Who Really Benefits from Export Processing Zones?

Estimating Distributional Effects Within Nicaraguan Municipalities1

Nathalie Picarelli2

London School of Economics

November, 2014

Abstract: This paper evaluates the distributional effects resulting from the establishment of

Export Processing Zones (EPZs) at the subnational level, by analysing Nicaragua’s program

from 1993 to 2009. Using a municipal-level panel, I estimate the average treatment effect on

a set of within-municipality inequality measures and quantile treatment effects across the

deciles of the real expenditure distribution of treated areas. I find that the overall impact on

disparity within treated municipalities is ambiguous, with a sustained increase in top-end

inequality that is mitigated by a milder, fluctuating decline in disparities at the lower-end of

the distribution. Primarily, I find that individuals at the 90th and 80th percentiles are the main

beneficiaries of the policy. The effect is heterogeneous in time for the rest of the segments,

diffusing at the 10th and middle-range percentiles after eight years of a firms’ establishment in

a given municipality. The time dynamics suggest that the initial inequality-enhancing effect of

the policy fades with the more years of operation, eventually reaching the middle and poorer

segments. The estimates fail to capture any spillover effects on the real expenditure levels

across the distribution of adjacent municipalities.

JEL Codes: C21, D31, D63, F13, O15, O24

Keywords: Export Processing Zones, Inequality, Distribution, Quantiles

1. Introduction

1 This is a preliminary draft not intended for citation or circulation without the author’s permission. I am indebted to Olmo Silva and Henry Overman for their guidance. I thank participants of seminars at the LSE for their helpful comments. All errors remain my own. This work was supported by the Economic and Social Research Council. 2 Email address: n.picarelli@lse.ac.uk

1

The use of Export Processing Zones (EPZs)3 as a policy for trade and economic development has

exponentially grown in the past decade, particularly but not solely in developing countries. In 1986, the

International Labor Organization (ILO) estimated that there were 176 zones in 47 countries. By 2008, the

number reached more than 3,000 zones in 135 countries, accounting for over 40 million direct jobs and

over US$200 billion in global exports (Farole and Akinci 2011). The attractiveness of this type of policy

lies in its wide scope: EPZs are considered a strategy to attract foreign direct investment (FDI), create

employment, support wider economic reform strategies, and as laboratories for new policies (FIAS 2008).

Despite the importance of the phenomenon, and the existence of a growing body of literature evaluating

their performance in achieving the latter objectives (Farole 2011, Engman and Onodera 2007, Aggarwal

2006, Glick and Roubaud 2006), little attention has been paid to understanding the distributional effects

resulting from the establishment of EPZs within countries, and more precisely within the municipalities

where these locate. Yet, if successful, EPZ programs are prone to have a significant influence not only at

the macro level - through growth generated by exports, FDI, or the restructuring of the business

environment –, but also on the local economies via micro channels (i.e. higher manufacturing wages,

employment generation, backward linkages with domestic suppliers, etc.). Disentangling in what way the

establishment of a zone profits the different segments of the income distribution within concerned areas is

central to apprehend the welfare dynamics of this popular tool and shed light on the ultimate beneficiaries

of the policy at the local level. It is also of great policy relevance, particularly in developing economies with

already large levels of overall income disparities.

In this sense, this analysis builds on a small literature studying the impact of EPZ policies on outcomes

measured at the subnational level in the context of developing countries. Findings in these papers are

revealing of the extent to which local economies are influenced by the establishment of an exporting zone.

For instance, using employment data and census statistics on educational outcomes at the municipal-level

for Mexico (1986-2000), Atkin (2012) finds an increase in school dropout in concerned municipalities

following the arrival of new better-paying export jobs (i.e. due to a higher opportunity cost of remaining at

school). He estimates that this negative effect could have been compensated had the firms located in

municipalities with lower levels of educational achievement. Similar in strategy to this paper, Wang (2013)

uses a Chinese municipal dataset to demonstrate that Special Economic Zones (SEZ) benefit local

economies in several ways. He finds that on average host municipalities have higher levels and growth

rates of per capita FDI and total factor productivity (TFP), as well as higher average wages that

compensate any increase in the local cost of living following the establishment of a zone. In many ways,

these findings highlight the various channels through which EPZs can have diverse distributional effects

within the areas where they locate.

This study takes advantage of the gradual establishment of EPZs in Nicaraguan municipalities during the

period 1993-2009 to assess the distributive pattern of this specific spatially-bound policy within the host

municipalities. Nicaragua provides an excellent setting to study this phenomenon. Not only is the program

considered a success in terms of generating employment and export growth, but the number of firms

operating under the regime has increased markedly since its inception across more than 20 of the 153

municipalities of the country. By 2010, EPZs jobs represented 25 percent of total formal work across the

3 When discussing EPZs, a variety of terminologies, such as Industrial Free Zones, Free Trade Zones, Special Economic Zones and Maquiladoras are used interchangeably in the literature. Although each has their own particularity, we will consider the broad definition of them being “demarcated geographic areas contained within a country’s national boundaries where the rules of business are different from those that prevail in the national territory. These differential rules principally deal with investment conditions, international trade and customs, taxation, and the regulatory environment; whereby the zone is given a business environment that is intended to be more liberal from a policy perspective and more effective from an administrative perspective than that of the national territory” (Farole and Akinci 2011:23).

2

country, with on average firms directly employing around 7.5 percent of the total labor force of the

municipalities where they are located4. Furthermore, at traditionally high levels, inequality in the country

registered an important decline throughout the period, offering an interesting setting to study the

distributional effect of EPZs at the subnational dimension.

For the analyses, I exploit an individual-level repeated cross-sectional dataset based on official household

surveys and construct a unique municipal-level panel that allows for the examination of inequality

dynamics before and after the policy implementation. By focusing on the aggregate outcomes at the

municipal level, the study captures general equilibrium impacts of EPZ establishment within a locality. The

identification strategy is straightforward. I use both the time and cross-section variation of zones’

establishment across municipalities to estimate their average effect on a set of within-municipality

inequality measures and on the levels of real expenditure per capita across all the deciles of the expenditure

distribution of treated municipalities. In the main approach, I use a difference-in-differences (DID)

strategy extended to quantiles as discussed by Athey and Imbens (2006). To increase the robustness of the

results, I also exploit the original repeated cross-sectional data to compute unconditional quantile

treatment effects (Firpo 2007) on the entire expenditure distribution of treated municipalities. An

important feature of the overall empirical strategy needs to be emphasized. The approach does not aim at

measuring the impact of the EPZ policy on inequality across Nicaragua as a whole. Rather, it looks at the

question of knowing whether within municipalities exposed to the establishment of EPZs, certain

segments of the distribution capture more or less of the resulting gains or losses.

The validity of the findings rely on the assumption that the different empirical methods manage to account

for the underlying differences in municipalities’ observable and unobservable characteristics that are likely

to explain the non-random choice of EPZs location across time and space. In the absence of a credible

instrument, I take care of unobservable confounders by allowing for time and municipalities fixed-effects

across specifications. Additionally, I use information on accessibility and socio-economic indicators to

build alternative sets of control municipalities that are likely to be more meaningful comparable groups

under the DID setting. The assumption of historical parallel trends of main outcomes holds for the key

control groups.

Additionally, the ‘experiment’ created by this setting may differ from the ideal case where the influence of

EPZs is contained within the static administrative borders of the treated municipality. As such, two other

mechanisms also pose a potential threat to the consistency of the regression estimates: the existence of

possible spillover effects to neighbouring areas and people’s mobility throughout the period. I test for the

existence of spillovers on the levels of expenditure per capita across the distribution of neighbouring areas

using varying degrees of distance to capture dynamics resulting either from firm creation/diversion or

commuting patterns. I find the effect to be close to zero, which is consistent with findings in the literature

registering small backward linkages generated by EPZ policies (ILO 2008)5. Also, data available for 2001

pinpoints to relatively low levels of commuting (3.5 percent of sample). The situation is more complex

concerning larger displacements. Information available at the individual-level allows controlling for within-

country migration in specifications using the repeated-cross sectional dataset. I add a grouped measure as

control in the municipal-level estimations. However, a potential pitfall remains. As noted by Angrist and

Piscke (2009), time-varying variables measured at the aggregated level are usually affected by treatment,

and can introduce endogeneity. Still, while concerns regarding possible population adjustments towards

4 Data obtained from the Comisión Nacional de Zona Franca (CNZF) and Nicaragua’s Central Bank (BCN). 5 Some of the main reasons for this failure are found to be related to the difficulty for domestic firms of supplying low-cost, high-quality inputs necessary for global production at EPZs, as well as to low technological spillover resulting from the unskilled labor nature of the EPZ work (ILO 2008).

3

EPZ municipality are valid, these are likely to be negligible in the Nicaraguan context. Research conducted

with U.S. data has found that regional adjustments to labor market shocks are slow and incomplete, mostly

due to the cost of mobility (Autor, Dorn and Hanson 2012). The latter is likely to be even higher in the

context of developing countries with poor infrastructure and rigid labor markets. In Nicaragua, during the

period studied, the urbanization rate remained low (increasing form 62 percent in 1995 to 69 percent in

2010), and there was an overall decline in the urban population growth rate from 3.8 in the period 1990-

1995 to 2.83 in the 2005-2010 period. Testing for an effect of EPZ establishment on within-country

migration seems to confirm this hypothesis. Overall, I do not find evidence of population adjustments

towards municipalities hosting EPZs.

The analysis reveals interesting conclusions concerning distributional dynamics of EPZ policies at the local

level. First, I find that the overall effect on within-treated municipalities’ inequality is ambiguous, with a

sustained increase in upper-end inequality that is mitigated by a smaller varying decline in bottom-end

disparities. Above all, individuals located at the 90th and 80th percentiles of the distribution in treated areas

are the main beneficiaries of the policy over the period covered. Interestingly, the effect is heterogeneous

in time for the remaining percentiles, diffusing to the first and middle-range deciles only after eight to

eleven years of an EPZ establishment in a given municipality. The time dynamics suggest that the initial

inequality-enhancing effect of the policy fades with the more years of operation, eventually reaching the

middle and poorer segments. As mentioned, the estimates fail to capture any spillover effect resulting from

EPZ establishment across the real expenditure distribution of neighbouring areas.

This paper is original for three reasons. First, it is to my knowledge the first time that the impact of EPZs

on income distribution at the subnational level is formally assessed. Second, it contributes to add empirical

evidence to the evaluation of place-based policies in the context of developing countries. Finally,

methodologically, it combines panel and individual-level analysis to estimate treatment effect on the entire

income distribution with two alternative estimators.

The remainder of the paper is organized as follows. Section 2 lays out the theoretical framework. Section 3

briefly describes the EPZ program in Nicaragua, while section 4 introduces and examines the data. Section

5 discusses the empirical strategy and section 6 presents the results. Robustness checks are shown in

section 7. The final section concludes and looks at policy implications.

2. Theoretical Grounds

To start, I look at the theoretical underpinnings that guide this empirical work. While there has been a

strand of literature that has looked at the overall welfare effects of the establishment of EPZs on host

countries using basic trade models and general equilibrium approaches6, the analyses have failed to

consider distributional issues. For this, I consider the main channels described by the rich body of research

that has studied the influence of trade openness on interpersonal inequality through similar conceptual

frameworks. I focus on the principal mechanism, namely the labor-market effects, and then briefly

summarize secondary channels (i.e. effect on asset ownership, changes in household production and

consumption decisions, and general equilibrium effects through changes in relative prices).

2.1 Labor-Market Channels

6 A review by Jenkins et al. (1998) concludes that largely, the implications of EPZs on national welfare are found to be ambiguous and depend on the complexity of the model used as well as their specific assumptions. For more details see Hamada (1974), Hamilton and Svensson (1982), Wong (1986), Chaudhuri and Adhikari (1993), and Din (1994).

4

The most straightforward effect of EPZs on income distribution is likely to happen through changes on

labor income, resulting from the impact on the relative demand for factors of production – here that is

skilled vs. unskilled labour. The basis to understand these dynamics is the standard 2x2 Heckscher-Ohlin

(HO) model and its related theorem Stolper-Samuelson (SS). Indeed, even though the theoretical and

empirical shortcomings of this basic model are widely recognized7, it has been the foundation to think

about distributional effects of trade policies over the past decades. Together, in the simplest form, they

predict that trade openness pushes countries to specialize in the production of goods where they have a

relatively-abundant factor of production. This, in turn, boosts the demand for the relatively-abundant

factor and raises its relative returns. In developing countries, the basic form of the theorem therefore

implies that unskilled-labor (the abundant factor of production) will be favoured by the resulting

distributional changes.

The dynamics alas are much more complex and depend on the consideration of additional interactions.

There are fundamental limits to this factor-endowment-based trade theory. In the context of developing

countries the most salient one is the de facto constraints to labor reallocation – one of its fundamental

mechanisms of adjustment – due to labor market rigidities, the existence of prominent informal sectors

and low labor mobility. In fact, recent trade models have shown that liberalization can reduce the wages of

unskilled labor even in labor abundant economies (Banerjee and Newman 2004). Goldberg and Pavcnik

(2007) offer a good review of alternative dynamics. These effects reverse the distributional predictions of

the basic HO model. I describe some of them next.

First, altering the simple HO model by introducing non-traded goods or additional factors of production

overturns the simple SS theorem predictions. In these cases, if the production of tradable goods requires a

higher ratio of skilled-to-unskilled workers, then trade openness will benefit skilled workers. Harrison and

Hanson (1999) have shown for instance that exporting plants in Mexico employ a higher share of white-

collar workers than non-exporting firms, thus inferring a higher ratio of skilled-to-unskilled labor than

similar firms in the domestic economy. A similar outcome occurs when adding capital mobility between

countries, with the idea that the increase in capital flows is associated with a higher demand for skilled

workers.

Adding technology to the equation also increases the relative demand for skilled-labour: Feenstra and

Hanson (1997) show that cheaper access to foreign technology allows developing countries to compete

internationally in more skill-intensive goods, which in turn allows firms in high-income countries to

minimize costs by outsourcing intermediate inputs to these countries. While outsourced goods would be

characterized as unskilled-labor-intensive from a developed country’s perspective, the authors argue that

they are actually skilled-labor-intensive when compared with existing domestic activities in developing

economies. As such, outsourcing increases the average skill intensity of production in developing

economies, thus increasing the relative demand for skilled labor and inducing an increase in the skill

premium. Similar dynamics have been shown by Verhoogen (2008), who argues that trade openness leads

to an upgrading of the average product quality in exporting plants, which in turn generates demand for a

better qualified workforce. Finally, an additional channel includes the existence of an industry wage

premium. The latter may arise from productivity improvements resulting from the higher exposure to

trade in a specific industry together with labor market rigidities that prevent labor reallocation (Heckman

and Pages 2000).

7 On the theoretical side the model relies on restrictive assumptions such as perfect mobility of factors within a country, perfect competition, and fixed technology; while on the empirical side there has been no support for the predictions in its basic form.

5

Extending these conceptual frameworks to EPZs is relatively straight-forward given that employment

generation in the zones correspond to ‘export jobs’, and thus it impacts labor income through an increase

in the relative demand for a specific factor of production or by the possible existence of a wage premium

in the exporting sector. As described for the general case of trade openness, the distributional dynamics

will depend on which factor of production relatively benefits from EPZs’ jobs. The answer is far from

obvious. The simple HO-SS predictions imply that the zones will concentrate in unskilled-labor-intensive

sectors such as manufacturing or agriculture - as is manifestly the case in Nicaragua and other developing

economies -. In its basic form, this concentration should favour unskilled labor and thus lead to a decrease

in inequality. Using the case of Madagascar, Glick and Roubaud (2006) find indeed evidence that

employment in the zones benefit workers with low-levels of schooling. However, the more subtle

mechanisms are also very much at play with EPZs (i.e. outsourcing, capital flows, export-related

productivity improvements, labor market rigidities), and they may in practice induce a skill premium that

modifies the distributional predictions of the standard model, as discussed. There is in fact more

documented evidence on the latter. Engman (2011) finds evidence of higher TFP levels in the EPZ sector

of Honduras. Atkin (2012) and Wang (2013) also provide solid evidence that the levels of wages are equal

or higher in EPZs for a similar set of skills than outside the zones. The broader literature exploring the

impact of exporting firms and multinational corporations on labor market outcomes in developing

countries provides further evidence of the existence of a sector premium8.

2.2 Secondary Channels

There are seemingly many additional mechanisms by which EPZs may impact income distribution; the

overall general equilibrium effect on inequality depending in the end on the net balance of forces. I discuss

the two most relevant next.

First, EPZs may affect income distribution in benefiting the owners of the factors of production other

than the skilled and unskilled labor (Meschi and Vivarelli 2008). An immediate example comes with land

ownership or natural resources. In this case, EPZs could contribute to income polarization. Anderson

(2005) imagines mechanisms that could affect the amount of inequality in the ownership of the factors of

production per se, either via income effects that relax credit constraints for instance, or via the reaction of

individuals to the way they adjust their holdings of assets in response to changes in the return of those

assets (i.e. land).

Second, in the medium to long term, important complementary channels are the ones linked to

households’ production and consumption decisions. The more direct one relates to the employment

choices for households’ members, which is particularly important in low-income countries. The

establishment of EPZ affect the choice between employment in the export sector and self-employment in

the informal sector, or with respect to educational choices. At the same time, consumption decisions may

be affected through general equilibrium effects reflected in local prices. For empirical evidence on these

two channels see Atkin (2012) and Wang (2013), respectively.

2.3 Spatial Considerations

An important concern here is the spatial aspect. The spatially-fixed nature of EPZs and their propensity to

hire or outsource in the area where they operate implies that the micro channels described above are likely

to have a local dimension. Limitations to factor mobility may help emphasize this dimension. These

8 Atkin refers to the findings of this literature, asserting that the most robust stylized facts that has emerged from research in this area using micro-level firm data, has been that exporting firms pay higher wages. For Mexico he cites Bernard (1995), Zhou (2003) and Verhoogen (2008).

6

characteristics justify the subnational approach taken in this paper. In this sense, results complement

findings in the strand of literature that has looked at the effect of trade liberalization on outcomes in local

and regional labor markets (Autor, Dorn and Hanson (2012), Kovak (2011), Topalova (2010))9.

3. Nicaragua’s Export Processing Zone Program

3.1 Ley de Zona Franca and Stipulations

Resembling analogous initiatives, Nicaragua’s EPZ regime (‘Zonas Francas’) was put in place with the aim

of generating employment, attracting FDI, increasing non-traditional exports, acquiring new technologies

and expanding international trade. The present program is relatively young compared to other Central

American countries as a previous initiative was abandoned for a decade during the civil conflict (1979-

1990) and the program was only reintroduced following the shift towards democracy and the market

economy in 1990. In 1991 Decree N.46-91 formally established the existing regime, with the appropriate

regulation passed in the following year (Decree N.31-92). Together, these two decrees created the legal

framework concerning the organization of EPZs, the fiscal incentives, the rights and obligations of

operators and users, as well as the regulatory body in the form of the National Commission of Free Trade

Zones (CNZF). The law was amended over the years, with its most notable reform taking place in 2005

(Decree N.50-2005). The latter significantly simplified procedures, and together with the signing of a Free

Trade Agreement with the U.S. (DR-CAFTA) in the same year, is considered instrumental in the increase

of FDI and firms’ participation in the EPZ program in the country (World Bank 2011).

Under the Nicaraguan legislation, foreign and national private firms are considered eligible for EPZ status

if they are oriented towards the production and export of goods and services. The law distinguishes three

different categories of firms: administrators (industrial parks operators), users (situated within industrial

parks) and “single-factory” EPZs. Technically, the regime institutes the entire territory as a possible host

for an EPZ, with firms applying to the status and locating themselves freely either as stand-alone plants, in

the form of industrial parks or within the latter. For the empirical analysis, I primarily define as an EPZ

single factories and industrial parks hosting one or more firms, which have been designated as such, and

are reported by the CNZF10.

Following standard practice, the program is based on attracting and generating business by offering

particular tax incentives to qualifying firms. These include the full exemption on (1) income tax, (2) taxes

on municipal property sales, (3) imports taxes of machinery, equipment and intermediate goods, as well as

of transportation and support services for the zone, (4) municipal taxes, (5) indirect taxes, taxes on

selective sales or consumption, (6) export taxes on processed products in the area; as well as the (7)

unrestricted repatriation of capital. Under Nicaraguan law, EPZs are allowed to sell intermediate goods

and inputs to other exporting firms in the country, provided they pay custom duties. However, they are

strictly forbidden from selling final goods in the local market. The legislation also provides conditions

under which EPZ developers may benefit from tax exemptions, as well as the possibility for EPZs to

subcontract to local Nicaraguan firms which are then exempted from VAT. These characteristics of the

program are important. On the one hand, the restriction to sell in the local market as well as the barriers to

sell intermediate inputs means that the larger effect of EPZs is going to be contained to the exporting

sector, without hampering local producers. On the other hand, the possibility to outsource to local

9 These have emphasized the limitations of perfect factor mobility across geographical areas, and focused on treating local labor market or subnational entities as sub-economies subject to differential trade shocks 10 From the CNZF Annual Yearbooks, I compiled information on EPZs location by the end of each year of interest, origin and sector of operation.

7

providers, which are then exempted from VAT, indicates that there are incentives in place to generate

backward linkages with local producers. Both these dynamics suggest that the impact of zone

establishment on local economies will be positive, either through business creation or employment

generation. The relative price effects on consumption decisions or local costs of living are harder to

predict.

Figure 3.1 Gradual Establishment of EPZs in Nicaraguan Municipalities (1993-2009)

1993

1998

2001

2005

2009

Notes: GIS digital maps. Municipalities are highlighted following the establishment of an industrial park or single factory under the EPZ regime. During the period considered some municipalities may contain more than one EPZ. Data was compiled from CNZF Annual Yearbooks.

8

3.2 Characteristics and Results

Since the launch of the program, the number of EPZs in the country visibly picked up. For the period

under study, the number of single factories and industrial parks under the regime increased from 1 in 1993

(only publicly-owned park) to 64 in 2009, located in 20 (of 153) different municipalities and 10 (of 18)

provinces (Figure 3.1). Approximately 90 percent of these EPZs operate in the manufacturing sector, with

a predominance of firms producing textile and apparel goods, followed by cigar production and light

electronics. Near 35 percent of the firms are from Nicaraguan origin, followed by North American (28

percent), Asian (19 percent) and other Central American countries (12.5 percent).

The policy has been successful in terms of generating employment, and has contributed significantly to

exports growth. Exports from EPZs increased from representing only 1 percent of GDP in 1994 to 15

percent of GDP in 2010. By that same year, the zones accounted for 50 percent of total exports in the

country (Figure 3.2). More precisely, mirroring their strong concentration in the manufacturing sector, they

were responsible for 93 and 95 percent of Nicaragua’s exports of apparel and machinery, respectively. The

growing contribution of manufacturing to the total productive value-added over the period confirms the

expansion of the sector: manufacturing went from representing near 16.5 percent in 1994 to around 20.5

percent in 2010, just behind agriculture which remained first throughout the period. It is important to note

that there seems to be a relatively low to moderate level of backward linkages with local producers. Their

share in EPZ’s exports is estimated to be close to 30 percent (ECLAC, 2012), putting Nicaragua far

behind leading Latin American countries such as Uruguay (61 percent)11.

Figure 3.2 EPZs’ Exports as a share of GDP and Total Exports FOB (1994-2011)

Figure 3.3 Number of EPZs and Employment (1991-2009)

More notably, EPZs have had a significant impact in terms of employment generation (Figure 3.3). In

2010, EPZs jobs represented 25 percent of total formal work in the country. The zones are estimated to

directly employ near four percent of Nicaragua’s labor force (2.2 million in 2009), and the CNZF estimates

that the impact through indirect employment is three times as high. This mechanism is probably the

principal channel through which EPZs may impact the local economy of municipalities where they locate.

On average in 2009, EPZs directly employed nearly 7.5 percent of the total labor force of concerned

11 In Central America, Nicaragua remains amongst the best performers. Honduras has the highest local contribution at 35 percent and Panama de lowest at 7 percent.

Notes: Data compiled from CNZF Annual Yearbooks, Pro Nicaragua and Nicaragua’s Central Bank; * in Figure 3.3 specifies years when household survey data is available.

9

municipalities, and reached more than 20 percent of their labor force indirectly (the numbers stay high at

the province level as well at 3.1 and 9.3 percent, respectively). Women accounted 55 percent of people

employed in EPZs in 2010, a decline from its early beginnings when they averaged above 60 percent. Still,

these numbers are well below popular cases such as Bangladesh (85 percent) or the Philippines (74

percent) and imply a wider reach of EPZs’ employment in the country.

Nicaraguan labor laws regulate work in EPZs and minimal wages have been in place since 1999 (Table

3.1). Despite the country having the lowest manufacturing wages in the region, a source of its comparative

advantage, minimum wages in EPZs have been traditionally higher than the ones in the overall

manufacturing and agricultural sectors throughout the period. Seemingly, this results from the sector

having registered the highest productivity levels in the country during the past decade12, pushed by the

expansion of EPZs’ activity in agribusiness and textiles.

Table 3.1 Average National Wage and Minimum Wages in Selected Sectors (1997-2009)

A final note regarding Nicaragua’s general labor market is warranted. Labor force has expanded rapidly

since the 1990s, at annual average rates that have outstripped population growth (3.6 percent vs. 1.7

percent, respectively). Positively, employment generation has followed a similar rate, with unemployment

remaining stable at around 6 percent for most of the 2000s. However, despite the sustained growth of the

manufacturing sector, the employment structure of the country remained largely unchanged in the last

decade, dominated by the labour-intensive low-skill agricultural sector. The latter even registered an

increase in participation in the 2000s (World Bank, 2012). These trends mirror the relatively more rural

nature of the country13 and its relatively low education level. In fact, while average educational attainment

12 In 2010, manufacturing accounted for nearly 21 percent of output and 11 percent of total employment (Central Bank of

Nicaragua). 13 Urbanization rate increased from 62 percent in 1995 to near 69 percent in 2010 (UN-Habitat), and there was an overall decline in urban population growth rate from 3.8 in the period 1990-1995 to 2.83 in the 2005-2010 period.

Minimum Wages

Agriculture Manufacturing EPZ

1997 1,617.3 300.0 500.0 .

1998 1,964.1 . . .

1999 2,282.3 450.0 600.0 800.0

2000 2,585.0 . . .

2001 2,897.1 550.0 670.0 895.0

2002 3,134.5 580.0 730.0 960.0

2003 3,388.2 615.0 825.0 1,037.0

2004 3,686.3 669.3 897.9 1,128.6

2005 4,266.2 769.4 1,032.7 1,298.4

2006 4,823.6 869.4 1,212.7 1,478.4

2007 4,957.2 1,025.9 1,431.0 1,744.5

2008 5,341.9 1,392.1 1,941.8 2,367.5

2009 6,010.3 1,573.1 2,155.5 2,556.7

Average National

Notes: Data compiled from Nicaragua’s Central Bank, Nominal Córdobas.

10

increased significantly during the past decade it remained one of the lowest in Latin America14.

Additionally, human capital accumulation in Nicaragua is unequal and inefficient with most of the

population falling short of even starting secondary education and only a small proportion completing

tertiary levels (World Bank, 2012).

4. Data and Basic Trends

4.1 Datasets

The data for this analysis were drawn from several sources. I compiled a repeated cross-sectional dataset

for the period 1993-2009 based on five LSMS Household Surveys15 (Encuesta Nacional de Hogares para la

Medición del Nivel de Vida, EMNV) collected by the Nicaraguan Institute of Statistics (Instituto Nacional

de Información de Desarrollo, INIDE). These contain detailed information on living standards of

individuals, occupation, industrial affiliation and various other household and individual characteristics.

The complete cross-sectional dataset amounts to 138,429 individual observations, from a total of 26,162

households, randomly selected according to the INIDE’s methodology to be representative at the national,

regional, urban and rural levels. I used the information from these repeated cross-sectional surveys to

construct a panel at the municipal-level and calculate inequality indices and various expenditure deciles at

this level of aggregation. These measures include a correction for the sample bias using sampling weights.

Municipalities with less than 30 observations per period were excluded to reduce measurement error.

Overall, the sample comprises observations for an unbalanced panel of between 80 to 136 municipalities

per year (Table 4.1).16 Information on Export Processing Zones is compiled from various CNZF Annual

Yearbooks, which contain information on location and sector of operation of each zone.

14 For the population 25 years and older, average educational attainment reached 6.2 years in 2011. While this was three times higher than in the 1960s, it remained well below the Latin American average of 9 years (World Bank, 2012). 15 The Living Standards Measurement Surveys were developed by the World Bank to improve the type and quality of household data collected by statistical offices in developing countries. The years of each EMNV are: 1993, 1998, 2001, 2005, and 2009. 16 Only 5 municipalities are omitted because of N<30 rule across the years. As a robustness check, I run the different specifications with a more balanced panel, omitting observations observed less than 4 times. Conclusions remain unchanged.

Sample Size 1993 1998 2001 2005 2009 Total

Municipalities 83 124 124 136 108 575

with EPZs 1 4 9 17 20 51

Individuals

Inviduals (>15 years)

Households

Table 4.1 Sample Description (Cross-Section and Panel Datasets)

Notes: Not all municipalities are observed every year. Municipalities with less than 30 individual observations

each year were dropped. On average, there are 250 individuals surveyed per municipality across the years, with

min 60 and max 5,000.

25,081 23,620 22,763 36,614

26,162

58,286

138,429

4,447 4,202 4,159 6,861 6,493

30,351

11,909 10,825 9,799 15,248 10,505

11

4.2 Conceptual Issues: Measuring Inequality

Assessing interpersonal inequality has always been a delicate issue, the ideal measure being based on

comparisons of an individuals’ wellbeing over their entire lifespan. Given the obvious constraints, an array

of different methods for analysing dispersion have been suggested and used in the empirical literature, with

most studies using cross-section regression analysis on changes or levels of synthetic or parametric

inequality measures based on income, wages or expenditure data.

In the present analysis I will use two popular measures of inequality for an overall appraisal of the effect of

EPZs establishment on the inequality of a given municipality: the Gini and Theil (T) indices17. Still,

different parts of the income distribution may have diverging dynamics and impact differently overall

disparity. This reality suggests that using a single distributional statistic may be unduly restrictive and

pinpoints to the importance of looking at the shape of the distribution more generally (Voitchovsky 2005).

In this sense, while the synthetic measures offer a starting point, I will also compute the ratios of income

percentiles on either side of the median to measure top and bottom-end inequality. More precisely, I use

the 50/10 percentile ratio to measure bottom-end inequality and the 90/50 ratio to estimate top-end

values. Additionally, following Milanovic (2002, 2008) I will focus on the pattern of change across all the

deciles shares of the distribution. This approach not only presents a more nuanced and accurate picture of

the entire distribution than any single number, but has the advantage of showing the precise beneficiaries

of the policy.

This study follows common practice in referring to the measure of living standards as income, which

should be interpreted broadly to encompass all the characteristics associated with geographical location

including climate or local public good provision, in order to assume that individuals with the same level of

income at different locations are equally well-off (Shorrocks and Wan 2004). Here, following best practice

income is proxied by consumption expenditure. The choice for using consumption expenditure as a

measure of income has been widely discussed in the literature. It is arguably the most appropriate variable

for capturing lifetime wellbeing (Deaton 1997) as it better captures intertemporal shifts of resources and

incorporates changes in purchasing power. Furthermore, expenditure data are of better quality in LSMS

Household Surveys (Deaton 2004; Banerjee and Duflo 2007), given that reporting problems are less

pronounced and consumption is less affected by the redesigning of surveys, which hampers comparability

across years.

The Gini and Theil indices, as well as obviously the percentile ratios, are hence all computed based on per

capita expenditure data. I use constant annual per capita expenditure (in Córdobas of 1999, C$)18, defined

as the household annual net expenditure divided by the number of persons in the household. All

households surveyed and all of their members are included (though consumption is adjusted depending on

members being children or adults). Missing values and zero incomes have been excluded. Additionally,

expenditure has been adjusted for difference in prices across regions. As noted by Shorrocks and Wan

(2004), adjusting for spatial price differences will tend to lower the between-group inequality, while not

17 Given the popularity of these measures I do not define them with more precision (See World Bank 2009). The Gini coefficient is derived from the Lorenz curve, and ranges from 0 (perfect equality) to 1. Theil indices belong to the family of the general entropy (GE) inequality measures, and can take values from zero to infinity (zero representing perfect distribution). The main determinant of GE measures is the alpha parameter, with higher alpha values increasing the sensibility of the measure to variations in the upper-tail of the distribution, and vice versa. Theil’s T corresponds to a GE with alpha equal to 1. I sometimes include Theil’s L (alpha equal to zero), as well as the GE measure with alpha 2, for comparison. 18 I use the Managua CPI from the Central Bank of Nicaragua.

12

altering inequality within regions, and are an important factor to take into account given the strong

correlation of prices with welfare levels.

A final note, empirically the treatment of space in the analysis of income inequality and regional income

distributions is relatively recent (Rey and Janikas 2005). This paper follows the common approach of

looking at aggregated income differentials partitioned into a set of spatial units, and analyses the data in

terms of the inequality observed within each unit, i.e. municipalities. The choice of using municipalities as

main unit of analysis stems from the fact that they are the ultimate discrete policy unit for which welfare

data is available in Nicaragua, and they remain the reference framework for delivering public services and

distributing government funds (Lall and Chakravorty 2005).

4.3 Descriptive Statistics & Basic Trends in Inequality

Tables 4.2 to 4.5 contain key summary statistics for the main variables used in this paper. They are

grouped according to municipalities’ EPZ status (or treatment status). The descriptive statistics for control

variables are presented in Table.A.1 in Appendix.

The first two tables below (4.2-4.3) offer a noticeable contrast regarding inequality and real per capita

expenditure measures at the beginning and end of the period studied. To begin with, there is a sharp

reduction in inequality between both years, independently of the method used and across all municipalities.

The fall is in line with the national trend: between 1993 and 2009 the Gini coefficient declined from 49.4

to 37.1 in the country as a whole (World Bank, 2009).

Second, one can also notice similar levels of expenditure per capita at the tails of the distribution in both

groups at the beginning of the period. Interestingly, for the same year the difference for the mean and

Table 4.2 Selected Outcome Statistics for Municipalities according to EPZ status, 1993

t-test

N=

Mean SD Mean SD Mean SD

Gini Coefficient 36.58 10.01 39.22 9.06 36.1 10.16 -1.03

Theil's T Index (GE1) 22.96 11.01 23.45 7.88 22.87 11.54 -0.18

Percentile Ratio 90th/10th 5.71 3.05 6.33 2.51 5.59 3.14 -0.8

Percentile Ratio 50th/10th 2.1 0.67 2.3 0.64 2.06 0.67 -1.16

Percentile Ratio 90th/50th 2.68 1.12 2.7 0.66 2.68 1.18 -0.07

Expenditure per capita

10th percentile 1,068.15 664.13 1,351.26 451.51 1,015.57 686.15 -1.69

50th percentile 2,155.74 1,236.90 3,078.41 1,200.08 1,984.38 1,173.63 -3.07

90th percentile 6,141.49 7,088.38 8,456.50 3,981.40 5,711.57 7,467.10 -1.28

Mean 3,108.71 1,908.33 4,403.47 1,908.33 2,868.25 2,371.87 -2.2

Entire SampleMunicipalities

Ever EPZ

Municipalities

Never EPZ

Notes: The data is for 83 municipalities in 1993, 13 ever designated EPZ. Expenditure measures are expressed in constant Cordobas

of 1999 (Exchange rate US$1=C$11.8). Theil's T index is a Generalized Entropy (GE) inequality measures, with α=1. PR90/10,

PR50/10 and PR90/50 correspond to percentile ratios of the 90th to 10th percentile, 50th to 10th, and 90th to 50th, respectively.

701383

Selected Outcome Measures

13

median are statistically significant, implying to some extent already higher level of real per capita

expenditure at the mid-percentiles of treated municipalities in 1993. By 2009, all deciles and the mean

display highly significant differences (at 1 percent levels) between treated and non-treated groups, in

favour of higher levels of expenditure per capita in EPZ municipalities. Finally, it is interesting to look at

the evolution of within-group inequality. While both groups had similar within-area disparity levels in

1993, the municipalities hosting a zone reveal statistically significant higher levels for both the Gini and

Theil indices in 2009. A look at the percentile ratios seems to indicate that these higher disparity levels

result from dynamics between both tails and at the top-end of the distribution. Largely, these trends

presuppose the existence of a heterogeneous effect of EPZ establishment across the different income

segments in treated areas, one which appears not to be distributional-neutral.

Table 4.4 below offers a more detailed assessment of the same variables. It displays the difference in

selected outcome variables between No-EPZ and EPZ municipalities up until the specific period of firm

establishment – i.e. by establishment sequence. Supporting the results of the previous tables, all four

groups display non-statistically significant differences regarding inequality indices and percentile ratios in

the period before firm establishment compared to the non-treated group. The picture is more nuanced

concerning levels of expenditure per capita. In fact, the first group shows no statistically significant

difference with respect to the non-treated group, while the subsequent groups display a similar result only

at the 90th percentile. The median and mean are the only percentile statistically significantly higher for

treated municipalities for all groups 2 to 4 prior to firm establishment, suggesting that municipalities

treated later already displayed higher levels of real per capita expenditure across the mid-range percentiles

than non-treated ones.

Above all, these results confirm that treated and non-treated municipalities had comparable within-area

inequality levels as well as similar levels of real expenditure per capita at the tails of the distribution prior to

Table 4.3 Selected Outcome Statistics for Municipalities according to EPZ status, 2009

t-test

N=

Mean SD Mean SD Mean SD

Selected Outcome Measures

Gini Coefficient 28.22 6.69 33.39 4.6 27.19 6.59 -3.8

Theil's T Index (GE1) 13.44 5.5 16.66 4.75 12.79 5.43 -2.81

Quantile Ratio 90th/10th 4.2 1.74 4.87 1.15 4.07 1.81 -1.8

Quantile Ratio 50th/10th 2.06 0.64 2.05 0.33 2.07 0.691 0.11

Quantile Ratio 90th/50th 2.05 0.56 2.4 0.57 1.98 0.53 -3.04

Expenditure per capita

10th percentile 2,921.71 906.09 3561.35 906.56 2,793.78 1,106.54 -2.76

50th percentile 5,707.41 1,905.50 7,129.75 1,305.00 5,423.01 1,883.80 -3.66

90th percentile 11,709.70 5,200.08 17,175.93 5,702.00 10,616.45 4,364.32 -5.51

Mean 6,554.00 2,228.38 8,769.49 1,637.93 6,095.62 2,057.86 -5.17

Entire SampleEPZ

Municipalities

No EPZ

Municipalities

Notes: The data is for 108 municipalities in 2009, 20 designated EPZ. Expenditure measures are expressed in constant Cordobas of 1999

(Exchange rate US$1=C$11.8). Theil's T index is a Generalized Entropy (GE) inequality measures, with α=1. PR90/10, PR50/10 and

PR90/50 correspond to percentile ratios of the 90th to 10th percentile, 50th to 10th, and 90th to 50th, respectively.

108 20 88

14

zone establishment. They reinforce the hypothesis of an inequality-enhancing effect in treated areas

resulting from EPZ institution and justify the interest of formally testing such a relationship.

In order to better understand the underlying mechanisms behind the changes in inequality within each

cluster, I look next at the average annual per capita growth rates, calculated across treated and non-treated

municipalities, at each percentile of the municipalities’ expenditure distribution. Figure 4.1 shows the

average values for the entire period (1993-2009). It is straightforward to see that expenditure at all deciles

grew over the period irrespective of zone formation. Interestingly, both groups show similar trends in that

they display higher growth rates at the bottom-end of the distribution (10th to 50th percentiles). Never-

EPZs municipalities also show higher annual growth rates at every percentile but the 90th, with the biggest

range at the 50th and 70th. When decomposing in two sub-periods (Figure 4.2) relatively similar dynamics

Timing of Treatment

A. Granting Sequence

Municipalities with newly estasblished EPZs

Total Municipalities with EPZs

B. Dispersion Measures by Group,

(for period prior to EPZ establishment)

Expenditure per capita

10th percentile

50th percentile

90th percentile

Mean 3,904

[-1.48]

4,137

[-1.84]

4,984

[-2.25]

5,650

[-2.07]

[-1.11] [-1.07] [-1.27]

8,827 7,855 8,718 9,853

35.5

[-0.16] [-0.58] [-0.24]

[-1.03] [-1.73] [-0.94]

[-0.23] [-1.32]

[-0.32] [0.04]

[-0.09] [0.33]

Group 3

[2002-2005]

Group 4

[2006-2009]

Group 1

[<1999]

Group 2

[2000-2001]

4

4 9

8

17

3

40.63 35.62

1,680

3,825

20

5

2,279

4,357

22.53

5.83

Notes: Expenditure measures are expressed in constant Cordobas of 1999 (Exchange rate US$1=C$11.8).The t-statistics

in brackets measure the statistical significance of the difference between treated and non-treated group for the selected

measure up until year of treatment, by treatment group. Values for each non-treated group are not represented. Theil's T

index is a Generalized Entropy (GE) inequality measures, with α=1. PR90/10, PR50/10 and PR90/50 correspond to

percentile ratios of the 90th to 10th percentile, 50th to 10th, and 90th to 50th, respectively.

Quantile Ratio 90th/10th

[-1.93]

[-1.52] [-2.40] [-2.19] [-1.90]

[-0.89]

Quantile Ratio 50th/10th 2.21 2.38 2.30

1,448

3,147

20.38

6.13

Theil's T Index (GE1)

1,498

3,325

34.15

17.42

4.37

[0.19]

[0.67]

[0.38]

1.93

[-1.31] [0.67]

5.44

[0.51] [0.38]Quantile Ratio 90th/50th

2.64 2.44 2.34 2.26

Table 4.4 Selected Outcome Statistics for Municipalities by Establishment Sequence

[-1.06] [-0.31] [-1.04] [-0.57]

18.12

Gini Coefficient

15

exist for the first half (1993-2001) – the period with also the higher rates of growth -, while the second half

exhibits slower rates with a lesser dominance of the No-EPZ group. In fact, during this period real per

capita expenditure grew at a higher rate for the EPZ group at the 10th and 90th percentiles.

Figure 4.1 Average Annual Growth Rates of Expenditure per capita

at Different Points of Municipalities Distribution, by EPZ Status (1993-2009)

Figure 4.2 Average Annual Growth Rates of Expenditure per capita at

Different Points of Municipalities Distribution, by EPZ Status (subperiods)

It is hard to draw conclusions from these various pictures. The higher growth rates of expenditure per

capita in never-treated municipalities throughout the period primarily impact inequality between the

groups. In this sense, while not the main focus of this analysis, a separate note on inequality between

municipalities is granted. In fact, the ratio of mean expenditure per capita between treated and non-treated

groups’ does decrease throughout the years19. Within-group inequality trends are harder to derive,

particularly as they seem to be heterogeneous in time. In this sense, a more formal analysis is needed to

infer any causal effect of the policy on within-area disparity and across the different income segments in

host municipalities.

19 The ratio of mean real expenditure per capita between treated and never-treated municipalities decreases from 1.53 in 1993, to 1.43 in 1998-2001 and finally 1.37 in 2009.

16

5. Empirical Strategy

The main empirical analysis relies on the Difference-in-Differences (DID) framework. The interest in

using this framework stems from not being able to observe what would have happened in the treated

municipalities had there not been any EPZ firm established there. For the identification strategy, I rely on

the significant variation provided by the timing of zone creation across the sample of municipalities. To

tackle the challenge regarding unobservable factors I allow for municipality and year fixed-effects, as

discussed next.

5.1 Basic Specifications

I follow Angrist and Pischke (2009) and define equation (1) as the extension of the simple DID model for

multiple periods and groups, in order to estimate the Average Treatment Effect on the Treated (ATT)

with the panel dataset. I denote EPZ as a dummy variable that switches to one when a municipality m gets

treated in year t. The basic specification takes the following form:

𝑌𝑚𝑡 = 𝛾𝑚 + 𝜇𝑡 + 𝛿𝐸𝑃𝑍𝑚𝑡 + 𝑋′𝑚𝑡𝛽 + 휀𝑚𝑡 (1)

Where m and t index for municipality (m=1, ..., 145) and year (t=1993, 1998, 2001, 2005, and 2009)

respectively, 𝑌𝑚𝑡 is the outcome variable of interest in municipality m at year t; 𝛾𝑚 are municipality-level

fixed-effects and 𝜇𝑡 are time effects. The fixed-effects capture the permanent differences in the

municipalities’ observed and unobserved characteristics that may influence the location of EPZs as well as

any time varying shocks. X’ is a vector for time-varying individual characteristics aggregated at the

municipality level. 휀𝑚𝑡 is the error term. Following Bertrand, Duflo and Mullainathan (2004) I use robust

standard errors clustered at the municipal level to account for the fact that outcomes are likely to be

correlated at this subnational level. The parameter of interest is 𝛿, which measures the average effect of the

establishment of an EPZ on the outcome Ymt (i.e. mean real expenditure per capita, inequality indices, and

decile ratios). The effect is identified by comparing municipal outcomes among treated and non-treated

groups before and after EPZ establishment.

I measure the effect across the expenditure distribution by extending the DID method to each quantile (𝜏)

instead of the mean (Athey and Imbens 2006). The specification takes the following form:

𝑌𝑚𝑡| 𝐸𝑃𝑍𝑚𝑡,𝜌𝑚𝑡 (𝜏) = 𝛾𝑚(𝜏) + 𝜇𝑡(𝜏) + 𝛿𝐸𝑃𝑍𝑚𝑡(𝜏) + 𝑋′

𝑚𝑡(𝜏)𝛽 + 휀(𝜏,𝜌𝑚𝑡) (2)

Where notations are the same as previously defined in the simple DID cases but 𝜏 now stands for each

decile of the expenditure distribution estimated at the municipality level. The robust standard errors

remain clustered at the municipal-level. This method referred to as the QDID allows comparing

individuals across both groups and time according to their quantile. Here, the quantiles are fixed and the

DID estimator compares changes in the levels of expenditure around them, under the identifying

assumption that the growth in expenditure from pre-EPZ to post-EPZ cohorts at each particular quantile

would have been the same in both treatment and comparison groups in the absence of treatment. It relies

on the assumptions of rank-preservation and homogeneity of treatment effect across individuals (Athey

and Imbens 2006).

17

5.2 Refining the validity of the DID Strategy

5.2.1 Counterfactuals

The strength of the DID approach for causal interpretation rests on the construction of an appropriate

control group that can act as a suitable counterfactual for the treated municipalities. The key assumption

behind a valid counterfactual being that the two groups’ measured outcomes would have followed the

same trends over time in the absence of policy (i.e. common trend assumption). Given that the EPZ

regime in Nicaragua exempts the zones from taxes at all subnational levels, it is rational to assume that the

choice of locating in a particular municipality remains conditioned on other observables and unobservable

characteristics measured at this spatial level. The export-oriented nature of EPZs and its human capital

needs imply that accessibility conditions (i.e. ports, airports and roads) and basic socio-economic

characteristics are likely to play an important role in the choice of location. As such, the baseline

comparison group of all never-treated municipalities may not necessarily stand as the best counterfactual.

To address this issue, I construct a set of alternative control groups using detailed information on

municipalities’ characteristics (Figure 5.1) and test their validity.

Figure 5.1 Control Groups

I base the first one on the conditions of road infrastructure. Given the limited development of Nicaraguan

ports, airports and railways, road networks remain the main mode of transport for most domestic

movements and a large share of imports and exports (World Bank, 2013). As such, good roads are not

only likely to be one key determinant for EPZs locational choices, but they may also be a proxy for

additional infrastructure conditions. I use the meters of paved roads every 1 km2 as main measure for road

accessibility, and set the threshold at the minimum in treated municipalities (40 meters). The second group

is constructed based on a distance-buffer set on the municipality’s centroid distance to the capital city.

There are several reasons behind this choice. First, as in most developing countries, the capital city is likely

Road-Network Buffer Distance-Buffer RAAN/RAAS Exclusion

Notes: GIS digital maps and World Bank (2011)

.

18

to be a key player for exports and imports, as well as the epicentre of economic activity in the country.

EPZs may have an interest in locating in its vicinity to avoid higher transport costs and take advantage of

agglomeration economies. Second, using alternative criteria for the distance-buffer would sizeable limit the

pool of municipalities. The threshold is set to the maximum distance for an EPZ municipality (<220km).

Finally, I construct a control group excluding the two autonomous regions of the Atlantic coast (RAAN

and RAAS), which correspond to the poorest areas of the country and are thus likely to have lower

infrastructure and human capital endowments. Furthermore, these areas also have very particular ethnic

and economic characteristics that differentiate them from the rest of the country.

Table 5.1 above presents summary statistics for the relevant selected outcome variables using the different

control groups for the baseline year. First, consistently with previous tables, none of the proposed

counterfactuals has statistically significant differences with the treated municipalities in terms of the

inequality measures prior to EPZ establishment. Still, there are statistically significant discrepancies

regarding the levels of real expenditure per capita across the chosen points of the distribution for all

groups. For the empirical analysis I choose to present results with the baseline control group of all non-

treated areas and the road-network control group. Both exhibit the closer statistically significant similarities

with treated municipalities for the tested outcomes variables in 1993, and together they provide a further

robustness check for the estimations. Results with the other control groups are sometimes used for

additional support. Figure 5.2 below provides visual evidence for the existence of a common underlying

trend between treatment and control municipalities using the two chosen definitions. The figures display

the trend for municipalities treated in the later sequences (2001 and 2005) to add pre-treatment years.

Table 5.1 Selected Outcome Statistics for Municipalities according to EPZ status and Control Group, 1993

N=

Mean SD Mean SD t-test. Mean SD t-test. Mean SD t-test.

Selected Outcome

Measures

Gini Coefficient 39.22 9.06 35.8 8.27 -1.29 36.30 9.63 -0.9836.37 9.4 -0.99

Theil's T Index (GE 1) 23.45 7.88 23.36 11.7 -0.03 23.03 7.88 -0.13 23.71 11.43 -0.09

Percentile Ratio 90th/10th 6.33 2.51 5.77 3.41 -0.55 5.60 3.27 -0.74 5.68 3.18 -0.69

Percentile Ratio 50th/10th 2.3 0.64 2.07 0.97 -1.08 2.10 0.69 -0.95 2.09 0.68 -1.02

Percentile Ratio 90th/50th 2.71 0.66 2.72 1.24 0.04 2.59 1.03 -0.36 2.71 0.67 0.01

Expenditure per capita

10th percentile 1,351.26 451.51 1,006.2 774.5 -1.5 931.6 523.6 -2.6 970.9 692.0 -1.9

50th percentile 3,078.41 1,200.08 1,968.6 1,292.3 -2.7 1,873.9 9,191.4 -3.9 1,927.3 1,209.4 -3.1

90th percentile 8,456.50 3,981.40 6,038.1 8,950.1 -0.9 4,875.9 3,987.6 -3.5 5,710.8 7,964.7 -1.2

Mean 4,403.47 1,908.33 2831.97 2723.29 -1.9 2,612.5 1,374.0 -3.8 2,804.6 2,485.0 -2.2

13 47 63 61

Notes: The data is for 83 municipalities in 1993, 13 ever designated EPZ. Expenditure measures are expressed in constant Cordobas

of 1999 (Exchange rate US$1=C$11.8). The t-test measure the mean differences of each control group with respect to the treated

one. Theil's T index is a Generalized Entropy (GE) inequality measures, with α=1. PR90/10, PR50/10 and PR90/50 correspond to

percentile ratios of the 90th to 10th percentile, 50th to 10th, and 90th to 50th, respectively.

Municipalities

Ever EPZ

Different Control Groups

1. Road Network Buffer 2. Distance-Buffer 3. Excludes RAAN/RAAS

19

Figure 5.2 Parallel trends across Selected Points of Real Expenditure Distribution Between Treated and Non-Treated Municipalities

20

6 The Impact of EPZ Establishment on the Inequality Dynamics of Treated Municipalities

6.2 Average Effect of EPZ Establishment

6.2.1 Mean Levels of Expenditure per capita and Inequality Measures

I start by examining the effect of EPZs establishment on the average real expenditure per capita of treated

municipalities. Table 6.1 presents the results of different DID estimates. The first two columns report the

results from equations (1), while in columns (3-4) the same specification is applied using the individual-

level datasets20.

First, the results confirm that, all else equal, the formation of an EPZ increased the average levels of real

20 To gauge the relevance of the relationship more rigorously, I use the original individual-level repeated cross-sectional dataset to estimate ATT on the levels of real per capita expenditure. These estimations permit the inclusion of a precise set of controls measured at the individual-level that are otherwise not available (such as age, individual educational achievement, household size, gender and migration behaviour). Additionally, to control even more richly for differences across municipalities, I follow Besley and Burgess (2004) and add a municipality-specific time trend in the form of interactions between a municipality dummy and a linear time trend. The larger sample allows introducing these dummies

without compromising the degrees of freedom. The equation becomes: 𝑌 𝑖𝑚𝑡 = 𝛾0𝑚 + 𝛾1𝑚𝑡 + 𝜇𝑡 + 𝛿𝐸𝑃𝑍𝑚𝑡 +𝑋′𝑖𝑚𝑡𝛽 + 휀𝑖𝑚𝑡 , where most of the parameters are the same as in equation (1) but i now indexes for individuals in municipality m at year t. Standard errors remain clustered at the municipal-level. The proximity of the results in terms of size and significance in columns (1) and (3) provide visual confirmation of the validity of using the municipal-level aggregations despite not being able to include individual controls related to income.

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

All Municipalities

601.52* 576.33* 517.51** 834.70**

(1.71) (1.67) (2.05) (2.06)

Observations 575 575 107,725 107,725

R-sq. 0.64 0.18 0.27 0.28

Road Network Buffer

498.79 427.75 459.5* 738.95**

(1.33) (1.15) (1.65) (2.05)

Observations 417 417 78,385 78,385

R-sq. 0.64 0.1 0.26 0.27

Year FE Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes

Municipality-Time Trend No No No Yes

Province-Time Trend No Yes No No

Covariates Yes Yes Yes Yes

Table 6.1 DID Estimates of the Effect of EPZs on Average Real Expenditure

per capita

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities). Covariates for

columns (3 & 4) include dummies for time-varying individual characteristics: educational level,

age square, gender, urban, sector of work, relationship with head of household and having

migrated within the past five years. In columns (1 & 2) eq. 1 uses dummies for percent of

population by educational achievement, share of urban population & migrants, at the

municipality level.

*** p<0.01, ** p<0.05, * p<0.1

EPZ

EPZ

21

expenditure per capita of individuals in treated areas over the period under consideration. The size of the

effect is relatively large in the range of C$450-840, and corresponds to around 15 to 30 percent of the

average expenditure level at the beginning of the period. It is reassuring to see that despite varying in

significance level, the results have a similar size in columns (1) and (3) across both control groups. The fact

that the coefficients are no longer significant in column (1 & 2) when using the road-network buffer to

delimit the counterfactual, suggests however that the effect might be less important when considering

similar municipalities. It might also imply a time-varying dynamic or a heterogeneous effect across the

different segments of the distribution that fails to be captured by the average estimates. Finally, it is

interesting to note that the inclusion of municipality-specific time trends in column (4) increases the size of

the coefficients, while keeping the statistical significance across both control groups. The fact that results

are robust to the inclusion of the trends is encouraging, as it suggests that the expenditure difference is not

resulting from individuals being in municipalities where expenditure levels were increasing anyway,

irrespective of the treatment21. To test whether municipal-level estimations do not suffer from any

potential negative omitted variable bias (OVB) without the inclusion of the trends, I estimate equation (1)

adding flexible province-time trends (column 2). I find results to be unchanged in terms of significance and

smaller in size, while losing significantly the number of degrees of freedom.

Table 6.2 below presents the results of the DID estimates (equation 1) on the synthetic inequality

measures constructed at the municipal-level.

While positive, there is no effect visible on the Gini coefficients irrespective of the control groups. Theil’s

T conversely does register higher inequality levels at 10 percent significance for treated municipalities

against the road-network control group (column 5). Given the particularity of the Theil index, I also run

regressions using the index in its generalized form (Generalized Entropy) with varying levels of alpha22.

21 Nicaragua has 18 counties, 2 of them being the autonomous regions. 22 In General Entropy (GE) measures, like Theil indices, this parameter represents the weight given to distances between incomes at different parts of the income distribution. They can take any real value but the most common values of α used are 0, 1, and 2, with the first two corresponding to Theil’s L and Theil’s T.

Gini Coefficient Theil Index (T) GE (2) Gini Coefficient Theil Index (T) GE (2)

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

EPZ 1.032 1.756 3.410* 1.08 2.02* 3.894*

(0.66) (1.504) (1.864) (0.678) (1.686) (1.977)

Year FE Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes

Observations 575 575 575 417 417 417

R-sq. 0.21 0.20 0.20 0.21 0.22 0.21

Number of Muni. 145 145 145 107 107 107

All Municipalities

Table 6.2 DID Estimates of the Effect of EPZs on Within-Municipality Inequality – Indices

Road Network Buffer

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities). Covariates include illiteracy rate, primary and

tertiary education, as well as share of urban population and migrants at municipal-level. Theil's T and GE (2) are

generalized entropy (GE) inequality measures with α=1 adn 2, respectively. The higher the alpha parameter the

greater the sensitivity of the index to variations in the upper-tail of the distribution.

*** p<0.01, ** p<0.05, * p<0.1

22

Interestingly, with alpha equal to 223 - which gives higher weight to changes that affect the upper tail of the

distribution - the index registers statistically significant higher inequality levels on EPZ municipalities

regardless of the choice of control group. The difference on disparity levels is relatively high, at well-above

1 point for both measures (i.e. 5 percent of 1993 level). A first observation, and one to which I return

below, is thus that following EPZ establishment, treated municipalities seem to display higher inequality

levels against non-treated ones, when captured by measures sensible to dynamics in the upper tail of the

distribution.

The results of the DID estimates on the percentile ratios in table 6.3 below support this primary

conclusion. The impact on the 90/50 ratio is consistently significant at 5 percent level and positive across

specifications (columns 3 & 6). It indicates a 0.36 to 0.38 points higher ratio for upper-end inequality in

treated municipalities across the period (i.e. near 10 percent of baseline value). Interestingly, EPZ-

municipalities also register a decline on bottom-end inequality over the period, with the 50/10 percentile

ratio negative and statistically significant. The latter displays nonetheless a lower significance, at 10 percent

level, and a smaller size (-0.2 points) with both control groups. These two opposite tendencies are notably

driving the positive but not significant 90/10 ratio, and might explain the lack of significance of the Gini

coefficient with both dynamics cancelling each other out.

The results of these first specifications suggest an ambiguous effect of EPZ establishment on overall

inequality within host municipalities. The estimates that capture a statistically significant impact seem to

indicate that a rise on inequality is driven by dynamics at the upper tail of the distribution. Upper-end

inequality is evidently higher in treated municipalities over the period considered. The trend on bottom-

end disparity is less clear with negative but lower significant values across specifications that fail to

translate into a reduction of inequality in any of the overall measures. These developments imply a larger

impact of EPZ formation on the upper-tail of the distribution. To explore formally this hypothesis, I

estimate next the treatment effect across all deciles of the real expenditure distribution of treated areas.

23 Theil’s L (α=0), like the Gini, presents a positive but not statistical significant coefficient. It has been omitted throughout to avoid redundancy.

PR90/10 PR50/10 PR90/50 PR90/10 PR50/10 PR90/50

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

EPZ 0.249 -0.204* 0.38** 0.21 -0.215* 0.365**

(0.617) (-1.919) (2.205) (0.518) (-1.889) (2.114)

Year FE Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes

Observations 575 575 575 417 417 417

R-sq. 0.10 0.05 0.10 0.11 0.05 0.13

Number of Muni. 145 145 145 107 107 107

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities) Covariates include illiteracy rate, primary and tertiary

education, as well as share of urban population and migrants at municipal-level. PR90/10, PR50/10 and PR90/50 correspond

to percentile ratios of the 90th to 10th percentile, 50th to 10th, and 90th to 50th, respectively.

*** p<0.01, ** p<0.05, * p<0.1

Table 6.3 DID Estimates of the Effect of EPZs on Within-Municipality Inequality – Percentile Ratios

All Municipalities Road Network Buffer

23

This strategy is useful to directly evaluate the heterogeneity of the effect across income segments, and

decompose the results found with the aggregated indices and ratios.

6.2.2 Percentiles of the Real Expenditure Distribution.

Table 6.4 below presents the results of DID estimates across all deciles of the real per capita expenditure

distribution of treated municipalities (QDID, equation 2). The first column is estimated using all

municipalities in the sample (i.e. the never treated municipalities as comparison group), and the second

column uses the road-network buffer to define the control group. In interpreting these results, it should be

kept in mind that I estimate treatment effect on each decile of the expenditure distribution of treated

municipalities strictly with respect to its respective decile in the non-treated group.

All Municipalities Road Network Buffer

(1) (2)

263.4* 225.8

(1.845) (1.497)

160.9 100.5

(0.997) (0.593)

175.5 109.7

(0.825) (0.503)

305.0 235.7

(1.391) (1.042)

310.0 234.4

(1.266) (0.934)

360.2 316.0

(1.162) (0.954)

512.7 428.0

(1.509) (1.208)

1,085** 991.5**

(2.347) (2.046)

2,648** 2,539**

(2.352) (2.135)

Yes Yes

Yes Yes

Yes Yes

575 417

145 107

Real Expenditure

per capita

10th Percentile

20th Percentile

30th Percentile

40th Percentile

Table 6.4 QDID Estimates of the Effect of EPZs across the

Expenditure Distribution in Treated Municipalities

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities).

Covariates include illiteracy rate, primary and tertiary education, as well as share of

urban population and migrants at municipal-level. Rsq. for both columns ranged

between 0.35-0.71. Regression in column (2) limits observations using the road-

network buffer.

*** p<0.01, ** p<0.05, * p<0.1

Municipality FE

Covariates

Observations

Number of Muni.

50th Percentile

60th Percentile

70th Percentile

80th Percentile

90th Percentile

Year FE

24

These estimates provide supportive evidence to the hypothesis of inequality dynamics being directly

related to the impact of the policy on the highest percentiles of the distribution. The most salient feature in

table 6.4 is the containment of the treatment effect at the top-end of the distribution (80th and 90th

percentiles). These are consistent across all specifications, with both deciles particularly keeping a 5 percent

significance level in both columns. The size of the impact is high at around one fourth to one third of the

level of expenditure per capita at the respective decile for all municipalities in 1993. It remains high when

comparing with the average values across the period, at 15 percent of the level in treated-municipalities

and about one quarter for both groups of control municipalities. Above all, these results suggest that the

establishment of a zone profits the top deciles of the distribution in host municipalities, which are better-

off, compared to their respective percentiles in non-treated areas.

Additionally, the coefficient at the 10th percentile also displays a positive and significant value with the

baseline control group (at 10 percent level) (column 1). The effect is lost with the most restrictive road-

network buffer24 but remains positive and close to significance. The lack of consistency in terms of

significance level across the control groups is suggestive and implies a relatively smaller effect than on the

levels of expenditure at the other two percentiles. It should be considered carefully. One interpretation of

this finding could be that EPZ benefit the poorer segments not through direct job creation but indirect

informal ones, which might manifest with lower intensity across the sample or take longer to materialize.

In contrast, while the coefficients are positive, the estimates do not register any significant effect at the 20th

to 70th percentiles. These outcomes support the ambiguous effect of EPZ establishment found on overall

inequality measures, with the mid-percentiles contributing to smooth any impact of treatment registered at

the tails.

Taking everything into account, the primary conclusions from estimates in tables 6.2 to 6.4 remain that

contrarily to what might be naively expected from this type of policies, the establishment of an EPZ in a

given municipality primarily increases the levels of real expenditure per capita at the upper-end of the

distribution. Specifically, compared to their peers in non-treated areas, the high-income segments are

relatively better-off in treated municipalities, making them the first beneficiaries of the policy. Interestingly,

this effect is not translated into a definite increase of overall within-municipality inequality in treated areas.

As seen, the coefficients for the Gini and Theil indices are positive but with no (or little) statistical

significance. Decomposing inequality helps to shed light on a more nuanced picture. Mainly, as a direct

consequence of the treatment effect at the upper-end percentiles, top-end inequality clearly increases in

treated areas over the period. Additionally, the estimates with the percentile ratio 50/10 are statistically

significant and indicate a reduction on the bottom-end inequality of treated municipalities. Still, this effect

is smaller and less consistent throughout control groups, surely explaining the lack of an impact on overall

inequality dynamics, and resulting from the smaller, if significant, increase at the poorest decile of the

distribution in treated areas.

These first conclusions are interesting in that they provide an empirical underpinning for the expectations

of the more advanced HO models. In fact, two general explanations could be behind these findings. First,

the concentration of the effect at the upper-end of the expenditure distribution may indicate that the

owners of the factors of production are the ones actually benefitting from zone establishment, either

through direct investment or land ownership. This is easy to imagine in a country with already high levels

of inequality. Second, the results could also be the consequence of the existence of a skill premium for

white-collar workers in the EPZ sector. In line with some of the theoretical predictions, the skill premium

24 As a robustness check I also run the QDID with the distance-buffer and RAAN/RAAS exclusion control groups. Both estimations display similar coefficients in terms of size and sign. The top deciles show similar statistical significance, while the lowest one is only significant against the RAAN/RAAS exclusion group. See Appendix 2.

25

might stem from a higher skilled-to-unskilled labor ratio, as well as from higher productivity levels in the

exporting sector together with limited labor mobility. These hypotheses fit remarkably well the Nicaraguan

case. As previously described, the labor market is characterized by low levels of educational achievement, a

de facto constraint to labor mobility. At the same time, productivity levels have increased markedly in the

manufacturing sector over the past years, notably because of the exporting EPZ sector. These dynamics

would imply that the skill intensity in the zones is higher than in the rest of the manufacturing sector,

interestingly supporting the predictions of Harrison and Hanson (1999) and Verhoogen (2006). Strangely,

one would expect that mid-range percentiles could also benefit from this skill premium. The fact that they

do not seem to register any significant treatment effect could result from a segmentation of the high-skill

labor sector. A final consideration should be given to the effect captured at the 10th percentile of the

distribution. If indeed positive and significant, it is reasonable to assume that the effect is driven from

indirect spillovers or trickle down effects, possibly in the informal sector from services provided to EPZs.

These findings would be supportive of polarization theories. Still, more information on the skill

composition of the EPZ job market in Nicaragua would be needed to reach any further definite

conclusion.

6.3 Spillover Dynamics & Commuting

In applying the DID framework to the data, it is important to consider carefully the ‘experiment’ created

by the establishment of a zone in a given municipality. In the ideal case, the decisions would be

independent random events that varied in timing and had no spillover effects to non-treated areas.

Additionally, there would be no interaction between treated and non-treated areas. While the empirical

strategy already controls for the major sources of endogeneity (i.e. time-varying shocks, selection-bias) the

present analysis may still differ from the ideal case. This section addresses the concerns related to possible

interactions taking place between an EPZ municipality and its neighbouring ones, either through spillover

dynamics or commuting patterns. If existent, these mechanisms are expected to have a reasonable impact

on the expenditure per capita of neighbouring areas. Not taking them into consideration would lead to a

bias estimation of the previous EPZ effects.

To begin with, I define spillovers as any form of business diversion or creation that could take place

because of agglomeration forces between treated and non-treated municipalities adjacent to each other.

Backward linkages between EPZs and local producers are probably the main mechanism through which

spillovers can be expected to happen. In this sense, I conduct a preliminary test using the individual-level

repeated cross-sectional dataset to measure the impact of EPZ establishment on the likelihood of being

self-employed (as a proxy for business start-ups25) amongst the working age population. I find no evidence

of a statistically significant effect26 with both logit and OLS estimators. While marginally positive the

coefficients are almost no different from zero, which is consistent with findings in the literature (Farole

and Akinci 2011). Even though these results are relatively reassuring for the quasi-experimental setting

created here, spillovers to adjacent municipalities could still be taking place in alternative ways and the

intensity could be higher when considering the spatial dimension.

Additionally, I consider commuting behaviour as form of spillover to adjacent areas, given that individuals

working in a treated municipality but residing in a neighbouring area are likely to have a similar impact in

terms of consumption (or simply in accounting for expenditure levels), than business creation/diversion in

the locality where they reside. Commuting is hard to measure. Information is only available for 2001, when

25 This is in line with the related literature. 26 See Appendix 2.

26

it represented 3.56 percent of the total sample (5.05 and 3.02 percent in treated and never-treated

municipalities, respectively). Reassuringly, the number seems sufficiently small not to alter the results.

Next, I use a distance-definition to consider them in the formal analysis.

In effect, in order to account for the possibility of the above-mentioned dynamics, I measure treatment

effect across the real expenditure distribution of municipalities adjacent to a treated area, using different

levels of proximity. For this, I compute a modified version of equations (1 & 2) adding a set of dummy

variables that switch to one for municipalities neighbouring a treated area, at two different levels of

proximity. Because of the particular small size of Nicaraguan municipalities in the west part of the country,

considering the shared border as criteria for proximity would imply denoting the vast majority of non-

treated municipalities in the region as neighbors. Instead, I use a municipalities’ centroid distance to the

centroid of the nearest treated one with two levels of distance, one set at between 5 to 15km and the other

at 15 to 35km27. This differentiation additionally allows accounting for both dimensions, commuting with

the closer group, and more formal spillover effects with the one further away.

Table 6.5 provides the QDID estimates augmented with the neighbor dummies across the deciles of the

real per capita expenditure distribution. For simplicity, I show the tables with the road-network control

group only28. It is encouraging to find a close similarity in the size, sign and statistical significance of the

EPZ coefficients to the ones in previous estimations. Above all, these results further emphasize that

individuals at the top-end of the distribution benefit the most from the establishment of an EPZ in the

municipality where they reside throughout the period.

Regarding the coefficients for neighboring areas, they seem to ratify the previous suspicion that spillover

effects are in fact limited. None of the coefficients for neighboring municipalities display statistical

significance, implying there are indeed no spillovers from EPZ establishment on the levels of expenditure

per capita across the deciles of the distribution in adjacent municipalities. Truthfully, it could also mean

that they take longer to materialize and have not yet done so in the period studied. The pattern deserves a

special mention. Particularly, while not significant, the effect is negative at the bottom deciles in

municipalities situated at 5 to 15km from a treated area. This suggests that the poorer segments in small

municipalities adjacent to a treated area are worse-off than those in host municipalities as well as in zones

further away. One possible explanation could be related to business diversion in the informal sector

towards treated areas together with a higher cost of commuting for the poor. Conversely, coefficients for

the upper-deciles are consistently positive and the largest for all degrees of proximity. They decline in size

with distance, but none is statistically significant.

Overall, I find little evidence of spillover effects on the levels of expenditure per capita in neighbouring

areas as a result of EPZ establishment in a given municipality. While interesting for themselves, these

outcomes support previous findings and confirm that bias from these phenomena is likely close to zero.

27 The measure is of course imperfect. While using the centroid’s distance offers the advantage of reducing the noise from larger municipalities where a large proportion would not be in contact with treated areas, it might also exclude some bordering zones. Smaller units of analysis and geocoding would be needed for a more precise measure. For simplicity, I refer to the measure as the direct distance between municipalities, but it always registers the distance between the centroids of two municipalities. 28 The results are consistent across different control groups. The only exception being a positive and significant coefficient at the 10 percent level for the 1st decile in treated municipalities using all never-treated municipalities as control group. Coefficients share a similar size; see Appendix 4.

27

10th

percentile

20th

percentile

30th

percentile

40th

percentile

50th

percentile

60th

percentile

70th

percentile

80th

percentile

90th

percentile

(1) (2) (3) (4) (5) (6) (7) (8) (9)

EPZ 216.8 77.22 79.95 244.2 237.7 361.7 465.3 1,052** 2,562**

(1.374) (0.446) (0.368) (1.058) (0.929) (1.008) (1.231) (2.103) (2.109)

Neighboring Municipalities1 -157 -337 -313.3 135.3 26.71 138.8 326.5 733.6 532.7

(5 to 15km) (-0.550) (-1.247) (-0.931) (0.411) (0.091) (0.422) (0.743) (1.364) (0.884)

Neighboring Municipalities2 28.88 6.396 2.038 59.2 31.82 211.5 111.2 240.6 336.9

(15 to 35km) (0.185) (0.032) (0.007) (0.207) (0.105) (0.616) (0.319) (0.552) (0.462)

Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 417 417 417 417 417 417 417 417 417

R-sq. 0.635 0.686 0.675 0.708 0.707 0.605 0.602 0.587 0.344

Number of Muni. 107 107 107 107 107 107 107 107 107

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities). Covariates include dummies for illiteracy rate, primary and tertiary education. as well as share of urban

population at municipality level. The variables Neighbor Municipalities are dummy variables that equal one if a municipality's centroid is situated at 5 to 15km, and 15km to 35km

distance from a treated municipality, and zero otherwise. I use the centroid's distance to avoid noise from larger municipalities. Observations are limited by Road-Network buffer.

*** p<0.01, ** p<0.05, * p<0.1

Table 6.5 The Effect of EPZs Establishment & Spillover Dynamics

Road Network Buffer

28

6.3 Extended Results. The Heterogeneous Effect of EPZ Establishment

The subsequent empirical section of this paper explores whether the effect of EPZs on local

distributional dynamics is heterogeneous in time. It has actually been extensively documented in the

literature that the impact may take time to materialize. In some cases it has also been found to diminish

with the years, as low labor costs vanish. Evaluating the causal effects of EPZ establishment as they

grow or fade in time across the different segments of the expenditure distribution provides an

opportunity to capture with more detail the mechanisms through which the policy affects income

distribution.

Hence, to explore the time pattern of EPZs establishment on within-municipality inequality, I compute

a modified version of equation (1) on municipal-level outcomes Ymt, adding a set of dummy variables

for 3 leads or post-treatment years (Autor, 2003) in the following form:

𝑌𝑚𝑡 = 𝛾𝑚 + 𝜇𝑡 + 𝐸𝑃𝑍𝑚𝑡0+ ∑ 𝛿+𝜋

3𝜋=1 . 𝐸𝑃𝑍𝑚,𝑡+𝜋 + 𝑋′𝑚𝑡𝛽 + 휀𝑚𝑡 (3)

Where the sum in the right-hand side allows for 3 leads (𝛿+4+, … , +𝛿+𝜋) or post-treatment years. The

dummies are equal to one in only one year each, per treated municipality29. The treatment dummy

𝐸𝑃𝑍𝑚𝑡0 equals one only on the year of EPZ establishment, or year zero. All other variables are as

previously specified. QDID Equation (2) is augmented in the same fashion.

Table 6.6 provides the estimates augmented with the leads, using the various inequality measures as left-

hand side variables. For simplicity, I show the tables with the road-network control group only30. The

first three columns exhibit the results for the different within-municipality inequality indices. The

coefficients are all positive and show consistent increment with post-treatment years. Particularly, it is

noticeable a high-jump in inequality after 11 years of EPZ operation in a host municipality. Still, the

incremental time dimension registered here should be considered carefully. The fact that the Gini

coefficient or Theil index are not significant, or that the GE index is only significant for one period,

supports the hypothesis of nuanced dynamics on the different segments of the distribution throughout

the years.

Estimates with the percentile ratios display similar, if more discernable, trends (columns 4 to 6). In fact,

coefficients for the 90/50 percentile ratio are significant and positive throughout almost all post-

treatment years, supporting previous conclusions, and suggesting a growing effect on upper-end

inequality throughout the period. Indeed, here too I observe a constant increment with a larger size after

11 years of EPZ establishment. Interestingly, the coefficient is already significant at 5 percent level for

year zero. Just as with average estimations, the pattern on bottom-end inequality displays a slightly more

nuanced behavior. First, consistent with previous findings, there is a decrease in bottom-end inequality

of treated areas, which persist over time. Yet, while all coefficients are negative, the pattern of reduction

is not harmoniously incremental. In fact, the decline is higher for years 4 and 11, and only the first

period registers statistical significance. This suggests a varying effect of EPZ establishment at the lower-

end of the distribution. A final observation is granted regarding the fluctuation registered by the 90/10

29 I compute leads for +4 years, +8 years, and +11 years of EPZ establishment. +4 and +8 years also include municipalities with +3 and +7 years, respectively, in order to avoid noise from a lead with a few municipalities. Each lead contains 4 to 16 treated municipalities. 30 Results using the alternative control group show similar coefficients in terms of significance and size; see Appendix 5.

29

percentile ratio. While never statistically significant, the signs of the coefficients alternate every other

period of post-treatment years. They are negative for post-treatment years 4 and 11, and positive

otherwise. Undoubtedly, this reflects the heterogeneous time dynamics of the policy across the lower-

end of the distribution in treated areas, and provides further evidence of the nuanced impact of EPZ

formation on overall within-group inequality.

Table 6.7 below shows the results of the QDID estimates on all the percentiles of the expenditure

distribution of treated municipalities by post-treatment years. Four remarks deserve mention. First, the

positive sign and relatively large size of most of the coefficient seem to indicate a consistent incremental

effect on the level of real expenditure per capita at all percentiles of the distribution with the more years

of EPZ operation. Yet, not all coefficients display statistically significance. Particularly, no coefficient is

statistically significant for year zero. This is consistent with findings in the literature reporting a delay on

the repercussion of this type of policy, which is likely to be more pronounced when measured on real

per capita expenditure levels.

Second, the higher effect happens at the very upper-tail of the distribution, notably at the 80th to 90th

percentiles. These segments are the only ones that benefit from the establishment of a zone in the near-

term (after 4 years of establishment). The constant statistically significant impact starting from year four

implies a growing absolute effect on the expenditure per capita at the end of the upper-tail. Evidently,

this trend explains the higher upper-end inequality levels, even despite statistically significant increases in

the expenditure per capita at the 50th to 70th percentiles after 8 years of EPZ operation. This dynamic is

also the main driver of the average effect registered earlier, displaying statistical significance for the end

Gini CoefficientTheil Index (T) GE (2) PR90/10 PR50/10 PR90/50

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

EPZ establishment, t0 1.51 1.27 2.38 0.34 -0.16 0.35*

(1.00) (1.141) (1.332) (0.879) (-1.353) (1.928)

EPZ establishment, t+4 0.57 2.33 4.89* -0.05 -0.30** 0.33*

(0.293) (1.475) (1.949) (-0.111) (-2.289) (1.783)

EPZ establishment, t+8 1.584 2.17 3.84 0.190 -0.12 0.32

(1.089) (1.155) (1.295) (0.261) (-0.708) (1.493)

EPZ establishment, t+11 2.39 5.31 9.75 -0.30 -0.40 0.37**

(0.653) (1.310) (1.572) (-0.450) (-1.577) (2.115)

Year FE Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes

Observations 417 417 417 417 417 417

R-sq. 0.207 0.222 0.211 0.105 0.062 0.132

Number of Muni. 107 107 107 107 107 107

Table 6.6 The Time-Effect of EPZs Establishment on Within-Municipality Inequality Measures

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities). Covariates include illiteracy rate, primary and tertiary education

at municipality level. GE and Theil's T indexes are generalized entropy (GE) inequality measures. They respectively have α=1 and α=2.

The higher the alpha parameter the greater the sensitivity of the index to variations in the upper-tail of the distribution. PR90/10,

PR50/10 and PR90/50 correspond to percentile ratios of the 90th to 10th percentile, 50th to 10th, and 90th to 50th, respectively. All

leads are equal to one in only a year each per treated municipality. EPZ dummy equals one only for year of establishment. EPZ t+4

includes t+3, and EPZ t+8 includes t+7, to have balanced dummies. Include observations with Road-Network Buffer.

*** p<0.01, ** p<0.05, * p<0.1

30

of the upper-tail only.

Third, and one of the most interesting revelations of this time decomposition, is that the lower-end of

the distribution (10th) and mid-range percentiles (40th to 60th) start benefiting from EPZ establishment

after 8 years of EPZ operation. Different explanations are conceivable here. First, the trend could result

from the existence of backward linkages with local communities in the form of outsourcing to small and

medium-size firms. The 8 year delay is consistent with the time it would take to materialize this type of

relationship. Caution is warranted given that backward linkages have usually been found to be small in

EPZ evaluations, and I have found no evidence of an EPZ effect on the likelihood of being self-

employed. Still, the time lag necessary to measure this type of relationships, particularly on real

expenditure levels, may explain the difficulty in registering any average effect in previous estimations or

across the timeline considered in this study. It is interesting to note the local dimension of the possible

backward linkages, as I found no evidence of spillovers to neighbouring municipalities. Additionally,

more classic trickle down effects might also be taking place, resulting from both medium-term

consumption and production dynamics after the expansion of the zones or increases in productivity

over the years.

Finally, the significance registered at the 10th percentile after 8, and particularly 11 years of EPZ

establishment deserves a special mention. It is the only percentile at the very bottom of the distribution

that registers a significant positive and accumulative effect of EPZ formation for the last two periods.

This suggests that EPZ policies need time before any positive effect reaches the poorer segments. The

fact that a significant impact is registered at the same time than for middle range percentiles could also

imply it arises from indirect trickle down effects. Interestingly, I do not find any significant effect at the

20th and 30th percentiles, implying that only the very bottom of the distribution ultimately benefits from

EPZ establishment. Further disentangling the drivers behind these heterogeneous time dynamics should

be considered for further exploration but are beyond the scope of this work.

While findings of this section are consistent and reinforce previous results, they also shed light on the

more complex inequality dynamics of the EPZ policy. On the one hand, they highlight the

predominance of the effect at the upper-end of the distribution, particularly the very top percentiles.

This salient feature is reflected in the significant and accumulative increase of upper-end inequality

throughout the years. Yet, the time decomposition also reveals a diffusion of the treatment effect at the

middle of the distribution and 10th percentile, after 8 years of zone establishment. These heterogeneous

time dynamics of EPZ establishment for the bottom and mid percentiles are certainly behind the

ambiguous overall impact of the zones on within-municipalities inequality. Particularly, they explain the

changing significance level for bottom-end inequality. These findings are interesting in that they provide

a time nuance to the local distributional impact of EPZ establishment, which evolve to be less inequality

reinforcing with the more years of operation.

31

10th

percentile

20th

percentile

30th

percentile

40th

percentile

50th

percentile

60th

percentile

70th

percentile

80th

percentile

90th

percentile

(1) (2) (3) (4) (5) (6) (7) (8) (9)

EPZ establishment, t0 108.6 -40.91 40.37 108.2 51.98 196.2 408.5 671.3 2,210

(0.736) (-0.269) (0.271) (0.560) (0.219) (0.642) (0.958) (1.362) (1.604)

EPZ establishment, t+4 374.9 243.2 142.9 325.0 310.5 206.6 170.1 972.0* 2,354*

(1.426) (0.828) (0.383) (0.943) (0.916) (0.477) (0.374) (1.717) (1.808)

EPZ establishment, t+8 333.4* 411.7 362.5 601.2** 987.3** 1,171*** 1,102*** 2,228*** 3,860***

(1.948) (1.538) (1.236) (2.358) (2.436) (2.660) (2.771) (2.832) (3.813)

EPZ establishment, t+11 876.2*** 470.9 679.9 850.7* 1,174*** 1,150*** 1,235** 1,851*** 5,088***

(3.006) (1.223) (1.264) (1.721) (3.013) (2.725) (2.224) (2.893) (4.863)

Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 417 417 417 417 417 417 417 417 417

R-sq. 0.639 0.688 0.677 0.711 0.712 0.608 0.604 0.592 0.352

Number of Muni. 107 107 107 107 107 107 107 107 107

Notes: Uses Road-Network Buffer to delimit control group. Robust t-statistics in parentheses, Clustered SE ( municipalities). Covariates include illiteracy rate, primary and tertiary

education, as well as the share of urban population and migrants at the municipality level. All leads are equal to one in only year each per adopting state. EPZ dummy equals one only

for year of establishment. EPZ t+4 includes t+3, and EPZ t+8 includes t+7, to have balanced dummies.

*** p<0.01, ** p<0.05, * p<0.1

Table 6.7 The Heterogeneous Time-Effect of EPZ Establishment

Road Network Buffer

32

7. Robustness Checks

The final empirical section contains some alternative specifications and identification strategies that

examine the robustness of the conclusions and address remaining concerns of people’s mobility across

areas. In general, the results from the various checks are quite similar to the main estimates.

7.1 Quantile Treatment Effect (QTE)

The main results in this paper regarding the effect of EPZ establishment across the expenditure

distribution of treated municipalities are obtained by applying the DID to each quantile instead of the

mean (QDID). As discussed by Athey and Imbens (2006), this method permits measuring the treatment

effect by comparing individuals across both groups and time according to their quantile, under the

assumptions of rank-preservation and homogeneity of treatment across individuals.

To probe further these results, I next use the repeated cross-sectional individual level dataset to estimate

unconditional quantile treatment effects on the treated following Firpo (2007, et al. 2009), and controlling for

unobserved time and group effects. The two methods are similar but rely on different dimensions of the

distributions in the identifying assumptions31. Additionally, compared to commonly-estimated conditional

QTE, the added value of this method is that it does not rely on any parametric functional form

assumption and results do not change dependent on covariates (unconditional)32. In the absence of a valid

instrument, I use selection on observables but include time and province-level fixed effects to take care as

much as possible of unobservable characteristics.

I define the unconditional QTE on the treated for quantile 𝜏 by:

∆𝜏= 𝑄𝑌1𝜏 − 𝑄𝑌0

𝜏 (4)

Where in the presence of a binary treatment D ∈ {1, 0}, 𝑌𝑖1 would be realised if individual i were to receive

treatment 1 (i.e. EPZ establishment in the municipality where she resides), and 𝑌𝑖0 would be realised

otherwise. The method proposed by Firpo (2007) uses a quantile regressor estimator with weights in two

steps: first, the probability of treatment is estimated nonparametrically33 and second, estimators are

calculated as solutions of two separate minimization problems from observable data. The final estimators

are equal to the differences between respective quantiles for treated and non-treated municipalities

adjusting for participation probability, and expressed as solutions of minimization problems, where the

minimands are sums of check functions34 (Firpo, 2007). The method is based on the assumptions of

selection on observables and common support.

31 The QDID holds the quantile fixed and compares changes in the levels of expenditure around this quantile, under the identifying assumption that the growth in expenditure from pre-EPZ to post-EPZ cohorts at this particular quantile would have been the same in both treatment and comparison groups in the absence of treatment. UQTE compares changes in the distribution around a quantile, holding the level of earnings fixed. The identifying assumption coming down to the expenditure growth from pre-EPZ to post-EPZ cohorts around a particular expenditure level being the same in the treatment group as in the comparison group in the absence of policy. See Havnes and Mogstad (2010, 2014) for detailed discussion. 32 For a detailed discussion see Firpo (2007), Firpo et al. (2009) and Frolich and Melly (2010). 33 The first step is a nonparametric estimation of a propensity score to obtain the weighting33 estimator related to the conditional probability for a municipality of hosting an EPZ. The weights are introduced to correct for differences in the distribution of observable characteristics between treated and control groups, with the aim of comparing strictly similar municipalities. 34 As defined by Angrist and Pischke (2009) by: 𝑄𝜏(𝑌𝑖|𝑋𝑖) = arg min 𝐸[𝜌𝜏(𝑌𝑖 − 𝑞(𝑋𝑖)) ], where 𝜌𝜏(𝑢) =

(𝜏 − 1(𝑢 ≤ 0))𝑢 is called the check function. Quantile regression achieves the minimization by substituting a linear

model for q (Xi).

33

The first step weighting estimator of Firpo (2007) uses D as its own instrument and is therefore expressed

as:

(�̂�, ∆̂𝜏) = arg min

𝛼, ∆∑ 𝑊1

𝐹 × 𝜌𝜏(𝑌𝑖 − 𝛼 − 𝐷𝑖∆) (5)

𝑊𝑖𝐹 =

𝐷𝑖

Pr (𝐷 = 1|𝑋𝑖)+

1 − 𝐷𝑖

1 − Pr (𝐷 = 1|𝑋𝑖)

I use the method developed by Frolich and Melly (2010) that uses local logit estimator as preliminary

estimator Pr (D=1|Xi) for implementing (7), and allows standard errors to be robust for

heteroskedasticity. As mentioned, I include time and province-level fixed effects to take care as much as

possible of unobservable characteristics. Table 7.1 below presents the results across the real expenditure

distribution of treated municipalities.

50th Percentile

60th Percentile

70th Percentile

80th Percentile

90th Percentile

Real Expenditure per capita

10th Percentile

20th Percentile

30th Percentile

(-0.385)

104.6

(0.742)

Notes: z-statistics in parentheses. Propensity score and variance estimated by local

logit regression with h=0.2, 0.5, 1 and lambda=1. Estimations include year and

county fixed effect, as well as the following control variables: age sq., sector of

employment, number of household members; dummies for urban residence,

gender, educational level (illiteracy to terciary level), and migrant status over last 5

years. It also includes dummies for municipal-level infrastructural conditions (paved

roads for every 1km2 at threshold, distance to capital city at threshold, landlocked).

*** p<0.01, ** p<0.05, * p<0.1

(1.631)

146.9

(1.083)

219.0

(1.493)

261.0

40th Percentile

Observations 90,413

Table 7.1 Unconditional QTE Estimates of the Effect of EPZs across

the Expenditure Distribution of Treated Municipalities

1,273***

906.5***

(3.847)

(3.544)

(2.261)

448.7**

(1)

-100.6

(-0.889)

-58.66

34

Reassuringly, the results of the unconditional QTE are relatively close to the main QDID estimates.

Mostly, they suggest fairly robust findings regarding the concentration of the main effect of EPZ

establishment at the very upper-tail of the distribution. There is now a smaller (i.e. half of the previous

estimate at the last decile) but more significant positive effect on the level of real expenditure per capita for

individuals at the 80th and 90th percentiles. Additionally, the coefficient at the 70th percentile is now

significant at 5 percent level. Coefficients for the 30th to 60th percentiles remain positive and not

significant, as before. A major difference nonetheless with respect to the previous results, concerns the

bottom-end of the distribution. While not significant, coefficients for the 10th and 20th percentile have now

a negative sign, implying that the poorer deciles would be worst-off in treated municipalities compared to

their peers in non-treated ones. This dynamic would imply a bigger impact on overall inequality, with a

larger difference between the tails. One explanation for this difference is that Firpo’s estimator computes

marginal quantiles using a weighting and matching estimator in the first step regression. As such, the latter

might be putting a higher weight at the difference generated by the strong positive treatment effect at the

upper-tail of the distribution (Firpo et al. 2009).

Overall, these estimations confirm the robustness of the main findings through the QDID. While the size

of the effect is discussable, there is little doubt about a major effect happening at the very upper-end of the

distribution.

7.2 Balanced Panel & Weight of Capital City

Given the unbalanced nature of the panel and the fact that the identification strategy relies on the timing

and cross-section variation of EPZ establishment across municipalities, I next address the possibility that

my results are driven by the inclusion of a particular observation as treated or non-treated. Largely,

conclusions remain robust to the different changes in the sample size.

For this, I start by limiting the sample to municipalities observed for at least 4 of the 5 periods considered.

For brevity I only report results of treatment effect across the deciles of the expenditure distribution of

treated municipalities (table 7.2). Reassuringly, the change does not alter previous conclusions. There is a

marginal enlarging effect with a relatively higher size on the significant coefficients, and estimations against

all never-treated municipalities capture a significant treatment effect at the 70th percentile. Results on

inequality measures, as well as in specifications checking for spillover dynamics and time heterogeneity

remain unchanged.

Next, I test that the capital city (and municipality) of Managua is not pooling the results. The

agglomeration remains by far the largest in the country in terms of economic output and population, with

its average level of real expenditure per capita being twice the average of all other municipalities in the

country. The range is almost as high across all the decile of the expenditure distribution. Managua is also

the only municipality treated during the entire period. Table 7.3 reports the QDID estimates. Again, I find

that results are not altered, providing support to the main findings of this paper. Conclusions also hold for

the inequality measures, as well as in specifications checking for spillover dynamics and time heterogeneity.

Given the special weight the longer treated municipality may have on the different time leads, I report

results of the heterogeneous EPZ time effect across the expenditure distribution on table 7.435. As

mentioned, overall conclusions remain unchanged. It is interesting to note the positive and significant

effect now captured for individuals at the 90th percentile on year zero, as well as the fact that the

coefficient at the 1st decile is now only significant strictly after 11 years of EPZ establishment. There seems

to be a marginal attenuating effect in most of the coefficients, particularly below the 80th percentile.

35 See Appendix 6 for results with alternative control group.

35

All

Municipalities

Road Network

Buffer

All

Municipalities

Road Network

Buffer

(1) (2) (1) (2)

302.6** 250.0 266.4* 229.5

(2.087) (1.645) (1.852) (1.510)

216.0 146.8 160.4 106.8

(1.376) (0.901) (0.985) (0.625)

227.2 142.4 169.9 112.3

(1.093) (0.67) (0.789) (0.508)

364.7 291.8 299.4 238.4

(1.651) (1.285) (1.343) (1.034)

392.5 301.2 315.0 246.3

(1.601) (1.206) (1.273) (0.969)

490.4 437.7 363.0 334.8

(1.624) (1.356) (1.15) (0.993)

623.8* 523.0 518.3 447.6

(1.863) (1.486) (1.513) (1.246)

1,163** 1,050** 1,075** 985.5**

(2.474) (2.139) (2.322) (2.029)

2,767** 2,522** 2,614** 2,505**

(2.387) (2.055) (2.338) (2.124)

Yes Yes Yes Yes

Yes Yes Yes Yes

Yes Yes Yes Yes

486 359 570 412

107 80 144 106

Observations

Number of Muni.

Notes: Robust t-statistics in parentheses, Clustered SE

(municipalities). Covariates include illiteracy rate, primary and tertiary

education, as well as share of urban population and migrants at

municipal-level. Rsq. for both columns ranged between 0.35-0.71.

Regression in column (2) delimits observations using the road-

network buffer.

*** p<0.01, ** p<0.05, * p<0.1

70th Percentile

80th Percentile

90th Percentile

Year FE

Municipality FE

Covariates

Table 7.3 The Effect of EPZs across the Expenditure

Distribution in Treated Municipalities - Excl. Managua

Real Expenditure per

capita

10th Percentile

20th Percentile

30th Percentile

40th Percentile

50th Percentile

60th Percentile

Municipality FE

Covariates

Observations

Number of Muni.

Notes: Robust t-statistics in parentheses, Clustered SE

(municipalities). Covariates include illiteracy rate, primary and tertiary

education, as well as share of urban population and migrants at

municipal-level. Rsq. for both columns ranged between 0.38-0.71.

Regression in column (2) delimits observations using the road-

network buffer.

*** p<0.01, ** p<0.05, * p<0.1

Year FE

Table 7.2 The Effect of EPZs across the Expenditure

Distribution in Treated Municipalities - Balanced Panel

Real Expenditure per

capita

10th Percentile

20th Percentile

30th Percentile

40th Percentile

50th Percentile

60th Percentile

70th Percentile

80th Percentile

90th Percentile

36

10th

percentile

20th

percentile

30th

percentile

40th

percentile

50th

percentile

60th

percentile

70th

percentile

80th

percentile

90th

percentile

(1) (2) (3) (4) (5) (6) (7) (8) (9)

EPZ establishment, t0 147.9 -11.91 87.17 163.2 113.3 287.3 525.7 821.4 2,533*

(0.961) (-0.07) (0.571) (0.809) (0.461) (0.897) (1.190) (1.607) (1.783)

EPZ establishment, t+4 396.4 268.4 174.3 367.0 378.3 286.5 231.3 1,093* 2,565*

(1.411) (0.851) (0.434) (0.997) (1.053) (0.628) (0.474) (1.834) (1.871)

EPZ establishment, t+8 267.2 314.0 219.7 422.7* 868.9* 1,083** 1,030** 2,291** 3,652***

(1.412) (1.078) (0.714) (1.754) (1.886) (2.120) (2.156) (2.499) (3.297)

EPZ establishment, t+11 845.0** 310.7 517.0 601.3 1,004* 975.1* 922.1 1,662** 5,371***

(2.083) (0.608) (0.688) (0.909) (1.890) (1.774) (1.267) (2.030) (4.230)

Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 412 412 412 412 412 412 412 412 412

R-sq. 0.639 0.688 0.677 0.711 0.712 0.608 0.604 0.592 0.352

Number of Muni. 106 106 106 106 106 106 106 106 106

Notes: Uses Road-Network Buffer to delimit control group. Robust t-statistics in parentheses, Clustered SE (municipalities). Covariates include illiteracy rate, primary and tertiary

education, as well as the share of urban population and migrants at the municipality level. All leads are equal to one in only year each per adopting state. EPZ dummy equals one only

for year of establishment. EPZ t+4 includes t+3, and EPZ t+8 includes t+7, to have balanced dummies.

*** p<0.01, ** p<0.05, * p<0.1

Table 7.4 The Heterogeneous Time-Effect of EPZ Establishment - Excl. Managua

Road Network Buffer

37

7.3 Mobility Issues

A final concern that remains with regards to the ‘experiment’ created in this paper relates to possible bias

from people’s mobility across areas during the period studied. Information available at the individual-level

for some of the household surveys would allow controlling for some degree of within-country migration in

specifications using the repeated cross-sectional dataset. As mentioned, I add an aggregated measure as

control for municipal-level estimations. However, as noted by Angrist and Piscke (2009), time-varying

variables measured at the aggregated level are usually affected by treatment and can introduce endogeneity.

A potential pitfall thus remains. To understand the magnitude, I use the information available in the

repeated-cross sectional dataset on individual migration across municipalities.

During the entire period, only 10.7 percent of the sample reports having moved in the previous five years

to another municipality, with a higher proportion in never-treated municipalities (12.3 percent). The flow

however seems to have been under-measured in 1998 and 2009, both years displaying very low

percentages (2-5 percent). Excluding them, 16.7 percent of the remaining sample reports having migrated

in the 5 years prior to the survey, and more relevant, only 9.1 percent of the working age population (9.6,

7.3 and 7.7 in never-treated, treated and road-network buffer municipalities, respectively). The numbers on

overall within-country migration are relatively small, and mirror the mentioned characteristics of

Nicaragua’s society throughout the period (i.e. relatively low levels of urban population, declining

urbanization rates, and labour market segmentation). There is limited information regarding origin and

destination of the flows, with data available only for 1998. For that year, 7.8 percent of the migrants

moved to a municipality hosting an EPZ. This partial information seems to indicate limited internal

migration from non-treated areas towards treated ones, with most flows taking place among non-treated

areas. To test formally for displacement resulting from EPZ establishment I estimate the likelihood of

migrating for the working age population following the establishment of a zone in a given area using both

OLS and logit estimators36. I find no evidence of a statistically significant impact. Furthermore, the sign on

the coefficients are consistently negative, implying than if anything, EPZs would deter emigration flows.

These results together with the relatively small number of migrants in the sample provide support to

discard major bias from displacement. As noted by Topalova (2010) in a similar context, the very findings

of disparate effects (of EPZ establishment) across space suggest absence of perfect factor mobility across

the country.

8. Conclusion

The present analysis has exploited a rich dataset at both individual and municipal levels to estimate the

distributional effects of EPZ establishment within their host municipalities in Nicaragua for the period

1993-2009. Evidence amassed is robust to the different specifications and identification strategies used.

Above all, I find robust evidence that EPZ establishment in a given municipality increases the average

level of real expenditure per capita at the very top-end of the distribution, with the 80th and 90th percentiles

consistently better-off compared to their peers in non-treated areas across the years. Furthermore, I find

this effect to be incremental with the more years of EPZ operation. The impact on overall within

municipality inequality is difficult to determine, which partly results from diverging dynamics between both

ends of the distribution. I find a sustained increase in top-end inequality across the period, but also a

smaller reduction in bottom-end disparities that fluctuates across the years. The slower diffusion of

treatment effect at the 10th and mid-range percentiles, only significant after 8 to 11 years of EPZ

36 See Appendix 3.

38

establishment, seems to be behind this trend. Interestingly, the heterogeneous time dynamic across the

different segments of the expenditure distribution suggests that the initial inequality-enhancing effect of

the policy fades with the more years of operation, eventually reaching the middle and poorer deciles.

Spillover effects to neighboring areas appear to be small in the timeframe studied and in no way

compromise previous findings. I find no evidence of larger displacements taking place as a result of the

policy.

The findings in this study are important from a policy perspective. They contribute to elucidate the

complex inequality dynamics of a popular policy tool. In a sense, they are supportive of the theoretical

predicaments of the more advanced HO models and polarization models. They contradict popular belief

that EPZs are directly beneficial for the poorer income segments in the municipalities where they locate.

The time lag registered for the policy to have a more equal effect across the distribution leaves a door open

for policy intervention. There is a need for further research in the area in order to identify with more

precision the different mechanisms at play.

References

Abadie, Alberto, Joshua Angrist, and Guido W. Imbens. “Instrumental variables estimates of the effect of subsidized training on the quantiles of trainee earnings.” Econometrica 70 (1) (2002): 91–117.

Aggarwal, Aradhna. “Special Economic Zones: Policy Debate.” Economic and Political Weekly 41 (43) (2006): 4533–36.

Anderson, Edward. "Openness and inequality in developing countries: A review of theory and recent

evidence." World Development 33, no. 7 (2005): 1045-1063. Angrist, J. D. and Jörn-Steffen Pischke. Mostly harmless econometrics: An empiricist’s companion. Princeton

University Press, 2008.

Athey, Susan, and Guido W Imbens. “Identification and Inference in Nonlinear Difference in Differences Models.” Econometrica 74 (2) (2006): 431–97.

Atkin, David. Endogenous skill acquisition and export manufacturing in Mexico. No. w18266. National Bureau of Economic Research, 2012.

Banerjee, Abhijit V., and Esther Duflo. "The economic lives of the poor." The journal of economic perspectives: a journal of the American Economic Association 21, no. 1 (2007): 141.

Banerjee, Abhijit, and Andrew Newman. 2004. "Inequality, Growth and Trade Policy" Unpublished.

Banerjee, Abhijit, Angus Deaton, and Esther Duflo. "Health, health care, and economic development: Wealth, health, and health services in rural Rajasthan."The American economic review 94, no. 2 (2004): 326.

Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. How much should we trust differences-in-differences estimates?. No. w8841. National Bureau of Economic Research, 2002.

Besley, Timothy, and Robin Burgess. "Can labor regulation hinder economic performance? Evidence from India." The Quarterly Journal of Economics(2004): 91-134.

David, H., David Dorn, and Gordon H. Hanson. The China syndrome: Local labor market effects of import competition in the United States. No. w18054. National Bureau of Economic Research, 2012.

39

Deaton, Angus. The analysis of household surveys: a microeconometric approach to development policy. World Bank Publications, 1997.

Engman, Michael, Osamu Onodera, and Enrico Pinali. Export processing zones: Past and future role in trade and development. No. 53. OECD Publishing, 2007.

Farole, Thomas, and Gokhan Akinci. eds. Special Economic Zones: Progress, Emerging Challenges and Future Directions. World Bank Plublications, 2011.

Farole, Thomas. Special Economic Zones in Africa: Comparing Performance and Learning from Global Experiences. World Bank Publications, 2011

Feenstra, Robert C., and Gordon H. Hanson. "Foreign direct investment and relative wages: Evidence from Mexico's maquiladoras." Journal of international economics 42, no. 3 (1997): 371-393.

FIAS. “SEZ: Performance, Lessons Learned and Implications for Zone Development.”The World Bank Group, 2008.

Firpo, Sergio. "Efficient semiparametric estimation of quantile treatment effects." Econometrica 75.1 (2007): 259-276.

Firpo, Sergio, Nicole M. Fortin, and Thomas Lemieux. "Unconditional quantile regressions." Econometrica 77, no. 3 (2009): 953-973.

Frölich, Markus, and Blaise Melly. "Estimation of quantile treatment effects with Stata." Stata Journal 10, no. 3 (2010): 423.

Glick, P., and F. Roubaud. 2006. “Export Processing Zone Expansion in Madagascar: What Are the Labor Market and Gender Impacts?” Journal of African Economies 15 (4): 722–56.

Goldberg, Pinelopi Koujianou, and Nina Pavcnik. Distributional effects of globalization in developing countries. No. w12885. National bureau of economic research, 2007

Hanson, Gordon H. "Increasing returns, trade and the regional structure of wages." The Economic Journal (1997): 113-133.

Harrison, Ann, and Gordon Hanson. "Who gains from trade reform? Some remaining puzzles." Journal of development Economics 59, no. 1 (1999): 125-154.

Havnes, Tarjei, and Magne Mogstad. "Is universal child care leveling the playing field?." Journal of Public Economics (2014).

Heckman, James J. and Carmen Pages. The cost of job security regulation: evidence from Latin American labor markets. No. w7773. National bureau of economic research, 2000.

Milberg, William, and Matthew Amengual. "Economic development and working conditions in export processing zones: A survey of trends." ILO, Geneva(2008).

Jenkins, Mauricio, Gerardo Esquivel, and Felipe Larraín B. “Export Processing Zones in Central America" Harvard Institute for International Development Working Paper 646 (1998)

Johansson, Helena, and Lars Nilsson. 1997. “Export Processing Zones as Catalysts.” World Development 25 (12): 2115–28.

40

Kusago, Takayoshi. “Export Processing Zones : A Review in Need of Update Human Development Network The World Bank Group, 1998.

Kovak, Brian K. 2011. “Local Labor Market Effects of Trade Policy: Evidence from Brazilian Liberalization.” Mimeo, Carnegie Mellon University

Lall, Somik Vinay, and Sanjoy Chakravorty. "Industrial location and spatial inequality: Theory and evidence from India." Review of Development Economics 9, no. 1 (2005): 47-68..

Melly, Blaise. Decomposition of differences in distribution using quantile regression. Labor Economics, 12(4) (2005), pp.577–590.

Meschi, Elena, and Marco Vivarelli. "Trade and income inequality in developing countries." World Development 37, no. 2 (2009): 287-302.

Milanovic, Branko. "Can we discern the effect of globalization on income distribution? Evidence from household budget surveys." Evidence from Household Budget Surveys (August 2002). World Bank Policy Research Working Paper 2876 (2002).

———. 2008. “Reform and Inequality during the Transition: An Analysis Using Panel Houshold Survey Data,

1990-2006.”, No. 2010, 62. Working paper/World Institute for Development Economics Research, 2010.

Neumark, D. & Kolko, J., Do enterprise zones create jobs? Evidence from California’s enterprise zone program. Journal of Urban Economics, 68(1) (2010), pp.1–19.

Rey, S. J., and M. V. Janikas. “Regional Convergence, Inequality, and Space.” Journal of Economic Geography 5 (2) (2005): 155–76.

Shorrocks, Anthony, and Guanghua Wan. "Spatial decomposition of inequality."Journal of Economic Geography 5, no. 1 (2005): 59-81.

Topalova, Petia. "Factor immobility and regional impacts of trade liberalization: Evidence on poverty from India." American Economic Journal: Applied Economics 2, no. 4 (2010): 1-41.

Verhoogen, Eric A. "Trade, Quality Upgrading, and Wage Inequality in the Mexican Manufacturing Sector." The Quarterly Journal of Economics 123, no. 2 (2008): 489-530.

Voitchovsky, Sarah. "Does the profile of income inequality matter for economic growth?." Journal of Economic Growth 10, no. 3 (2005): 273-296.

Wang, Jin. 2013. “The Economic Impact of Special Economic Zones: Evidence from Chinese Municipalities.” Journal of Development Economics 101 (2013). Elsevier B.V. 133–47.

World Bank (2009) Nicaragua Poverty Assessment, Poverty Reduction and Economic Management Sector, Latin America and the Caribbean Region Report No. 39736-NI, Washington DC.

World Bank (2011), Lopez H.J., Shankar R. (eds). “Getting the Most Out of Free Trade Agreements in Central America”, Directions in Development (Trade). The World Bank Group, Washington DC.

World Bank (2012) “Better Jobs in Nicaragua: the role of human capital”. Human Development Department, Latin America and the Caribbean Region Report No. 72923-NI, Washington DC

41

APPENDIX 1. DESCRIPTIVE STATISTICS FOR SELECTED CONTROL VARIABLES

TableA.1 Key Descriptive Statistics of Municipality Characteristics according to EPZ status

Percentages Mean SD Mean SD Mean SD Mean SD

Primary Education 29.15 3.6 24.28 7.33 28.7 2.6 29.02 5.06

Secondary Education 22.9 6.24 11.66 6.97 31.64 6.8 18.68 8.5

Terciary Education 3.7 2.3 0.82 0.9 8.26 4.00 2.72 2.52

Illiteracy rate 18.29 8.03 34.7 12.9 14.32 5.56 27.3 10.3

Unemployment 19.6 3.19 15.4 7.5 4.09 1.57 4.33 4.55

Urban Population 63.77 20.5 30.58 22.1 66.05 21.45 34.5 21.83

1990s 2000s

Municipalities

Ever EPZ

Municipalities

Ever EPZ

Municipalities

Never EPZ

Notes: The data is for 83 municipalities in 1993, 13 ever designated EPZ. Expenditure measures are expressed in constant

Cordobas of 1999 (Exchange rate US$1=C$11.8)

Municipalities

Never EPZ

42

APPENDIX 2. QDID ESTIMATES WITH ALTERNATIVE CONTROL GROUPS

Distance-Buffer RAAN/RAAS Exclusion

(1) (2)

188.8 273.8*

(1.307) (1.875)

82.34 155.5

(0.493) (0.941)

105.8 179.6

(0.481) (0.833)

224.9 311.9

(1.015) (1.41)

242.0 328.6

(0.99) (1.336')

205.0 375.8

(0.693) (1.181)

370.5 530.6

(1.102) (1.542)

863.8* 1,071**

(1.887) (2.284)

2,185** 2,648**

(1.994) (2.298)

Yes Yes

Yes Yes

Yes Yes

432 500

107 126

Real expenditure per

capita

10th Percentile

20th Percentile

30th Percentile

40th Percentile

Table A.2 QDID Estimates of the Effect of EPZs across the Expenditure

Distribution in Treated Municipalities - Additional Control Groups

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities). Covariates

include illiteracy rate, primary and tertiary education, as well as share of urban population

and migrants measured at municpality level. Rsq. for both columns ranged between 0.52-

0.74. Regression in column (1) limits observations with respect to the distance-buffer

and in column (2) exclude municipalities in RAAN/RAAS.

*** p<0.01, ** p<0.05, * p<0.1

Municipality FE

Covariates

Observations

Number of Muni.

50th Percentile

60th Percentile

70th Percentile

80th Percentile

90th Percentile

Year FE

43

APPENDIX 3.A EFFECT OF EPZS ON BUSINESS CREATION

APPENDIX 3.B EFFECT OF EPZS ON INTERNAL MIGRATION

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

0.010 0.008 0.0418 0.035

(1.12) (0.85) (0.66) (0.47)

Year FE Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes

Covariates No Yes No Yes

Observations 78,166 76,668 78,168 76,668

R sq. 0.03 0.09 0.03 0.11

Number of Muni. 145 145 145 145

EPZ

Notes: Robust t and z-statistics in parentheses, Clustered SE ( municipalities)

Covariates includes dummy variables for level of educational achievement, illiteracy,

gender, urban residence, age square, number of household members, relationship

to household head. Observations are limited to working age popluation (15-65

years).Logit outputs are log-odds coefficients.

*** p<0.01, ** p<0.05, * p<0.1

Table A.3 (a) Effect of EPZs on Business Start-ups

(Self-employed) 1993-2009

OLS LOGIT

OLS LOGIT OLS LOGIT

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

-0.016 -0.192 -0.041 -0.197

(-1.01) (-0.70) (-0.50) (-0.67)

Year FE Yes Yes Yes Yes

Municipality FEYes Yes Yes Yes

Covariates Yes Yes Yes Yes

Observations 45,227 45,227 30,349 30,349

R sq./Pseudo Rsq. 0.17 0.28 0.19 0.19

Number of Muni. 145 145 145 145

Table A.3(b) The Effect of EPZs on Migration

1993-2009 Excl. 1998, 2009

Notes: Robust z and t-statistics in parentheses, Clustered SE ( municipalities).

Covariates includes dummy variables for level of educational achievement,

illiteracy, gender, urban residence, age square, number of household members,

relationship to household head. Observations are limited to working age

popluation (15-65 years).The left hand side measure for migration is a dummy

variable indicating if the person migrated within the past 5 years. Logit outputs are

log-odds coefficients.

*** p<0.01, ** p<0.05, * p<0.1

EPZ

44

APPENDIX 4. SPILLOVER DYNAMICS – ALTERNATIVE CONTROL GROUP

10th

percentile

20th

percentile

30th

percentile

40th

percentile

50th

percentile

60th

percentile

70th

percentile

80th

percentile

90th

percentile

(1) (2) (3) (4) (5) (6) (7) (8) (9)

EPZ 258.5* 147.3 157.8 320.3 327.6 386.9 551.2 1,165** 2,707**

(1.757) (0.897) (0.74) (1.438) (1.313) (1.191) (1.568) (2.488) (2.38)

Neighboring Municipalities1 -117.8 -275 -237.9 204.5 122.1 168 403.5 845.7 778.6

(5 to 15km) (-0.422) (-1.039) (-0.707) (0.587) (0.387) (0.531) (0.973) (1.653) (1.625)

Neighboring Municipalities2 24.18 46.53 58.67 142.8 126.4 202.5 208.3 453.8 545.2

(15 to 35km) (0.182) (0.276) (0.234) (0.59) (0.483) (0.707) (0.698) (1.161) (0.899)

Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 575 575 575 575 575 575 575 575 575

R-sq. 0.625 0.673 0.672 0.704 0.701 0.619 0.619 0.594 0.356

Number of Muni.145 145 145 145 145 145 145 145 145

Table A.4 The Effect of EPZs Establishment & Spillover Dynamics

All Municipalities

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities) Covariates include dummies for illiteracy rate, primary and tertiary education, as well as share of urban

population, at municipality level. The variables Neighbour Municipalities are dummy variables that equal one if a municipality's centroid is situated at 5 to 15km, and 15km to 35km

distance from a treated municipality, and zero otherwise. I use the centroid's distance to avoid noise from larger municipalities.

*** p<0.01, ** p<0.05, * p<0.1

45

APPENDIX 5. HETEROGENEOUS TIME EFFECTS – ALTERNATIVE CONTROL GROUP

10th

percentile

20th

percentile

30th

percentile

40th

percentile

50th

percentile

60th

percentile

70th

percentile

80th

percentile

90th

percentile

(1) (2) (3) (4) (5) (6) (7) (8) (9)

EPZ establishment, t0 124.2 -10.12 72.46 145.7 86.34 214.5 451.5 694.5 2,214*

(0.868) (-0.069) (0.486) (0.759) (0.369) (0.740) (1.063) (1.464) (1.665)

EPZ establishment, t+4 412.8 300.6 206.0 386.5 377.6 236.4 247.4 1,070* 2,484*

(1.580) (1.033) (0.554) (1.126) (1.108) (0.567) (0.561) (1.947) (1.968)

EPZ establishment, t+8 393.9** 508.9* 474.4* 718.1*** 1,108*** 1,250*** 1,267*** 2,447*** 4,128***

(2.407) (1.949) (1.696) (3.006) (2.740) (2.892) (3.253) (3.069) (4.209)

EPZ establishment, t+11 905.2*** 562.4 787.0 936.0* 1,246*** 1,170*** 1,371*** 2,051*** 5,381***

(3.182) (1.443) (1.462) (1.900) (3.247) (3.035) (2.617) (3.548) (5.099)

Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 575 575 575 575 575 575 575 575 575

R-sq. 0.629 0.676 0.677 0.708 0.706 0.623 0.622 0.598 0.364

Number of Muni. 145 145 145 145 145 145 145 145 145

Table A.5 The Heterogeneous Time-Effect of EPZ Establishment

All Municipalities

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities) Covariates include illiteracy rate, primary and tertiary education, as well as the share of urban population and

migrants at the municipality level.. All leads are equal to one in only one year each per adopting state. EPZ dummy (t0) equals one only for year of establishment. EPZ t+4 includes

t+3, and EPZ t+8 includes t+7, to have balanced dummies.

*** p<0.01, ** p<0.05, * p<0.1

46

APPENDIX 6. HETEROGENEOUS TIME EFFECTS – ALTERNATIVE CONTROL GROUP, EXCL. MANAGUA

10th

percentile

20th

percentile

30th

percentile

40th

percentile

50th

percentile

60th

percentile

70th

percentile

80th

percentile

90th

percentile

(1) (2) (3) (4) (5) (6) (7) (8) (9)

EPZ establishment, t0 162.3 16.81 107.2 186.0 142.8 288.9 555.9 804.7* 2,437*

(1.102) (0.109) (0.708) (0.943) (0.596) (0.960) (1.285) (1.680) (1.793)

EPZ establishment, t+4 438.2 335.5 248.6 438.5 458.8 320.1 325.0 1,207** 2,723**

(1.574) (1.081) (0.624) (1.199) (1.283) (0.730) (0.695) (2.102) (2.063)

EPZ establishment, t+8 325.8* 422.4 326.0 525.6** 985.8** 1,144** 1,177** 2,454*** 3,804***

(1.829) (1.470) (1.110) (2.404) (2.169) (2.312) (2.570) (2.680) (3.537)

EPZ establishment, t+11 851.3** 400.2 603.0 640.5 1,060** 937.7* 1,036 1,774** 5,485***

(2.237) (0.801) (0.826) (0.995) (2.018) (1.864) (1.565) (2.431) (4.016)

Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Municipality FE Yes Yes Yes Yes Yes Yes Yes Yes Yes

Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 570 570 570 570 570 570 570 570 570

R-sq. 0.629 0.676 0.677 0.708 0.706 0.623 0.622 0.598 0.364

Number of Muni. 144 144 144 144 144 144 144 144 144

Table A.6 The Heterogeneous Time-Effect of EPZ Establishment - Excl. Managua

All Municipalities

Notes: Robust t-statistics in parentheses, Clustered SE ( municipalities) Covariates include illiteracy rate, primary and tertiary education, as well as the share of urban population and

migrants at the municipality level.. All leads are equal to one in only one year each per adopting state. EPZ dummy (t0) equals one only for year of establishment. EPZ t+4 includes

t+3, and EPZ t+8 includes t+7, to have balanced dummies.

*** p<0.01, ** p<0.05, * p<0.1