws based on

an assessment of current needs.4 This assessment is based on all available information at the time

of the request, including data maintained by CBP on current enforcement levels and apprehensions

of undocumented migrants. ORBBP occurs annually, but the lag between initial requests and

resource allocation exceeds one year.

Although budget allocations determined through ORBBP follow a fairly rigid process, the

Department of Homeland Security may also address unexpected border enforcement needs through a

“surge” of agents or other resources to particular border sectors. Because these additional resources

may be contemporaneously correlated with immigrant flows, we are concerned about inconsistent

estimates obtained through OLS. However, initial budget requests are based on an assessment

of enforcement needs before such unexpected shocks are realized. If these initial requests are

uncorrelated with the change in unobserved factors realized two years later, then the identifying

assumption that ek,t−2 is uncorrelated with ∆φkt will hold. This approach also mirrors one that

has been used in the labor supply literature, as in Ziliak (1997).

The choice of control variables to include in X is also important to isolate the role of bor-

der enforcement from other factors influencing immigrant location choice. We include a host of

destination-specific controls for economic conditions most relevant to prospective immigrants: un-

employment rates, hourly wages, GDP per capita, manufacturing output, agricultural output,

construction output, and new housing permits. The economic sectors are chosen because of the

high concentration of Mexican immigrants employed in these industries. Moreover, including new

housing permits separately from current output helps to capture the role of economic expectations

4We base this section on information learned in discussions with former Department of Homeland Security officials.

9

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

in immigrant location decisions.

We also include measures of state-level legislation aimed at immigrants, which have proliferated

since 2004. Arguably in response to increasing unauthorized immigrant populations and federal

inaction on comprehensive policy reform, state legislatures have enacted hundreds of laws between

2004 and the present. Most immigrant-related state laws are intended to deter employment or

restrict services to unauthorized immigrants, and a few have been shown to be effective deterrents,

at least to immigrant location choice, if not to the law’s stated intent (Bohn et al. 2011). Because

policymakers see both border enforcement and state-level legislation as important deterrents to

unauthorized immigration, including data on this legislation is critical to isolate the role of border

enforcement in immigrant location decisions.

3 Data

To conduct the analysis, we need data on population shares of Mexican immigrants (and other

subpopulations) by U.S. destination; enforcement intensity by border patrol sector; choices of

border crossings and destinations by migrants to construct the weights used in the enforcement

index; and destination-specific control variables. We describe the sources of these data below, with

additional details in Appendix B.

The main source for population data is the U.S. Current Population Survey (CPS), 1994-2011.

We classify immigrants by place of birth, while natives are those born in the United States. We

also combine the 2000 U.S. Census and American Community Survey (ACS) 2001-2011 into an

alternate dataset to check the consistency of the CPS results. We work with state-level aggregates

derived from these sources.5 Relative to the Census/ACS, the CPS provides a longer time series,

including a set of years (1994-1999) with notable fluctuations in border enforcement. These features

5We prefer the U.S. state to other levels of geographic aggregation, such as the metropolitan statistical area (MSA),because there will be fewer state-year cells with zero immigrants than alternative geographic units. Passel and Cohn(2010) cautions against using the CPS and ACS for MSA-level analysis when focusing on unauthorized immigrants.States also leave greater scope to control for changing economic conditions because of greater data availability.

10

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

lead us to prefer the CPS despite the larger sample sizes available in the Census/ACS.

Figure A1 compares the numbers of Mexican immigrants between the Census, ACS, and CPS.

The CPS has notably lower counts of Mexican immigrants than the other sources over most of the

period, with a considerable dip in 2008 likely due to a change in the Census Bureau’s revision to pop-

ulation controls in that year. In particular, the Bureau made sizeable changes to the methodology

for estimating changes in population due to international migration—disproportionately affecting

estimates of foreign-born persons (Passel and Cohn 2010). Although these differences are prob-

lematic for estimates of Mexican immigrant population levels, they are less likely to bias results

for the national share of immigrants at a particular location, our outcome of interest. The lower

immigrant counts in the CPS will bias our current results compared to the Census/ACS only if the

two sources differ because the CPS is differentially correlated with border enforcement. We have

no reason to believe that this is the case, and will present results using both data sources to check

for consistent results. Additionally, excluding data for 2008 from the CPS sample does not alter

our findings (results not shown but available upon request).

Data on border enforcement are from the U.S. Department of Homeland Security (DHS). DHS

reports the number of border patrol agents employed through the Customs and Border Protection

agency annually in each sector of the southern U.S. border. We would prefer to measure border

enforcement using linewatch hours, a more direct measure of enforcement used in several related

studies (Hanson and Spilimbergo 1999, Orrenius 1999, Hanson, Robertson and Spilimbergo 2002,

Orrenius 2004, Gathmann 2008, Angelucci 2012, Lessem 2012), However, DHS stopped reporting

linewatch hours in mid-2004, and denied our repeated Freedom of Information Act requests to

obtain related information that could be used to extend the series. Figure A2 compares agent

counts and linewatch hours for the years in which there is overlap. Although the graph shows

slightly different trends in these series, reflecting higher average annual linewatch hours per agent

in more recent years, the series nonetheless track each other closely. This high correlation comports

with DHS reports indicating the primary activity of border agents is toward linewatch (Department

11

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

of Homeland Security 2002, Simanski and Sapp 2013), and makes us confident that an enforcement

index based on border patrol agents appropriately captures enforcement intensity.

Data on border crossing patterns used to construct weighted enforcement are from the Northern

Border Migration Survey (EMIF), a survey of migrants along the U.S.-Mexico border conducted

by the Mexican government annually since 1993. We use the survey to construct, for each border

patrol sector, the probability of entering each U.S. state. To do so, we assign each survey respondent

to a border sector and a U.S. state according to the crossing point and place of main U.S. residence

on his or her last trip to the U.S. We drop any respondents whose last trip was more than 10 years

prior to the interview date in order to mitigate recall bias. The shares of migrants whose last trip

to the U.S. occurred between 1983-1993 at border crossing s whose last U.S. residence was in state

k are used to construct the crossing probabilities ω that appear in (2).6

Control variables used in the analysis come from various sources, with details in Appendix

B. Data on state-level economic conditions are from U.S. government sources. Data on state-

level legislation aimed at immigrants was compiled from quarterly reports on all state laws related

to immigrants from the National Conference of State Legislatures (NCSL). Our controls include

passage of any deterrent state laws related to employment or enforcement, as of the date a law was

signed by the state governor.

The diffusion of Mexican immigrants to new U.S. destinations, the phenomenon that motivates

our inquiry, is documented in Table 1. The concentration of Mexican immigrants in traditional

destinations—as measured by the shares in the top 5 states, top 10 states, and in California and

Texas—was nearly unchanged between 1980 and 1990. These shares dropped considerably between

1990 and 2000, however, with the share in the top 5 states falling from 90% to 76%. A further

though less precipitous drop occurred between 2000 and 2010. Figure 1 gives a sense of which areas

6Given the availability of data on an immigrant’s crossing location and U.S. destination in the EMIF, one mightreasonably ask why we do not use the EMIF to construct our outcome measures in addition to the enforcementweights. We prefer the CPS (and Census/ACS) for the outcome data because the much larger sample sizes (morethan 1.5 million annually in the CPS compared to around 15,000 in the EMIF) will lead to more accurate measures ofpopulation shares. A similar argument applies to the Mexican Migration Project (MMP), which covers only selectedMexican communities, in addition to its relatively smaller sample.

12

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

absorbed these new migrants, with states in the Southeast, Great Plains, and Midwest experiencing

the fastest growth in Mexican immigration. As Figure 2 shows, over most of this period, declining

concentration of immigrants in traditional destinations was correlated with increased investment

in border enforcement. This correlation, as noted above, prompts the hypothesis that directed

increases in enforcement diverted Mexican immigrant flows away from traditional destinations.

Figure 3 presents data broadly consistent with this story. Panel (a) shows border enforcement in

selected sectors (only a subset are shown for clarity), including a substantial increase in enforcement

in the San Diego border patrol sector in the mid-1990s that leveled off later in the decade. The Rio

Grande Valley (eastern Texas) sector also experienced an increase throughout the period, ending

on a similar level as San Diego. The sharpest increase was in the Tucson sector, however. Panel (b)

shows the share of unauthorized Mexican immigrants crossing at each sector. After remaining flat

for most of the period 1980-1995, San Diego began to lose share beginning the mid-1990s, while Rio

Grande Valley ended the period at a similar level as its historical average. Tucson’s share increased

considerably over the same period as San Diego’s decline. Although the evidence is circumstantial,

the figures do show a clear shift in enforcement and crossing activity from the traditional gateways

on the western and eastern edges of the border towards the center.

This paper seeks to determine if these patterns also led to changes in the residential locations

of Mexican immigrants. If changes in border enforcement during our sample period led immigrants

to change their crossing patterns but not their destinations, then we would expect to see a weaker

link between crossing location and destinations over time. In fact, we observe the opposite. Across

all border sector-U.S. state pairs, the correlation coefficient between the state’s share of migrants

from a crossing location (the weights ω in (2)) and the distance between them is -.29 during 1983-

1993, the period on which our weights are based. In the period 1994-2011, this correlation rose in

magnitude to -.31. If migrants were changing their crossing locations in response to enforcement but

not their destinations conditional on crossing, then border enforcement led to changes in immigrant

locations. Although this simple correlation is not a substitute for a formal analysis, it does suggest

13

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

that our premise matches basic patterns in the data.

The index we use to measure the border enforcement intensity faced by potential migrants to

each U.S. state consists of two components: 1) enforcement intensity by border patrol sector and

2) weights representing the propensity of immigrants crossing at a sector to locate in a particular

U.S. state. We have already presented data on (1). Figure 4 shows data on (2), in the form of maps

showing Mexican immigrant destinations for selected border crossings. Panel (a) shows the loca-

tions chosen from 1983-1993 by migrants crossing in the Rio Grande Valley sector (eastern Texas,

with representative city Brownsville circled). Unsurprisingly, the modal destination is Texas, with

southeastern states also popular. Panels (b) and (c) show the analogous maps for the El Paso

(western Texas and New Mexico) and San Diego sectors. As in panel (a), immigrants crossing in

these sectors choose destinations that are geographically proximate. This variation in U.S. destina-

tions, conditional on border crossing location, allows us to transform the variation in enforcement

across border patrol sectors into state-specific measures of border enforcement intensity.

Figure 5 shows the resulting enforcement index for a representative state, Arizona. The solid line

shows the enforcement index, which may be interpreted as the number of border patrol agents as-

signed to prevent unauthorized immigrants from entering Arizona. Enforcement in the Rio Grande

Valley, San Diego, and Tucson sectors are also plotted. As shown in the graph, the correlation be-

tween Arizona’s enforcement index and enforcement intensity in the Tucson sector is much higher

than that for the other sectors. This is the result we would expect if enforcement in the Tucson

sector is more relevant for potential migrants to Arizona than enforcement in the other sectors.

We close this section by presenting summary statistics in Table 2 on the panel of U.S. states used

in the analysis. The mean Mexican immigrant share is (approximately) 2%, which is a mechanical

result of the sample size of 50 states and the District of Columbia; we omit reporting shares of

other population groups for this reason. The next several rows show average levels of various

subpopulations (sample sizes vary because of state-year cells with zero shares, in accordance with

the sample used in the regression analysis). The average state has 184,480 Mexican immigrants,

14

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

compared to 4.1 million natives. Levels of other subpopulations are mostly as expected, although

our estimate of the unauthorized Mexican immigrant population implies that 34% of Mexican

immigrants are unauthorized, which is low compared to Hanson’s (2006, p. 870) estimate of 56%.

The average level of the enforcement index is 0.21, indicating 210 border patrol agents assigned to

prevent unauthorized immigration to an average state annually. An alternate enforcement index

that replaces border patrol agents with apprehensions of unauthorized migrants in (2) shows 21,200

apprehensions intended for an average state per year. The final rows of the table show border-sector

specific enforcement, measured by number of agents. There is considerable variation across sectors,

with San Diego, Tucson and El Paso assigned the largest numbers of agents.

4 Results

4.1 Main results

In estimating (4), we include in the vector of controls (X) unemployment rates, hourly wages,

(log) GDP per capita, (log) manufacturing output, (log) agricultural output, (log) construction

output, (log) new housing permits, and an indicator for passage of any punitive legislation aimed

at immigrants, all in first differences. Unemployment rates and hourly wages are specific to the

subpopulation whose population shares are under analysis. We also include a constant and year

fixed effects. We cluster standard errors by state.

Before discussing results of estimating (4), we present in Table 3 the results of the first stage, in

which we regress the first difference of the enforcement index on its second lag, with the same set

of controls as described above. In column (1), the coefficient on the instrument is .073, indicating

that every 1,000 border patrol agents assigned to a state two years ago corresponds to an increase

of 73 agents in the past year.7 The coefficient is precisely estimated, with an F -statistic of 74.2.

In column (2) we restrict attention to the years 2000-2011, corresponding to the period of the

7For ease of exposition, the enforcement index based on border patrol agents is specified in absolute numbers ofagents in Table 3, but in thousands in all other results.

15

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Census/ACS sample. The coefficient falls slightly to .068, with an F -statistic of 59.

An alternate instrument that replaces border patrol agents with apprehensions of unauthorized

migrants (in thousands) also produces strong first stage results. Column (3) shows the coefficient for

the enforcement index based on apprehensions is 0.57, representing the additional agents assigned

to a state for every 1,000 apprehensions two years ago. The coefficient using the Census/ACS

sample in column (4) is very similar, and both have F -statistics greater than 25. We prefer the

agent-based instrument for the second stage analysis, however, because it provides a stronger first

stage.

Table 4 presents the main results from estimation of (4), with OLS results in Panel A and IV

results in Panel B.8 Column (1) uses a state’s share of all Mexican immigrants located in the U.S.

as the dependent variable. The OLS coefficient of -.176 indicates that an increase of 1,000 border

patrol agents assigned to a state is correlated with a 17.6% decrease in a state’s share of Mexican

immigrants. In an average state with a 2% share, this would reduce the share to 1.65%. The

corresponding IV coefficient in Panel B is -.219, meaning that a 1-unit increase in the enforcement

index leads to a 21.9% decrease in a state’s Mexican immigrant share. The larger magnitude of the

IV coefficient is as we would expect if OLS coefficients are upwardly biased because enforcement

intensity responds to immigrant inflows. Both coefficients are statistically significant at 1%.9

In subsequent columns of Table 4 we focus on subpopulations of Mexican immigrants to look

for differential responses to border enforcement. In column (2) we restrict attention to males aged

16-50 with a high school education or less, a group with a high propensity to migrate. The IV

coefficient is nearly 1.5 times both the corresponding OLS coefficient and that for all Mexican

8First-stage F -statistics reported in Table 4 do not correspond exactly to those in Table 3, column (1) becauseestimation samples vary due to state-years with a zero population share, for which the log population share isundefined. Cells with a zero share also explain the uneven sample sizes across columns. We check the sensitivity ofresults to exclusion of these observations in Table A2.

9To give a further sense of the magnitudes of our estimates, the average change in the enforcement index is 0.017,representing an annual increase of 17 border patrol agents assigned to a state. Multiplying this figure by our IVestimate of -.219 results in a predicted annual decline of 0.37% in an average state’s Mexican immigrant share. Inthe average state with a 2% Mexican immigrant share, this will result in a decline to 1.99% in one year, or a declineto 1.88% when compounded over the 17 years of our sample.

16

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

immigrants in column (1). The larger magnitude is as expected if this group is more likely to

be affected by border enforcement. In column (3) we investigate the response of unauthorized

immigrants. U.S. government surveys do not ask about immigrants’ legal status. Instead, we use

state-level estimates of unauthorized immigrants from Warren and Warren (2013), multiplied by

the proportion of immigrants who are Mexican (according to the state-year cell of the CPS panel)

to obtain an estimate of a state’s share of unauthorized Mexican immigrants.10 The IV coefficient

is -.324, significant at 1% and considerably larger than the coefficient for all Mexican immigrants,

indicating a greater responsiveness of unauthorized Mexican immigrants to border enforcement, as

we would expect.

Despite the lack of information about immigrant legal status in the data, data on U.S. citizenship

can be used to identify a subgroup of immigrants with certain legal status. These immigrants are

not at risk of deportation and therefore should not respond to border enforcement in the same

manner as non-citizens. Splitting the sample of those born in Mexico into naturalized citizens

and non-citizens in columns (4)-(5), we find that naturalized citizens are not responsive to border

enforcement when deciding where to reside in the U.S., but non-citizens are. The non-citizen

response is of similar magnitude as the unauthorized immigrant group examined in column (3).

Mexican immigration to the U.S. is characterized by high rates of circular migration, with

migrants cycling back and forth between countries with some regularity (Rendon and Cuecuecha

2010). For migrants currently at a U.S. destination, greater enforcement increases the cost of

return migration to Mexico by making it more difficult to engage in circular migration. This

increases the incentives for migrants to remain at their U.S. destination when border enforcement

tightens (Kossoudji 2002, Angelucci 2012). Although U.S. government surveys do not ask directly

about circular migration, they do ask for a respondent’s migration status one year ago. Mexican

immigrants who report being abroad last year were presumably residing in Mexico, and are likely

re-entrants or newly arrived migrants to the U.S., compared to those who report residing in the

10Appendix B provides more detail on the methodology used by Warren and Warren (2013). These data end in2010, leading to fewer observations than the main sample.

17

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

U.S. the previous year. We split Mexican immigrants into groups by their migration status one

year ago in columns (6)-(7). In column (6), the IV coefficient for those residing abroad one year

ago is nearly zero, while the coefficients for those not abroad one year ago are similar to those for

the full sample. The results are consistent with border enforcement leading to postponement of

return migration to Mexico, rather than deterring re-entry or new migration to the U.S.

If migrants are responding to local shocks other than border enforcement, then they may switch

locations within the U.S. rather than change their entry or exit decision. While our specification

controls for many shocks at the state level, the data allow us to further test these responses by

classifying Mexican immigrants as internal migrants if they resided in a different U.S. state one

year ago, and a non-internal migrant otherwise. We expect border enforcement to exert a greater

influence on non-internal migrants. The IV coefficient for non-internal migrants in column (9) is

almost identical to that for the full sample in column (1). In column (8), however, the IV coefficient

for internal migrants is positive but not statistically significant. These findings show that the effect

of border enforcement on location choice is driven by movements across the border, not between U.S.

states. This differential response helps alleviate concerns that the enforcement index is correlated

with a more general, but unobserved, adverse environment for all Mexican immigrants at a location.

In Table 5 we repeat the specification of (4) using additional population groups. In these re-

gressions, we replace controls for the Mexican immigrant unemployment rate and hourly wage with

those for the relevant subpopulation, but all other covariates are unchanged. In columns (1)-(2), we

analyze shares of all non-Mexican immigrants and non-Mexican unauthorized immigrants, respec-

tively, where the latter are constructed by multiplying the Warren and Warren (2013) estimates by

the state’s proportion of immigrants who are non-Mexican. In both cases, the OLS and IV coef-

ficients are statistically indistinguishable from zero, in contrast to our earlier findings for Mexican

immigrants. The results are sensible because unauthorized immigrants from countries other than

Mexico are probably more likely to arrive by air, sea, or through the northern border.

Column (3) shows the response of natives to border enforcement. Although fear of deportation

18

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

should not lead natives to respond to border enforcement, they might nonetheless respond indirectly

through the effect of border enforcement on the location decisions of other groups. The results show

that this is the case, with the IV coefficient positive and significant. This result is consistent with

natives engaging in wage arbitrage as immigrants relocate from high to low enforcement intensity

states. The response is relatively mild, however: the IV coefficient of .037 implies that an increase

of 1,000 border patrol agents assigned to a state increases the native population share by 3.7%, or

from 2% to 2.07% in the average state.

In column (4) we examine the response of Puerto Ricans, who provide a useful falsification test

for our main results because of their linguistic and cultural similarities with Mexicans and their U.S.

citizenship. We find no statistically significant movements of Puerto Ricans in response to border

enforcement, as might be expected. Column (5) shows the response of Central Americans, with

the IV coefficient on the enforcement index of -.201 significant at 5%. This is an interesting result,

suggesting that Central Americans respond to border enforcement in similar fashion as Mexicans,

consistent with anecdotal evidence of relatively large flows of unauthorized Central Americans into

the U.S. through Mexico and the southern U.S. border.

4.2 Robustness checks

In Section 3, we discussed the reasons we preferred the longer panel based on the CPS relative

to the shorter Census/ACS panel. However, the Census/ACS panel provides larger sample sizes

than CPS, and thus is a better source for the years over which the panels overlap. Table A1 shows

results analogous to Table 1 using the Census/ACS panel, which covers the years 2000-2011. The

results are quite similar to those from the CPS panel. In particular, all negative and statistically

significant IV coefficients on the enforcement index from Table 1 are also negative and statistically

significant when using Census/ACS data. The negative coefficient on naturalized citizens is now

significant at the 5% level in column (4), suggesting that border enforcement influences location

choices for Mexican immigrants beyond concerns about legal status. This effect would be consistent

19

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

with immigrant preferences to live with others from their home country (Bartel 1989, Munshi 2003).

In all results presented to this point, state-year observations with a zero population share were

omitted from the analysis, because the log population share is undefined for these cells. We check

whether including these observations leads to different results by adding one person to all state-year

subpopulations and recalculating the population shares, so that the log share is defined for all cells.

The results, presented in Table A2, are similar to Table 1. Although some IV coefficients that are

statistically significant in Table 1 lose significance when including all state-year cells, the signs and

magnitudes are mutually consistent.11

4.3 Discussion: how much does border enforcement matter for location choice?

Given the robustness of our results, it would be instructive to determine the extent to which

U.S. border enforcement accounted for the spatial diffusion of Mexican immigrants during the

sample period. To quantify the effect of border enforcement, we compare actual state shares of

the Mexican immigrant population to those implied by our estimates under a counterfactual of

no change in enforcement. To calculate these counterfactual population shares, we subtract our

baseline estimate of the border enforcement effect (the IV coefficient reported in Table 4, column

[1], multiplied by observed changes in border enforcement during the sample period) from actual

changes in population shares. Details of the calculation appear in Appendix C.

Table 6 presents results of this exercise. Columns (1)-(2) show each state’s observed share of

the Mexican immigrant population at the beginning and end of the sample period. Column (3)

shows the end-period share if border enforcement had not changed over the same period. Taking

the first state in the list, Alabama, as an example, we observe that between 1994 and 2011 its share

11Because our model stems from a multinomial logit formulation, it will be misspecified if the independence ofirrelevant alternatives (IIA) is violated, for instance if nearby states are closer substitutes for a given destinationthan states farther away. To allow for this possibility, we run specifications with enforcement in nearby states asan additional control variable. In one version of this specification, we include the a population-weighted averageenforcement index of neighboring states; in another, we include the average enforcement of all other states, weightedby the inverse distance between state centroids. Results, which are not shown but available upon request, are similarin magnitude to our main estimates, with the coefficients on the enforcement index estimated at -.200 and -.373,respectively, and significant at the 1% level.

20

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

of the Mexican immigrant population rose more than tenfold, from 0.05% to 0.51%. This change

is indicative of the diffusion of Mexican immigrants to southeastern states. In column (3), we see

that our estimates imply that if border enforcement had not changed since 1994, Alabama’s share

would be only 0.26%. The corresponding 0.25 percentage point discrepancy reported in column (4)

indicates that border enforcement played an important role in the increased presence of Mexican

immigrants in Alabama during the sample period.

Similar insights appear throughout the table. Of particular note are our estimates for the

southern border states. We find that Mexican immigrant shares in California and Texas would be

considerably higher if border enforcement had remained static, by more than 8 percentage points

in each case. Conversely, immigrant shares in Arizona and New Mexico would be lower, consistent

with the Massey et al. (2002) hypothesis of enforcement in high-traffic areas of the border leading to

increasing crossing and settlement in border areas with less historical traffic. In fact, our estimates

imply that all states would have a lower (or unchanged) share of Mexican immigrants if enforcement

had not changed, with the exceptions of California and Texas. The maps presented in Figure 6 help

to visualize the results presented in the table. Panel (a) shows the empirical change in Mexican

immigrant shares, while panel (b) presents our estimates from Table 6, column (4). The map shows

that the Mexican immigrant population would not have diffused as extensively across the country

if enforcement had remained unchanged.

Although the estimates in this section stem from an empirical specification derived from a

theory of immigrant location choice, several caveats are in order. First, we ignore the effects of

border enforcement on aggregate flows between Mexico and the United States, and focus only on

the spatial distribution of immigrants across states. Second, the empirical specification embeds

policy and economic variables, such as state-level legislation targeted to immigrants and conditions

in industries with large concentrations of immigrant workers, that would also likely change in

response to any changes in border enforcement. Nonetheless, we think this exercise is instructive

to gauge the relative importance of border enforcement in the diffusion of Mexican immigrants to

21

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

new U.S. destinations in the past two decades.

5 Conclusion

To our knowledge, no causal analysis has been conducted on the impact of border enforcement on

immigrant location choice. The influential Massey et al. (2002) hypothesis is difficult to evaluate

because of data limitations, the presence of competing factors in location choice, and in particular

the endogeneity of border enforcement. We attempt to overcome this latter problem by proposing

an instrumental variables approach. We construct an index of enforcement intensity that varies

across time and place, that, on a lagged basis, we argue provides exogenous variation. Our strategy

controls for a host of location choice factors including state economic conditions and deterrent state

immigration policy.

We find evidence that increases in border enforcement decrease the share of Mexican immigrants,

on the order of a 21.9% decrease in share for every 1,000 additional border patrol officers. These

results are stable across subgroups, with slightly stronger effects for likely or estimated unauthorized

population shares, and null effects for immigrants less likely to be border crossers. Our estimates

imply that California and Texas lost shares of Mexican immigrants to other parts of the country

due to border enforcement, consistent with the hypothesis in Massey et al. (2002).

Understanding the causal effect of border enforcement efforts on immigration location choice

has implications for policy at various levels of government. Major efforts are devoted to controlling

who enters the U.S. across its southern border, with arguably great success. U.S. Secretary of

Homeland Security Janet Napolitano recently testified that the Southwest border has “never been

stronger,” with illegal crossings on the decline (Dinan 2013). However, border policy is presumed to

have less (if any) control over where immigrants settle once they cross the border. These results tell

a different story—we quantify an economically sizeable effect of border enforcement on destination

choice. As such, the results may be useful to policymakers at the federal and state level, concerned

with the size and nature of immigration flows into the U.S. and into individual states.

22

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

References

Angelucci, Manuela, “US Border Enforcement and the Net Flow of Mexican Illegal Migration,”Economic Development and Cultural Change, January 2012, 60 (2), 311–357.

Bartel, Ann P., “Where Do the New U.S. Immigrants Live?,” Journal of Labor Economics, October1989, 7 (4).

Bohn, Sarah, Magnus Lofstrom, and Steven Raphael, “Did the 2007 Legal Arizona Workers ActReduce the State’s Unauthorized Immigrant Population?,” IZA Discussion Paper, April 2011,(5682).

Borger, Scott, Gordon H. Hanson, and Bryan Roberts, “The Decision to Emigrate from Mexico,”2012.

Borjas, George J., “Self-Selection and the Earnings of Immigrants,” The American EconomicReview, September 1987, 77 (4).

, “Does Immigration Grease the Wheels of the Labor Market?,” Brookings Papers on EconomicActivity, January 2001, 2001 (1).

Boushey, Graeme and Adam Luedtke, “Immigrants across the U.S. Federal Laboratory: ExplainingState-Level Innovation in Immigration Policy,” State Politics & Policy Quarterly, 2011, 11 (4),390–414.

Cadena, Brian C., “Native Competition and Low-Skilled Immigrant Inflows,” Journal of HumanResources, September 2013, 48 (4), 910–944.

, “Recent Immigrants as Labor Market Arbitrageurs: Evidence from the Minimum Wage,”July 2013.

and Brian K. Kovak, “Immigrants Equilibrate Local Labor Markets: Evidence from the GreatRecession,” Working Paper 19272, National Bureau of Economic Research August 2013.

Card, David and Ethan G. Lewis, “The Diffusion of Mexican Immigrants during the 1990s: Ex-planations and Impacts,” Mexican Immigration to the United States, 2007, pp. 193–227.

Chiquiar, Daniel and Gordon H. Hanson, “International Migration, Self-Selection, and the Distri-bution of Wages: Evidence from Mexico and the United States,” Journal of Political Economy,April 2005, 113 (2), 239–281.

Cornelius, Wayne A., “Death at the Border: Efficacy and Unintended Consequences of US Immi-gration Control Policy,” Population and Development Review, 2001, 27 (4), 661685.

Department of Homeland Security, “Yearbook of Immigration Statistics,” Technical Report 2002.

, “Budget-In-Brief, Fiscal Year 2013,” Technical Report 2013.

23

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Dinan, Stephen, “Napolitano approves Senates border security plan,” The Washingtion Times,June 2013.

Gathmann, Christina, “Effects of Enforcement on Illegal Markets: Evidence from Migrant Smug-gling along the Southwestern Border,” Journal of Public Economics, October 2008, 92 (10-11),1926–1941.

Hanson, Gordon H., “Illegal Migration from Mexico to the United States,” Journal of EconomicLiterature, December 2006, 44 (4).

and Antonio Spilimbergo, “Illegal Immigration, Border Enforcement, and Relative Wages:Evidence from Apprehensions at the U.S.-Mexico Border,” The American Economic Review,December 1999, 89 (5).

, Raymond Robertson, and Antonio Spilimbergo, “Does Border Enforcement Protect U.S.Workers from Illegal Immigration?,” Review of Economics and Statistics, February 2002, 84(1), 73–92.

Hopkins, Daniel, “Politicized Places: Explaining Where and When Immigrants Provoke LocalOpposition,” American Political Science Review, 2010, 104 (1), 40–60.

Ibarrraran, Pablo and Darren Lubotsky, “Mexican Immigration and Self-Selection: New Evidencefrom the 2000 Mexican Census,” in “Mexican Immigration to the United States,” Chicago, IL:University of Chicago Press, 2007.

Kossoudji, Sherrie A., “Playing Cat and Mouse at the U.S.-Mexican Border,” Demography, May1992, 29 (2), 159–180.

, “Playing Cat and Mouse at the U.S.-Mexican Border,” in Klaus F. Zimmermann and ThomasBauer, eds., The economics of migration. Volume 1. The migration decision and immigrationpolicy, Elgar Reference Collection. International Library of Critical Writings in Economics,vol. 151., 2002, pp. 384–405.

Lessem, Rebecca, “Mexico-US Immigration: Effects of Wages and Border Enforcement,” 2012.

Lozano, Fernando A. and Mary J. Lopez, “Border Enforcement and Selection of Mexican Immi-grants in the United States,” Feminist Economics, January 2013, 19 (1), 76–110.

Massey, Douglas S., Jorge. Durand, and Nolan J. Malone, Beyond smoke and mirrors: Mexicanimmigration in an era of economic integration, New York: Russell Sage Foundation, 2002.

Munshi, Kaivan, “Networks in the Modern Economy: Mexican Migrants in the U. S. LaborMarket,” The Quarterly Journal of Economics, May 2003, 118 (2).

Orrenius, Pia M., “The role of family networks, coyote prices and the rural economy in migrationfrom Western Mexico: 1965-1994,” 1999.

24

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

, “The effect of US border enforcement on the crossing behavior of Mexican migrants,”Crossing the Border: Research from the Mexican Migration Project, 2004, p. 281298.

and Madeline Zavodny, “Self-Selection among Undocumented Immigrants from Mexico,”Journal of Development Economics, October 2005, 78 (1), 215–240.

Passel, Jeffrey S. and DVera Cohn, “U.S. Unauthorized Immigration Flows Are Down SharplySince Mid-Decade,” 2010.

Pena, Anita Alves, “Locational Choices of the Legal and Illegal: The Case of Mexican AgriculturalWorkers in the U.S.1,” International Migration Review, 2009, 43 (4), 850880.

Pugatch, Todd and Dean Yang, “The Impact of Mexican Immigration on US Natives: Evidencefrom Migrant Flows Driven by Rainfall Shocks,” Mimeo. University of Michigan., 2011.

Rendon, Silvio and Alfredo Cuecuecha, “International Job Search: Mexicans In and Out of theUS,” Review of Economics of the Household, March 2010, 8 (1), 53–82.

Reyes, Belinda, Hans Johnson, and Richard Van Swearingen, “Holding the Line? The Effect ofRecent Border Build-up on Unauthorized Immigration,” Public Policy Institute of CaliforniaReport, 2002.

Roberts, Bryan, Gordon H. Hanson, Derekh Cornwell, and Scott Borger, “An Analysis of MigrantSmuggling Costs along the Southwest Border,” DHS Office of Immigration Statistics WorkingPaper, November 2010.

Scanlon, Dennis P., Michael Chernew, Catherine McLaughlin, and Gary Solon, “The impact ofhealth plan report cards on managed care enrollment,” Journal of Health Economics, January2002, 21 (1), 19–41.

Simanski, John and Lesley Sapp, “Immigration Enforcement Actions: 2012,” Technical Report,Department of Homeland Security Office of Immigration Statistics December 2013.

Singer, Audrey, “The Rise of New Immigrant Gateways,” 2004.

Sjaastad, Larry A., “The Costs and Returns of Human Migration,” Journal of Political Economy,October 1962, 70 (5).

Sorensen, Todd and Carmen Carrion-Flores, “The Effects of Border Enforcement on Migrants’Border Crossing Choices: Diversion or Deterrence?,” 2007.

Warren, Robert and John Robert Warren, “Unauthorized Immigration to the United States: An-nual Estimates and Components of Change, by State, 1990 to 2010,” International MigrationReview, 2013, 47 (2), 296329.

Ziliak, James P., “Efficient Estimation With Panel Data When Instruments Are Predetermined:An Empirical Comparison of Moment-Condition Estimators,” Journal of Business & EconomicStatistics, 1997, 15 (4), 419–431.

25

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

A Border enforcement and location choice model

We adapt the pioneering work of Sjaastad (1962) and Borjas (1987), and assume that a prospectivemigrant from Mexico will relocate to the United States if his utility net of migration costs in theU.S. exceeds his utility in Mexico, such that the following inequality holds:12

uUS − c(e, d) > uMX (6)

where uUS is utility in the U.S., uMX is utility in Mexico, and c(·) is the cost of migrating to the U.S.from Mexico. Assuming the migrant cannot cross the border legally, the cost is comprised of bor-der enforcement e and the distance d to the choice of location, and is increasing in each argument.13

The model is easily extended to multiple potential destinations in the U.S. Initially assume thatthere is a unique mapping between border crossing locations and destinations, so that migrantscrossing successfully in a particular location are constrained to locate in the associated U.S. desti-nation. The migrant will choose U.S. destination k over U.S. destination l if the following inequalityholds (and the left-hand side is non-negative):

uk − uMX − c(ek, dk) > ul − uMX − c(el, dl) (7)

where enforcement is made destination-specific in the sense developed in the paper. It followsimmediately that changes in enforcement at particular crossing points will alter the relative costsassociated with U.S. destinations, leading marginal migrants to change their locations.

The effect operates through two channels, as shown in Sorensen and Carrion-Flores (2007): deter-rence from any migration to the U.S., and diversion of some migrants from one crossing location(and hence destination) to another. Both effects will increase the share of migrants locating at thedestination with decreased relative enforcement.

Now consider the more realistic setting in which a potential migrant to a destination may chooseamong multiple crossing locations. Crossing locations vary in their associated smuggling fees andprobabilities of apprehension. For each destination, crossing locations also vary in their distancecost, where “distance” is defined broadly to include direct travel costs, ease of travel throughpre-existing networks associated with the destination, foregone earnings during travel time, andprobability of apprehension in the interior. For each U.S. destination k, the value of migration is:

Vk = uk −mins

[c(es, dsk)] (8)

where s indexes border crossing sector and dsk is the distance cost from crossing s to destination

12We thank Scott Borger for discussions that helped develop this section.13Enforcement affects smuggling fees f and the probability of apprehension p. Although not all illegal migrants

choose to pay for smuggling services, in general they will face a trade-off between paying for smuggling or facing ahigher probability of apprehension, as in Gathmann (2008). We abstract from this decision and assume that themigrant chooses the cost-minimizing combination of (f, p), summarized as e.

26

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

k. A migrant will choose destination k if Vk > uMX and Vk > Vl for all l 6= k. As in the case of asingle crossing location per destination, with multiple crossing locations changes in enforcement atborder sector s can lead to changes in migrant destinations. Because the cost-minimizing crossinglocation can vary across destinations, crossing sector-specific changes in migration costs can alter amigrant’s ranking of destinations. For instance, suppose San-Diego/Tijuana is the preferred cross-ing location for migrants to California (CA). Increased enforcement at this sector may decrease VCA

sufficiently to switch a migrant’s preferred destination from California to an alternate destination,such as Arizona or New Mexico.

Moreover, if the shape of the value function varies in the population of potential migrants, as islikely, then the composition of migrants may also change. Relatively risk-averse migrants may bedeterred from attempting to locate in traditional U.S. destinations as enforcement increases, whilemore risk-tolerant migrants will be induced to settle in non-traditional destinations where the eco-nomic and social environment is less familiar.

Specifying the right-hand side of (8) as a linear function of border enforcement, observable char-acteristics, and an idiosyncratic error, and adding individual and time subscripts i and t, leads to(1), our point of departure in Section 2.

B Detail on data sources

Population data sources used to construct outcome variables are the U.S. Current Population Sur-vey (CPS) 1994-2011, U.S. Census 2000, and American Community Survey (ACS) 2001-2011. Ineach case, microdata maintained by the Minnesota Population Center’s Integrated Public Use Mi-crodata Series (IPUMS) are aggregated by U.S. state (including the District of Columbia) and year.The 2000 Census are a 5% random sample of the full Census. CPS data for 2011 are from Marchonly, as this was the only month available at the time the analysis was conducted; all other yearsinclude CPS data from all months.

Estimates of the unauthorized immigrant population by state-year are from Warren and Warren(2013). Warren and Warren (2013) use a residual method to estimate unauthorized populationby state for each year from 1990 to 2010. This method, common to the migration and demog-raphy literature, compares the total number of immigrants based on survey data (in this case,ACS and Decennial Census) to the total number of authorized immigrants based on administrativeinformation (DHS records of legal permanent and non-immigrant residents). The Warren and War-ren (2013) estimates are the most detailed state-year estimates of unauthorized immigrants to date.

In constructing the weights ω used to create the enforcement index in (2), we use the NorthernBorder Migration Survey (EMIF) modules on migrants to the U.S. and voluntary returnees fromthe U.S. to Mexico. Because the weights span those whose last trip to the U.S. was between 1983-1993 and we drop any respondents whose last trip was more than 10 years prior to the interviewdate, we use EMIF waves conducted between 1993-2003.

27

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Control variables used in the analysis come from various sources. Unemployment rates, hourlywages, and annual income come from the IPUMS versions of the U.S. Current Population Survey1994-2011, U.S. Census 2000, and American Community Survey 2001-2011, in concordance withthe dataset used in the analysis. Data on gross domestic product (total, agricultural, and manu-facturing) are from the U.S. Bureau of Economic Analysis. Data on new housing permits are fromthe U.S. Census Bureau Building Permits Survey. Data on state legislation regarding immigrantsare from the National Conference of State Legislatures Immigration Policy Project. Since 2005, theNCSL has collected information on laws and resolutions enacted by state legislatures to addressimmigrant-related issues. NCSL uses a comprehensive search to identify all immigrant-related laws,both deterrent and attractive, as well as those signed or vetoed by the governor. Our policy controlvariables account for the passage of any deterrent laws, signed by the governor and related onlyto employment and enforcement. Similar laws have been shown to have significant impacts onthe location choice of immigrants, as in (Bohn et al. 2011). Additional laws, for example thoseconstricting immigrant access to public services, may also be relevant to immigrant location choice.

C Calculation of counterfactual population shares

We use our estimates of the effect of border enforcement to calculate state shares of Mexicanimmigrants that would prevail if enforcement had not changed over the sample period. To do so,we subtract our estimated border enforcement effect from each state’s (log) population share overthe sample period:

log(SkT )− log(Sk0) =T∑t=0

(∆ log(Skt)− θ∆ekt

)(9)

where t = 0 and t = T denote the beginning and end of the sample, θ is the estimated IV coefficienton the enforcement index, and all other notation is as in (5). We solve for SkT by rewriting theleft-hand side of (9) and performing some algebraic manipulations:

exp

[

log

(SkTSk0

)]× Sk0 = SkT (10)

Finally, we normalize SkT so that the shares sum to 1.

28

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Table 1: Shares of Mexican Immigrants

Year Share in top 5 states Share in top 10 states Share in California and Texas

1980 0.92 0.96 0.801990 0.90 0.95 0.792000 0.76 0.86 0.632010 0.71 0.82 0.58

Table shows shares of Mexican immigrants in state groupings by year. Sources: U.S. Census for 1980-2000, American CommunitySurvey for 2010.

29

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Table 2: Summary statistics

Variable N Mean S.D.Mexican immigrantsshare 840 0.02 0.07all 840 194,931 631,422males 16-50 810 73,869 217,852unauthorized 790 65,322 203,200naturalized citizen 790 47,152 164,748not naturalized citizen 816 154,978 481,895abroad last year 220 1,674 3,951not abroad last year 840 194,396 630,026internal migrant 280 834 2,058not internal migrant 839 194,814 631,187Other population groupsnon-Mexican immigrant 867 466,837 941,395non-Mexican unauthorized 816 126,225 255,302natives 867 4,876,839 4,898,778Puerto Rican 743 29,911 70,619Central American 836 47,433 121,534Enforcementindex (agents) 840 0.21 0.75index (apprehensions) 840 21.2 76.9weights (state-sector pairs) 423 0.02 0.10border patrol agents (sector-years)

all sector-years 153 1,177 810Rio Grande Valley (TX) 17 1,491 615Laredo (TX) 17 1,035 496Del Rio (TX) 17 963 421Big Bend (TX) 17 298 194El Paso (TX & NM) 17 1,529 747Tucson (AZ) 17 2,048 1,065Yuma (AZ) 17 492 303El Centro (CA) 17 682 337San Diego (CA) 17 2,058 365

Table shows summary statistics from U.S. state-years 1995-2011 (unless otherwise indicated). Population data fromCurrent Population Survey. Unauthorized Mexican/non-Mexican immigrants are estimates from Warren and Warren(2013), multiplied by the proportion of Mexican/non-Mexican immigrants in the state-year cell. Enforcement index =∑

s Pr(US destination|cross at border sector s) × enforcement at sector s, where enforcement is thousands of border patrolagents or apprehensions of unauthorized migrants. Index may be interpreted as amount of enforcement dedicated to preventingarrival of unauthorized migrants at destination. Enforcement weights (crossing probabilities) calculated from EMIF crossings1983-1993. Border patrol agents and apprehensions from Department of Homeland Security.

30

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Table 3: First Stage

enforcement index, first difference

(1) (2) (3) (4)

agent index, t− 2 0.073 0.068(0.009)*** (0.009)***

apprehension index, t− 2 0.57 0.60(0.095)*** (0.120)***

Observations 863 612 914 612R-squared 0.38 0.37 0.39 0.331st stage F -statistic 74.2 59.0 36.3 25.4data source CPS ACS CPS ACS

Table shows regressions of first difference of enforcement index on its second lag. Sample is U.S. state-years (including D.C.) fromCurrent Population Survey, 1995-2011 in column (1), U.S. Census 2000 and American Community Survey 2001-2011 in column(2). Enforcement index =

∑s Pr(US destination|cross at border sector s)× enforcement at sector s. Index may be interpreted

as amount of enforcement dedicated to preventing arrival of unauthorized migrants at destination, where enforcement is numberof border patrol agents or apprehensions of unauthorized migrants (thousands), as indicated. Crossing probabilities calculatedfrom EMIF crossings 1983-1993. Border patrol agents from Department of Homeland Security. All regressions include yearfixed effects and the following controls (in first differences): Mexican immigrant unemployment rate, Mexican immigrant hourlywage, log GDP per capita, log agricultural GDP, log manufacturing GDP, log construction GDP, log new housing permits,and a dummy for passage of any punitive immigration legislation. Robust standard errors in parenthesis, clustered by state. *significant at 10%; ** significant at 5%; *** significant at 1%

31

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and PugatchT

able

4:M

exic

anim

mig

rant

pop

ula

tion

shar

esan

db

ord

eren

forc

emen

t

all

mal

es16-5

0u

nau

thori

zed

natu

rali

zed

citi

zen

ab

road

last

year

inte

rnal

mig

rant

HS

or

less

yes

no

yes

no

yes

no

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Pan

elA

:O

LS

enfo

rcem

ent

index

-0.1

76-0

.204

-0.2

00

-0.0

75

-0.2

09

-0.3

-0.1

75

-0.5

71

-0.1

75

(0.0

58)*

**(0

.057)*

**

(0.0

57)*

**

(0.0

50)

(0.0

56)*

**

(0.2

72)

(0.0

58)*

**

(0.3

50)

(0.0

58)*

**

R-s

qu

ared

0.04

0.0

30.0

30.0

20.0

30.0

70.0

40.1

20.0

4P

anel

B:

IVen

forc

emen

tin

dex

-0.2

19-0

.316

-0.3

24

-0.0

47

-0.3

04

-0.0

03

-0.2

14

0.2

05

-0.2

18

(0.0

75)*

**(0

.078)*

**

(0.0

75)*

**

(0.0

66)

(0.0

90)*

**

(0.3

16)

(0.0

74)*

**

(0.2

92)

(0.0

73)*

**

Ob

serv

atio

ns

839

809

789

789

815

220

839

280

838

1st

stag

eF

-sta

t75

.675.6

80.8

75.4

75.7

49.4

75.6

53.8

75.7

Tab

lesh

ow

sre

gre

ssio

ns

of

log

pop

ula

tion

share

on

enfo

rcem

ent

ind

ex,

infi

rst

diff

eren

ces.

Sam

ple

isU

.S.

state

-yea

rs(i

ncl

ud

ing

D.C

.)fr

om

Cu

rren

tP

op

ula

tion

Su

rvey

,1995-2

011.

Pop

ula

tion

share

isst

ate

’ssh

are

of

Mex

ican

imm

igra

nts

wit

hin

each

cate

gory

.E

nfo

rcem

ent

ind

ex=

∑ sP

r(U

Sd

esti

nati

on|c

ross

at

bord

erse

ctors)×

enfo

rcem

ent

at

sect

ors,

wh

ere

enfo

rcem

ent

isth

ou

san

ds

of

bord

erp

atr

ol

agen

ts.

Ind

exm

ay

be

inte

r-p

rete

das

am

ou

nt

of

enfo

rcem

ent

ded

icate

dto

pre

ven

tin

garr

ival

of

un

au

thori

zed

mig

rants

at

des

tin

ati

on

.In

stru

men

tin

IVsp

ecifi

cati

on

sis

seco

nd

lag

of

enfo

rcem

ent

ind

ex.

Cro

ssin

gp

rob

ab

ilit

ies

calc

ula

ted

from

EM

IFcr

oss

ings

1983-1

993.

Bord

erp

atr

ol

agen

tsfr

om

Dep

art

men

tof

Hom

elan

dS

ecu

rity

.U

nau

thori

zed

Mex

ican

imm

igra

nts

calc

ula

ted

by

mu

ltip

lyin

gu

nau

thori

zed

imm

igra

nt

esti

mate

of

Warr

enan

dW

arr

en(2

013)

by

the

pro

port

ion

of

imm

i-gra

nts

wh

oare

Mex

ican

inth

est

ate

-yea

rce

ll.

All

regre

ssio

ns

incl

ud

eyea

rfi

xed

effec

tsan

dth

efo

llow

ing

contr

ols

(in

1st

diff

eren

ces)

:M

exic

an

imm

igra

nt

un

emp

loym

ent

rate

,M

exic

an

imm

igra

nt

hou

rly

wage,

log

GD

Pp

erca

pit

a,

log

agri

cult

ura

lG

DP

,lo

gm

anu

fact

uri

ng

GD

P,

log

con

stru

ctio

nG

DP

,lo

gn

ewh

ou

sin

gp

erm

its,

an

da

du

mm

yfo

rp

ass

age

of

any

pu

nit

ive

imm

igra

tion

legis

lati

on

.R

ob

ust

stan

dard

erro

rsin

pare

nth

esis

,cl

ust

ered

by

state

.*

sign

ifica

nt

at

10%

;**

sign

ifica

nt

at

5%

;***

sign

ifica

nt

at

1%

32

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Tab

le5:

Pop

ula

tion

shar

esan

db

ord

eren

forc

emen

t

non

-Mex

ican

nat

ive

Pu

erto

Cen

tral

imm

igra

nt

un

auth

oriz

edR

ican

Am

eric

an

(1)

(2)

(3)

(4)

(5)

Pan

elA

:O

LS

enfo

rcem

ent

ind

ex0.

017

0.00

60.

017

0.00

4-0

.095

(0.0

30)

(0.0

17)

(0.0

14)

(0.0

68)

(0.0

70)

R-s

qu

ared

0.04

0.03

0.05

0.03

0.04

Pan

elB

:IV

enfo

rcem

ent

ind

ex-0

.03

-0.0

270.

037

0.01

7-0

.201

(0.0

44)

(0.0

53)

(0.0

17)*

*(0

.105

)(0

.084

)**

Ob

serv

atio

ns

863

812

863

739

832

1st

stag

eF

-sta

tist

ic75

.680

.975

.275

.575

.7

Tab

lesh

ow

sre

gre

ssio

ns

of

log

pop

ula

tion

share

on

enfo

rcem

ent

ind

ex,in

firs

td

iffer

ence

s.S

am

ple

isU

.S.st

ate

-yea

rs(i

ncl

ud

ing

D.C

.)fr

om

Cu

rren

tP

op

ula

tion

Su

rvey

,1995-2

011.

Pop

ula

tion

share

isst

ate

’ssh

are

wit

hin

each

cate

gory

ind

icate

d.

En

forc

emen

tin

dex

=∑ s

Pr(

US

des

tin

ati

on|c

ross

at

bord

erse

ctors)×

enfo

rcem

ent

at

sect

ors,

wh

ere

enfo

rcem

ent

isth

ou

san

ds

of

bord

erp

atr

olagen

ts.

Ind

exm

ay

be

inte

rpre

ted

as

am

ou

nt

of

enfo

rcem

ent

ded

icate

dto

pre

ven

tin

garr

ival

of

un

au

thori

zed

mig

rants

at

des

tin

ati

on

.In

stru

men

tin

IVsp

ecifi

cati

on

sis

seco

nd

lag

of

enfo

rcem

ent

ind

ex.

Cro

ssin

gp

rob

ab

ilit

ies

calc

ula

ted

from

EM

IFcr

oss

ings

1983-1

993.

Bord

erp

atr

ol

agen

tsfr

om

Dep

art

men

tof

Hom

elan

dS

ecuri

ty.

Un

au

thori

zed

imm

igra

nts

are

esti

mate

sof

Warr

enan

dW

arr

en(2

013),

mu

ltip

lied

by

the

pro

port

ion

of

imm

igra

nts

wh

oare

non

-Mex

ican

inth

est

ate

-yea

rce

ll.

All

regre

ssio

ns

incl

ud

eyea

rfi

xed

effec

tsan

dth

efo

llow

ing

contr

ols

(in

1st

diff

eren

ces)

:u

nem

plo

ym

ent

rate

of

ind

icate

dgro

up

,h

ou

rly

wage

of

ind

icate

dgro

up

,lo

gG

DP

per

cap

ita,

log

agri

cult

ura

lG

DP

,lo

gm

anu

fact

uri

ng

GD

P,

log

con

stru

ctio

nG

DP

,lo

gn

ewh

ou

sin

gp

erm

its,

an

da

du

mm

yfo

rp

ass

age

of

any

pu

nit

ive

imm

igra

tion

legis

lati

on

.R

ob

ust

stan

dard

erro

rsin

pare

nth

esis

,cl

ust

ered

by

state

.*

sign

ifica

nt

at

10%

;**

sign

ifica

nt

at

5%

;***

sign

ifica

nt

at

1%

33

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Table 6: State shares of Mexican immigrants, 1994-2011

State observed counterfactual change

1994 2011 2011(1) (2) (3) (3)-(2)

Alabama 0.05% 0.51% 0.26% -0.25Alaska 0.03% 0.04% 0.02% -0.02Arizona 4.50% 5.03% 4.72% -0.30Arkansas 0.15% 0.29% 0.15% -0.14California 54.02% 38.16% 46.40% 8.24Colorado 0.57% 1.78% 1.02% -0.76Connecticut 0.00% 0.18% 0.12% -0.05Delaware 0.06% 0.19% 0.10% -0.09District of Columbia 0.02% 0.06% 0.03% -0.03Florida 2.17% 1.95% 1.09% -0.86Georgia 1.28% 1.95% 1.02% -0.92Hawaii 0.01% 0.08% 0.04% -0.04Idaho 0.36% 0.57% 0.30% -0.27Illinois 6.53% 5.56% 3.02% -2.54Indiana 0.09% 0.82% 0.42% -0.40Iowa 0.11% 0.54% 0.28% -0.26Kansas 0.12% 0.63% 0.33% -0.30Kentucky 0.03% 0.32% 0.16% -0.15Louisiana 0.07% 0.21% 0.11% -0.10Maine 0.00% 0.01% 0.00% -0.01Maryland 0.09% 0.43% 0.22% -0.21Massachusetts 0.17% 0.05% 0.03% -0.02Michigan 0.12% 0.87% 0.45% -0.42Minnesota 0.28% 0.48% 0.25% -0.23Mississippi 0.03% 0.12% 0.06% -0.06Missouri 0.11% 0.25% 0.13% -0.12Montana 0.01% 0.03% 0.02% -0.01Nebraska 0.13% 0.54% 0.28% -0.26Nevada 1.14% 1.93% 1.02% -0.91New Hampshire 0.00% 0.00% 0.00% 0.00New Jersey 0.37% 1.80% 0.92% -0.88New Mexico 1.08% 0.99% 0.56% -0.43New York 1.33% 1.84% 0.95% -0.89North Carolina 0.53% 2.03% 1.09% -0.94North Dakota 0.01% 0.00% 0.00% 0.00Ohio 0.11% 0.34% 0.18% -0.17Oklahoma 0.56% 0.44% 0.23% -0.21Oregon 1.09% 0.74% 0.39% -0.35Pennsylvania 0.19% 0.38% 0.20% -0.19Rhode Island 0.02% 0.04% 0.02% -0.02South Carolina 0.06% 0.54% 0.28% -0.26South Dakota 0.00% 0.03% 0.01% -0.01Tennessee 0.03% 0.65% 0.34% -0.32Texas 20.85% 22.45% 30.63% 8.17Utah 0.32% 0.64% 0.33% -0.31Vermont 0.00% 0.00% 0.00% 0.00Virginia 0.12% 0.85% 0.44% -0.41Washington 0.47% 1.84% 0.97% -0.88West Virginia 0.00% 0.01% 0.00% -0.01Wisconsin 0.56% 0.79% 0.41% -0.38Wyoming 0.05% 0.05% 0.03% -0.02

Table shows state shares of Mexican immigrants in 1994, 2011, and 2011 under counterfactual in which border enforcement didnot change, in columns (1)-(3), respectively. Final column shows percentage-point change in state share due to enforcement,found by subtracting column (3) from column (2). Column (3) found by using estimated IV coefficient on enforcement indexfrom Table 4, column (1) to determine predicted change in population share in each state due to changes in enforcement. Thischange is then used to predict each state’s population share in 2011 as though no change in enforcement occurred. Details inAppendix C. Data source: CPS, 1994-2011.

34

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and PugatchT

able

A1:

Mex

ican

imm

igra

nt

pop

ula

tion

shar

esan

db

ord

eren

forc

emen

t(C

ensu

s/A

CS

)

all

mal

es16-5

0u

nau

thori

zed

natu

rali

zed

citi

zen

ab

road

last

year

inte

rnal

mig

rant

HS

or

less

yes

no

yes

no

yes

no

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Pan

elA

:O

LS

enfo

rcem

ent

index

-0.0

83-0

.005

-0.0

72

-0.0

24

-0.0

2-0

.028

-0.0

80.1

45

-0.0

87

(0.0

61)

(0.0

37)

(0.0

56)

(0.0

49)

(0.0

53)

(0.0

77)

(0.0

59)

(0.0

94)

(0.0

60)

R-s

qu

ared

0.08

0.1

70.0

80.0

70.0

90.0

80.0

70.1

10.0

9P

anel

B:

IVen

forc

emen

tin

dex

-0.2

22-0

.191

-0.2

05

-0.1

25

-0.1

95

-0.0

51

-0.2

25

0.1

26

-0.2

24

(0.0

83)*

**(0

.075)*

*(0

.081)*

*(0

.062)*

*(0

.074)*

**

(0.1

93)

(0.0

79)*

**

(0.1

60)

(0.0

86)*

**

Ob

serv

atio

ns

557

547

506

550

552

396

557

423

557

1st

stag

eF

-sta

t74

.874.8

98.6

74.5

74.9

90.5

74.8

91.6

74.8

Tab

lesh

ow

sre

gre

ssio

ns

of

log

pop

ula

tion

share

on

enfo

rcem

ent

ind

ex,

infi

rst

diff

eren

ces.

Sam

ple

isU

.S.

state

-yea

rs(i

ncl

ud

ing

D.C

.)fr

om

U.S

.C

ensu

s2000,

Am

eric

an

Com

mu

nit

yS

urv

ey2001-2

011.

Pop

ula

tion

share

isst

ate

’ssh

are

of

Mex

ican

imm

igra

nts

wit

hin

each

cate

gory

.E

nfo

rcem

ent

ind

ex=

∑ sP

r(U

Sd

esti

nati

on|c

ross

at

bord

erse

ctors)×

enfo

rcem

ent

at

sect

ors,

wh

ere

enfo

rcem

ent

isth

ou

san

ds

of

bord

erp

atr

ol

agen

ts.

Ind

exm

ay

be

inte

rpre

ted

as

am

ou

nt

of

enfo

rcem

ent

ded

icate

dto

pre

ven

tin

garr

ivalof

un

au

thori

zed

mig

rants

at

des

tin

ati

on

.In

stru

men

tin

IVsp

ecifi

cati

on

sis

seco

nd

lag

of

enfo

rcem

ent

ind

ex.

Cro

ssin

gp

rob

ab

ilit

ies

calc

ula

ted

from

EM

IFcr

oss

ings

1983-1

993.

Bord

erp

atr

olagen

tsfr

om

Dep

art

men

tofH

om

elan

dS

ecu

rity

.U

nau

thori

zed

Mex

ican

imm

igra

nts

calc

ula

ted

by

mu

ltip

lyin

gu

nau

thori

zed

imm

igra

nt

esti

mate

of

Warr

enan

dW

arr

en(2

013)

by

the

pro

port

ion

of

imm

igra

nts

wh

oare

Mex

ican

inth

est

ate

-yea

rce

ll.

All

regre

ssio

ns

incl

ud

eyea

rfi

xed

effec

tsan

dth

efo

llow

ing

contr

ols

(in

1st

diff

eren

ces)

:M

exic

an

imm

igra

nt

un

emp

loym

ent

rate

,M

exic

an

imm

igra

nt

hou

rly

wage,

log

GD

Pp

erca

pit

a,

log

agri

cult

ura

lG

DP

,lo

gm

anu

fact

uri

ng

GD

P,

log

con

stru

ctio

nG

DP

,lo

gn

ewh

ou

sin

gp

erm

its,

an

da

du

mm

yfo

rp

ass

age

of

any

pu

nit

ive

imm

igra

tion

legis

lati

on

.R

ob

ust

stan

dard

erro

rsin

pare

nth

esis

,cl

ust

ered

by

state

.*

sign

ifica

nt

at

10%

;**

sign

ifica

nt

at

5%

;***

sign

ifica

nt

at

1%

35

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Tab

leA

2:M

exic

anim

mig

rant

pop

ula

tion

shar

esan

db

ord

eren

forc

emen

t(a

llce

lls)

all

mal

es16-5

0u

nau

thori

zed

natu

rali

zed

citi

zen

ab

road

last

year

inte

rnal

mig

rant

HS

or

less

yes

no

yes

no

yes

no

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Pan

elA

:O

LS

enfo

rcem

ent

ind

ex-0

.18

-0.2

08

-0.2

00

-0.1

39

-0.1

75

0.2

62

-0.1

79

-0.8

46

-0.1

74

(0.0

58)*

**(0

.080)*

*(0

.058)*

**

(0.0

90)

(0.0

98)*

(0.4

75)

(0.0

58)*

**

(0.8

41)

(0.0

63)*

**

R-s

qu

ared

0.05

0.0

40.0

30.0

10.0

60.1

60.0

50.1

40.0

5P

anel

B:

IVen

forc

emen

tin

dex

-0.2

45-0

.242

-0.3

75

-0.1

-0.2

59

0.3

45

-0.2

4-0

.063

-0.1

67

(0.1

05)*

*(0

.157)

(0.0

83)*

**

(0.0

92)

(0.1

42)*

(0.7

77)

(0.1

04)*

*(0

.911)

(0.1

36)

Ob

serv

atio

ns

863

863

812

863

863

863

863

863

863

1st

stag

eF

-sta

t75

.675.6

80.9

75.6

75.6

75.6

75.6

75.6

75.6

Tab

lesh

ow

sre

gre

ssio

ns

of

log

pop

ula

tion

share

on

enfo

rcem

ent

ind

ex,in

firs

td

iffer

ence

s.S

am

ple

isU

.S.st

ate

-yea

rs(i

ncl

ud

ing

D.C

.)fr

om

Cu

rren

tP

op

ula

tion

Su

rvey

,1995-2

011.

Pop

ula

tion

share

isst

ate

’ssh

are

of

Mex

ican

imm

igra

nts

wit

hin

each

cate

gory

.O

ne

per

son

ad

ded

toall

state

-yea

rce

lls

soth

at

log

of

zero

pop

ula

tion

share

isw

ell-

defi

ned

.E

nfo

rcem

ent

ind

ex=

∑ sP

r(U

Sd

esti

nati

on|c

ross

at

bord

erse

ctors)×

enfo

rcem

ent

at

sect

ors,

wh

ere

enfo

rcem

ent

isth

ou

san

ds

of

bord

erp

atr

ol

agen

ts.

Ind

exm

ay

be

inte

rpre

ted

as

am

ou

nt

of

enfo

rcem

ent

ded

icate

dto

pre

ven

tin

garr

ival

of

un

au

thori

zed

mig

rants

at

des

tin

ati

on

.In

stru

men

tin

IVsp

ecifi

cati

on

sis

seco

nd

lag

of

enfo

rcem

ent

ind

ex.

Cro

ssin

gp

rob

ab

ilit

ies

calc

ula

ted

from

EM

IFcr

oss

ings

1983-1

993.

Bord

erp

atr

ol

agen

tsfr

om

Dep

art

men

tof

Hom

elan

dSec

uri

ty.

Un

au

thori

zed

Mex

ican

imm

igra

nts

calc

ula

ted

by

mu

ltip

lyin

gu

nau

thori

zed

imm

igra

nt

esti

mate

of

Warr

enan

dW

arr

en(2

013)

by

the

pro

port

ion

of

imm

igra

nts

wh

oare

Mex

ican

inth

est

ate

-yea

rce

ll.

All

regre

ssio

ns

incl

ud

eyea

rfi

xed

effec

tsan

dth

efo

llow

ing

contr

ols

(in

1st

diff

eren

ces)

:M

exic

an

imm

igra

nt

un

emp

loym

ent

rate

,M

exic

an

imm

igra

nt

hou

rly

wage,

log

GD

Pp

erca

pit

a,

log

agri

cult

ura

lG

DP

,lo

gm

anu

fact

uri

ng

GD

P,

log

con

stru

ctio

nG

DP

,lo

gn

ewh

ou

sin

gp

erm

its,

an

da

du

mm

yfo

rp

ass

age

of

any

pu

nit

ive

imm

igra

tion

legis

lati

on

.R

ob

ust

stan

dard

erro

rsin

pare

nth

esis

,cl

ust

ered

by

state

.*

sign

ifica

nt

at

10%

;**

sign

ifica

nt

at

5%

;***

sign

ifica

nt

at

1%

36

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Figure 1: Mexican immigrant diffusion

Figure 2: Mexican immigrant diffusion and border enforcement

37

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Figure 3: Mexican migrant enforcement and crossing patterns

(a)

(b)

38

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Figure 4: Mexican migrant crossing patterns

(a) (b)

(c)

39

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Figure 5: Enforcement index

Figure 6: Change in state shares of Mexican immigrants, observed and counterfactual

(a) (b)

40

U.S. Border Enforcement and Mexican Immigrant Location Choice Bohn and Pugatch

Figure A1: Mexican immigrant population, by data source

Figure A2: Border patrol agents and linewatch hours

41