of 11 /11
1/11 Returns to Schooling in Egypt Meltem Dayıoğlu, Murat G. Kırdar, Abdurrahman Aydemir, Ragui Assaad Key contribution: Break endogeneity of schooling by using the 1988 reduction in primary schooling as an IV. Use two-stage least squares estimation to find the causal link between schooling and wages: i i i u X D s + + + = θ α α ' 1 0 (1) i i i v X s w + + + = δ β β ' 1 0 log (2) Data: 2012 round of the Egypt Labor Market Panel Study (ELMPS)

Embed Size (px)

### Transcript of Returns to Schooling in Egypt

1/11

Returns to Schooling in Egypt

Meltem Dayıoğlu, Murat G. Kırdar, Abdurrahman Aydemir, Ragui Assaad Key contribution: Break endogeneity of schooling by using the 1988 reduction in primary schooling as an IV. Use two-stage least squares estimation to find the causal link between schooling and wages:

iii uXDs +++= θαα '10 (1)

iii vXsw +++= δββ '10log (2)

Data: 2012 round of the Egypt Labor Market Panel Study (ELMPS)

2/11

When a dummy variable is used as the instrument, the IV estimator in the simple regression model is the Wald estimator, which can be calculated from knowledge of subsample means:

𝛽𝐼𝑉 =𝑊�1 −𝑊�0𝑆1̅ − 𝑆0̅

Here W denotes (log)wages, S years of schooling, subscripts refer to control (0) and treatment groups, bars to sample means.

Evidently average schooling went down (𝑆1̅ − 𝑆0̅ < 0) and average wages also went down (𝑊�1 −𝑊�0 < 0) so 𝛽𝐼𝑉 > 0.

In the case of females: 𝑆1̅ − 𝑆0̅ ≅ 0, so Wald estimator is not defined.

Weak instrument problem.

3/11

• Treatment and control groups differ in an important way: treated group is younger than the controls. Thus age and/or cohort effects have to be controlled for. Thus use of the Wald estimator is not appropriate. But it is still a useful benchmark worth reporting.

Paper: “The fact that birth-cohorts might have direct effects on wages means that the independence assumption of our instrument is more likely to hold conditional on birth-cohort groups. This issue is ignored in previous studies that use the variation in exposure to the policy across birth-cohorts generated by compulsory schooling laws.” References?

Do the previous papers use multiple or single cross-section?

Cubic in age vs. coarse cohort dummies: How results change

4/11

Assignment of treatment status:

Policy change came in 1988 and affected those born in 1978 and later – as long as the individual started school at age 6 and did not repeat grades. Since late starters and grade repeaters are present, some members of early cohorts may have been affected by the policy as well. So two cut-offs are used, 1978 and 1977.

Results favor use of 1977 (drop in schooling detected).

{1 = 1977 or later, 0 = 1976 or earlier},

it might be better to use

{1 = 1978 or later, 0 = 1976 or earlier}.

5/11

How about propensity score weighting? For a range of birth cohorts, membership probability in treatment group increases with year of birth, and may also depend on HH characteristics.

Can this function be estimated using other data sources that have information on school starting age and grade repetition?

Operational sample sizes:

All = 3300-7500, M = 2600-6300, F = 500-1200.

May not be possible to do much with the female sample.

6/11

Functional form: “There are a number of assumptions that the Mincer earnings equation makes. First, it postulates a linear relationship between schooling and the logarithm of wages. Card (1999) claims that this assumption is supported by empirical results for the US; however, Heckman et al. (2007) reports that several of the implications of the Mincer earnings equation are not supported by US earnings data. Second, the Mincer earnings framework imposes separability between the schooling and experience variables. In other words, it rules out that returns to schooling change with the experience level. Similar to other reduced-form studies, we also assume that these two assumptions hold.”

Why not say: We have one instrument, so we have to make some strong assumptions about functional form.

7/11

What does LATE capture?

“Imbens and Angrist (1994) illustrate that the 2SLS estimates give the effect for the set of compliers—those who are induced to change their schooling behavior as a result of the exogenous source of variation in schooling. The set of compliers may not be representative of the whole population and the estimates derived provide local treatment effects (LATE) for compliers. Oreopoulos (2006) notes that as the fraction of the population affected by the policy change increases, the LATE estimate approaches the average treatment effect (ATE): the marginal individual affected by the policy change becomes more similar to the average individual in the population.” (p.7)

8/11

Earlier: “Oreopoulos (2006) notes that as the fraction of the population affected by the policy change increases, the LATE estimate approaches the average treatment effect (ATE): the marginal individual affected by the policy change becomes more similar to the average individual in the population. Using a much larger set of compliers, Oreopoulos (2006) finds that the 2SLS estimates in the UK are not much different than that found by Angrist and Krueger. Devereux and Hart (2010), however, find much lower returns in the same context. Pischke and von Wachter (2008) similarly report much lower IV estimates (zero returns) than the OLS estimates in Germany where a large fraction of the population were affected by the compulsory school law change.” (p.4)

9/11

“Imbens and Angrist (1994) also note that if there are nonlinearities in the return to schooling, estimating the return via a policy-induced increase at a given schooling level may lead to differences between the estimated LATE and the ATE.”

In this case a weighted average of the different returns is estimated:

10/11

This paper: returns to schooling in Egypt for men ages 20-45 is between 2 to 5.7 percent.

Aydemir and Kırdar (2014) – similar exercise for Turkey. They find a wage return that is almost zero for men and 3.8 percent for women.

Should we give up an improvement of education as a policy goal?

Not yet.

LATE vs. ATE

Remember: Education is the “mother of all variables” (MM) – associated with good outcomes across the board.

11/11

Lingering questions:

Falsification exercise mentioned in section 6.2:

Did you do it?

Section 6.4 – Effect of schooling on employment type

Do some of the results pose challenges for the conditional independence assumption?