43
1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

Embed Size (px)

DESCRIPTION

3 Coalmining and Respiratory Symptom Each physiological system has a certain tolerance: tolerance vector Ashford and Sowden: ignoring important information if you estimate equations separately  Model in spirit of Seemingly Unrelated Regression (SUR)

Citation preview

Page 1: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

1

Multivariate Probit

An Analysis of Access to Amenities in Delhi’s Slums

Page 2: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

2

Example: Coalmining and Respiratory Symptoms Ashford and Sowden, Biometrics, 1970 Is there a relationship? Best model? Standard approach: two probit equations

1. Wheezing and years in mine (age)2. Breathlessness and years in mine (age)3. Does this approach overlook anything?

Page 3: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

3

Coalmining and Respiratory Symptom Each physiological system has a certain

tolerance: tolerance vector Ashford and Sowden: ignoring important

information if you estimate equations separately Model in spirit of Seemingly Unrelated Regression

(SUR)

Page 4: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

4

Multivariate Probit (MVP)

Extension of univariate probit (UVP) model Allows for:

1. Simultaneous estimation of multiple probit equations

2. Correlated disturbances across equations

Page 5: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

5

MVP vs. UVP

How is it better? Does not ignore information across equations Better prediction of conditional and joint

probabilities More consistent estimation More efficient estimation

Page 6: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

6

Why MVP?

Access to amenities in Delhi slums Correlation between access to sanitation

services and access to drainage for a given household

Separate UVP estimation would ignore this

Page 7: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

7

Where we’re going…

I. Univariate ProbitII. Bivariate Probit (i = 2)III. Multivariate Probit (i = T)IV. Delhi Slum Dwellers’ Access to AmenitiesV. Published Applications and Extensions

Page 8: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

8

Univariate Probit

Dichotomous dependent variable Takes on a value of either 0 or 1

Estimate with OLS?

Page 9: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

9

Linear Probability Model

),()|1Pr( xx FY

),(1)|0Pr( xx FY

'),( xx F 'xY

Page 10: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

10

Linear Probability Model

Shortcomings Cannot constrain probabilities to the 0-1

interval Negative variances Heteroscedasticity of ε that depends on X Logically not attractive Solutions?

Page 11: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

11

Normal Cumulative Distribution FunctionProperties Bounded by 0 and 1 Nonlinear relationship between P and X

Page 12: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

12

Univariate Probit Model

)'()()|1Pr()|0*Pr('

xxxx

dttYy

y*: latent dependent variable

Y: binary dependent variable

x: vector of explanatory variables

t: standardized normal variable

φ: normal pdf

β: measures impact of changes in x

Φ: normal cdf

Page 13: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

13

UVP Example

Y = 1: individual purchased refrigerator in last yearY = 0: individual has not purchased refrigerator in last

yearX = individual’s income per annum

)()|1Pr( XXY

Page 14: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

14

UVP: Mechanics

1. Group the data by RHS variable (income)2. Calculate Phati for each i grouping of income3. For each Phati, use the standard normal cdf to find Ii

4. Add 5 to each Ii

5. Use OLS to estimate β1, β2 in:

iii uXI 21

Page 15: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

15

Two probit equations

Y1 = 1: individual purchased refrigerator in last year, 0 otherwise

Y2 = 1: individual purchased dishwasher in last year, 0 otherwise

X = individual’s income per annum

Page 16: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

16

Bivariate Probit

0*1 y

, y 1 = 1 if , 0 otherwise

22'2

*2 xy , y 2 = 1 if , 0 otherwise

0],|[],|[ 212211 xxxx EE

1],|[],|[ 212211 xxxx VarVar

0,],|,[ 2121 xxCov

0*1 y11

'1

*1 xy

Page 17: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

17

BVP: Estimation Maximum Likelihood Bivariate Normal cdf:

212122211 ,),,(),Pr( dzdzzzxXxX

where φ2 represents the bivariate normal pdf and:

2/12

)1/()2)(2/1(

212 )1(2),,(

22

22

21

xxxxexx

Page 18: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

18

BVP: Estimation

Probabilities that enter the likelihood function:

21* iiiqq

),,(),|,Pr( *212212211 iiiii wwyYyY xx

Φ2: bivariate normal cdf

2,1,' jqw jijijij x12,12 2211 iiii yqyq

Page 19: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

19

BVP: Estimation

Function to be maximized:

n

iiii wwL

1212 ),,(lnlog *

Page 20: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

20

BVP: Estimation in Practice

Simulated Maximum Likelihood1. Markov chain Monte Carlo2. GHK simulator

Geweke-Hajivassiliou-Keane smooth recursive conditioning simulator

Greene (2003) discusses this in Appendix E Cappellari and Jenkins (2003)

Page 21: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

21

BVP: Is it Necessary? H0: ρ = 0 (estimate independent probit

equations separately) Test statistic (Kiefer 1982):

n

iiiii

ii

n

iii

iiii

wwwwww

wwww

qqLM

12211

221

2

121

2121

)()()()()()(

)()()()(

LM ~ χ2 with d.f. = (T)(T-1)/2 where T = # of equations

Page 22: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

22

BVP: More Test Statistics

z-statistic:

Likelihood ratio:

j = number of restrictions

)(~ˆ 2

ˆ

js

z

)(~)ˆlnˆ(ln2 2 jLLLR UR

Page 23: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

23

BVP: Properties of the Estimator Considers unobservable heterogeneity Random components of one equation are allowed to

be freely correlated with the random components of the other

Takes into account unobservable characteristics that might affect both dependent variables

More efficient and consistent than separate ML estimation of UVP models UVP does not account for the correlation between error

terms: assumes exogeneity of dep var covariates, so does not give consistent estimates of parameters (Maddala 1983)

Page 24: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

24

BVP: Measure of Goodness of FitMcFadden’s likelihood ratio index (LRI):

lnL: maximized value of log-likelihood function for specification at hand

lnL0: maximized value of log-likelihood function calculated with only a constant term

Bounded by 0, 1, increases as fit improves

0ln/ln1 LLLRI

Page 25: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

25

Multivariate Probit

iiiy '* xyi = 1 if yi* > 0, 0 otherwise, i = 1,…,T

,0~',,1 MVNT

where

TTT

T

1

111

and ρii = 1, ρij = ρji

for i, j = 1, …, T

Page 26: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

26

MVP Application: Delhi SlumsDelhi slum dwellers’ access to amenities

Page 27: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

27

Delhi Slums

Tiebout sorting (Charles Tiebout 1956) Individuals sort themselves into

communities based on preferences of provisions of public goods

Assumptions1. Unlimited mobility2. Unrestricted number of communities

Implication: Heterogeneous preferences

Page 28: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

28

Delhi Slums

Heterogeneity in community composition Impact on economic outcomes

1. Reduced participation to secure community grants in US (Vigdor 2004)

2. Decreased maintenance of infrastructure projects in Pakistan (Khwaja 2001)

3. Less spending on education, sewers, roads in US (Alesina et al 1999)

4. Slower growth in Sub-Saharan Africa (Easterly and Levine 2003)

Channels?

Page 29: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

29

Delhi Slums: Model (Alesina et al 1999)Model:

g*: amount of public good provided in equilibrium : median distance from the type of public good most

preferred by the median voter α: parameter from individual’s utility function (0<α<1) Punchline: g* and are inversely related:

)1/(1* ˆ1

milg

mil̂

mil̂

1/*

ˆ11ˆ

mim

i

llg

Page 30: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

30

Delhi Slums: H0 and H1

H0: Public goods provision is not affected by the degree to which preferences are polarized

H1:Public goods provision is negatively affected by polarization of preferences

Page 31: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

31

Delhi Slums: Data

Variable name Mean Median Min. Max. Std. dev.

Religious fractionalization 0.22 0.24 0 0.43 0.11Caste fractionalization 0.68 0.69 0.44 0.73 0.04Access to medical facility (DV) 0.55 1 0 1 0.50Access to sanitation services (DV) 0.70 1 0 1 0.46Access to drainage system (DV) 0.79 1 0 1 0.41Household PC income (monthly in Rupees) 841 686 186 5947 616Participation in political process (DV) 0.78 1 0 1 0.42Years in community of head of HH 16 16 0 84 0.42Mean to median income ratio 1.14 1.14 0.98 1.24 0.08

TABLE ISummary Statistics of Slum Data

Page 32: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

32

Delhi Slums: Public Goods

Provision of public goods is a latent variable Proxy with access to public goods

1. Medical facilities (MED)2. Sanitation services (SAN)3. Drainage (DRA)

Page 33: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

33

Delhi Slums: FractionalizationProxies for Fractionalization Religion

1. Hindu2. Muslim3. Sikh

Caste1. Backward castes and tribes2. Scheduled castes and tribes3. General Hindu4. Muslim, Sikh, other

Page 34: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

34

Delhi Slums: Econometric Model

: amount of public good a (latent) accessible by slum dweller i

Map to the observed realizations: 1 represents access, 0 otherwise

if , 0 otherwise Lose information

aaay '* X DSMa ,,

1aiy 0* aiy

*aiy

Page 35: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

35

Delhi Slums: Econometric ModelAssumption:

where

UVP:

,0~',, MVNDSM

11

1

SDMD

SDMS

MDMS

0 SDMDMS

Page 36: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

36

Delhi Slums: Econometric ModelX vector: Religious fractionalization (frd) Caste fractionalization (fcd) Per capita household income (pcinc) Education dummies (edu1, edu2, edu3) Mean-to-median income ratio (mminrat) Poverty dummy (poor) Political participation dummy (political) Years in community (yrincomm)

Proxies for lhatm

Page 37: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

37

RHS variables MVP UVP MVP UVP MVP UVP

REL -1.795 -1.712 -5.637 -5.03 -1.596 -1.464(1.38) (1.28) (3.45)** (3.11)** (1.15) (0.96)

CASTE -5.615 -5.546 16.345 15.702 9.837 9.104(1.87) (1.80) (3.49)** (3.36)** (3.36)** (2.81)**

PCINC 0 0 0 0 0 0(0.34) (0.67) (0.11) (0.16) (0.33) (0.20)

EDU1 0.206 0.238 0.225 0.278 0.186 0.158(0.89) (1.02) (0.97) (1.19) (0.81) (0.65)

EDU2 -0.044 -0.034 0.835 0.891 0.818 0.707(0.15) (0.12) (2.38)* (2.50)* (2.19)* (1.86)

EDU3 0.221 0.256 0.273 0.2 3.66(0.43) (0.49) (0.51) (0.37) -0.03

MMINRAT 4.487 4.545 2.558 2.456 0.297 0.487(3.47)** (3.47)** (1.65) (1.57) (0.20) (0.31)

POOR -0.246 -0.241 0.322 0.336 0.563 0.453(1.13) (1.09) (1.34) (1.36) (2.10)* (1.65)

POLITICAL 0.537 0.532 0.258 0.219 0.186 0.317(2.30)* (2.26)* (1.10) (0.93) (0.77) (1.27)

YRINCOMM 0.011 0.016 0.01 0.011 0.004 -0.001(1.04) (1.41) (1.01) (1.07) (0.35) (0.07)

Constant -1.391 -1.69 -13.068 -12.607 -6.416 -6.167(0.55) (0.64) (3.19)** (3.09)** (2.37)* (2.09)*

Coeff. z-stat.ρ MS 0.42 (3.81)**ρ MD -0.25 (1.94)ρ SD 0.52 (4.92)**

Absolute value of z-statistics in parentheses* significant at 5%; ** significant at 1%tedu3 dropped b/c it predicted success perfectly

Correlation Coefficients

TABLE II

n/at

Medical Facilities Sanitation Services Drainage

Dependent Variables are Access to the Specified Amenity

Page 38: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

38

Delhi Slums: MVP vs. UVP

Is MVP necessary? H0: ρMS = ρMD = ρSD = 0?

Stata reports the LR test statistic = 37.78 ~ χ2(3), so reject H0

Yes, MVP is an improvement on UVP Not ignore information contained in covariance matrix

goodness of fit: 137.033.35032.3021

LRI

Page 39: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

39

Future Research

Dependent variable More direct measure for spendingPanel data Changes in income v changes in

fractionalizationSemiparametric and nonparametric techniques Horowitz and Savin (2001)

1. Single-index modeling2. Median regression approach

Page 40: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

40

MVP in Practice

Ashford and Sowden (1970)

Zhao, X. and M. Harris (2004) “Demand for Marijuana, Alcohol and Tobacco: Participation, Levels of Consumption and Cross-equation Correlations,” The Economic Record, 80(251): 394-410.

Greene, W. (1998) “Gender Economics Courses in Liberal Arts Colleges: Further Results,” The Journal of Economic Education, 29(4): 291-300.

Christofides, L., T. Stengos, and R. Swidinsky (1997) “Welfare Participation and Labour Market Behavior in Canada,” The Canadian Journal of Economics, 30(3): 595-621.

Page 41: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

41

ReferencesAlesina, A., R. Baqir, and W. Easterly (1999) “Public Goods and

Ethnic Divisions”, The Quarterly Journal of Economics, 114(4): 1243-1284.

Ashford, J.R. and R.R. Sowden (1970) “Multi-variate Probit Analysis,” Biometrics, September: 535-546.

Cappellari, L. and S.P. Jenkins (2003) “Multivariate probit regression using simulated maximum likelihood,” Stata Journal, 3(3): 221-235.

Easterly, W., and R. Levine (1997) “Africa’s Growth Tragedy: Policies and Ethnic Divisions,” Quarterly Journal of Economics, 112(4), 1203-1250.

Greene, W. (2003) Econometric Analysis (Fifth Edition), Delhi:

Pearson Education.

Page 42: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

42

References (cont.)Horowitz, J. and N.E. Savin (2001) “Binary Response Models:

Logits, Probits and Semiparametrics,” Journal of Economic Perspectives, 15(4): 43-56.

Kiefer, N. (1982) “Testing for Dependence in Multivariate Probit Models,” Biometrika, 69(1): 161-166

Khwaja, A.I. (2001) “Can Good Projects Succeed in Bad Communities? Collective Action in the Himalayas,” John F. Kennedy School of Government Faculty Research Working Paper Series RWP01-043. URL: http://ssrn.com/abstract=295571

Maddala, G.S. (1983) Limited Dependent Variables in Econometrics, Cambridge: Cambridge University Press.

Page 43: 1 Multivariate Probit An Analysis of Access to Amenities in Delhi’s Slums

43

References (cont.)

Tiebout, C. (1956) “A Pure Theory of Local Expenditures,” Journal of Political Economy, 64(5): 416-424.

Vigdor, J. (2004) “Community Composition and Collective Action: Analyzing Initial Mail Response to the 2000 Census,” The Review of Economics and Statistics. 86(1): 303-312.