Partial Identification and Bound Estimation of the Average Treatment Effect of Education on Earnings for South Africa
| Date | 01 April 2015 |
| Published date | 01 April 2015 |
| Author | Juergen Meinecke,Martine Mariotti |
| DOI | http://doi.org/10.1111/obes.12059 |
210
©2014 The Department of Economics, University of Oxford and JohnWiley & Sons Ltd.
OXFORD BULLETIN OF ECONOMICSAND STATISTICS, 77, 2 (2015) 0305–9049
doi: 10.1111/obes.12059
Partial Identification and Bound Estimation of the
AverageTreatment Effect of Education on Earnings
for South Africa*
Martine Mariotti† and Juergen Meinecke‡
†Australian National University, Research School of Economics, HW Arndt Building 25a
Canberra, ACT 0200, Australia (e-mail:martine.mariotti@anu.edu.au)
‡Australian National University, Research School of Economics, HW Arndt Building 25a
Canberra, ACT 0200, Australia (e-mail:juergen.meinecke@anu.edu.au)
Abstract
We partially identify the average treatment effect of education on earnings and non-
parametrically estimate its upper bound for African, Coloured (mixed race) and White
males in South Africa. We address endogenous selection into education, cohort effects
and endogenous selection into work. Using the September 2007 South African Labour
Force Survey, the upper bound estimates are considerably lower than existing parametric
point estimates of the return to education. As a lesson, policy makers should focus less
on increasing the amount of education obtained by Africans and Coloureds, but rather on
measures that can grow the return to existing levels of education.
I. Introduction
Our objective is to partially identify and non-parametricallyestimate an upper bound on the
average treatment effect(ATE) of education on earnings forAfrican, Coloured (mixed race)
and White males in South Africa.While partial identification with its transparent and weak
non-parametric assumptions offers an attractive alternative to parametric estimation, it has
not been applied in this context before.The partial identification strategy follows three steps:
define a policy-relevantATE, provide a weakset of assumptions that leads to identification
of an upper bound on the ATE and propose a consistent upper bound estimator. Our upper
bound estimates fall well below existingparametric retur n to education estimates for South
Africa. An important policy implication of these results is that decision makers ought to
focus less on the amount of education obtained byhistorically disadvantaged race groups and
more on measures that can increase the marginal return of the current levels of education.
*We thank seminar participants at ESAM 2009, the ANU School of Economics Seminar, the RSSS brown bag,
seminar, UNSW, La Trobe University, University of Otago, Monash University, Labour Econometrics Workshop
2010 and the Econometrics Society WorldCongress 2010 for valuable comments. We are grateful to Gaurab Aryal,
Robert Breunig, Robert Gregory, TimothyHatton, Brian McCaig and Farshid Vahidfor suggestions. Guochang Zhao
provided excellent research assistance.All er rors are those of the authors.
JEL Classification numbers: C14, J31, O15.
ATEof education on earnings 211
We define theATE explicitly as a function of a person’s actual choice of education. For
example, we ask by how much expected earnings of a person with seven years of education
(end of primary school) could increase, at most, when treated with a completed high-school
degree (12 years of education).The answer to this question is of immediate policy relevance.
The educational attainment ofAfricans and Coloureds in South Africa substantially lags that
of Whites. One explanation is that some Africans and Coloureds are not able to choose an
optimal level of education – their observed educational outcomes are constrained optima –
because of rigidities such as credit constraints, institutional constraints, discrimination or
household decision making/time preferences. A policy that aims to reduce or remove these
rigidities can use theATE to learn about the maximum benefit for those people for which the
constraints/rigidities are binding.Another explanation is that the perceived opportunity cost
of dropping out of school is low (see Lam et al., 2011) and knowing theATE could inform
individuals about the benefits of additional education.This in turn could affect intertemporal
choices as well as time preferences and as a consequence lead to changes in human capital
investment.
The parametric literature for SouthAfrica generally estimates a high return to education,
especially for Africans and Coloureds. Leibbrandt et al. (2005) find an annual return to ed-
ucation for men (across race groups) of 11.8% in 1995 and 11.2% in 2000. Anderson et al.
(2001) argue that the return to education forAfricans lies between 20%–25% for each year of
secondary school. Mwabu and Schultz (2000), allowingfor nonlinearities in the conditional
expectation function, find a return to education that is particularly high for Africans and
Coloureds at the secondary and tertiary education levels. Theyestimate a retur n forAfricans
of 8.4% for each year of primary school, 15.8% for each year of secondary school and 29.4%
for each year of tertiary education. For Coloureds, the return equals 1.4% (primary), 18.7%
(secondary) and 18.6% (tertiary). For Whites, the return is lower at 8.4% (secondary) and
15.1% (tertiary). Only two articles attempt to correct for ability bias. Hertz (2003) finds an
annual return to education for Africans (men and women) of 11.4%.Applying a restrictive
parametric model of within-household differences to address ability bias (and measurement
error), the return to education estimates go down to 5.2% per year of education. Moll (1998),
using measures of cognitive ability (reducing his sample size to only 133 observations),
estimates a return of 15% for each year of primary and also secondary school in the absence
of measures of cognitive ability and 2.9% (primary) and 9.7% (secondary) when including
these measures.1
Three lessons emerge from the parametric literature. First, estimates of the effect of
education on earnings are sensitive to the parametric specification. Different specifications
yield vastly different estimates of the return to education. Second, the relationship between
education and earnings is likely nonlinear (with the return varying considerably by edu-
cation level). A simple linear approximation of the human capital earnings function does
not adequately reflect the substantial heterogeneity in the return to education. Third, unob-
served ability creates an upwardbias in estimating the causal effect of education on earnings.
The non-parametric partial identification frameworkproposed by Manski and Pepper (2000,
2009) lends itself well to this empirical setting. Following this approach, we derive an upper
1Other articles that estimate the return to education include Maitra and Vahid (2006) and Mwabu and Schultz
(1996). In a related context, Yamauchi (2008) highlights the importance of early childhood health capital on human
capital development and ultimatelythe return to education.
©2014The Depar tment of Economics, Universityof Oxford and John Wiley & Sons Ltd
To continue reading
Request your trial