Take the 2-minute tour ×
Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It's 100% free, no registration required.

Consider the following model:

  • $wage_i=\beta_0+\beta_1after_i+\beta_2female_i+\beta_3after_ifemale_i+u_i$

where

  • $after_i=1$ if date after gender-wage discrimination policy; $0$ if date before gender-wage discrimination policy.

  • $female_i=1$ if female; $0$ if male.

The 'treatment group' is females.

This model measures the effect of a gender-wage discrimination policy on the average wage of women relative to men.

The affect of the policy is captured by $\beta_3$ which can be estimated as follows:

  • $\hat{\beta_3}=(E[wage_i|after_i=female_i=1]-E[wage_i|after_i=1, female_i=0])-(E[wage_i|after_i=0, female_i=1]-E[wage_i|after_i=female_i=0])$

which is known 'difference-in-differences' estimator.

Questions:

  1. If $\hat{\beta_3}>0$, then does this means that the policy caused women's earning to increase relative to mens?
  2. For $\hat{\beta_3}$ to be interpreted as the causal effect of the policy on wages, does $E(u_i|female_i, after_i)=0$ have to be assumed?

Any guidance would be very much appreciated. Thank-you.

share|improve this question
    
How many time periods do you have? –  Andy May 17 at 16:12
    
A pool of two cross-sections, one from the year before the policy was implemented and one from the year after. –  ajohnrobertson May 17 at 16:23

1 Answer 1

up vote 4 down vote accepted

The way you have described the problem, females are not the control but the treatment group. Then a positive $\widehat{\beta}_3$ would imply that women's earnings have increased more relative to men's earnings in response to the discrimination policy. Equivalent your expression, this will then be captured by the coefficient: $$ \begin{align} \widehat{\beta}_3 &= (E[\text{wage}_i | \text{female}=1, \text{after}=1] - E[\text{wage}_i | \text{female}=1, \text{after}=0]) \newline &- (E[\text{wage}_i | \text{female}=0, \text{after}=1] - E[\text{wage}_i | \text{female}=0, \text{after}=0]) \end{align} $$

With respect to your second question: yes, you need this assumption. And it will be very difficult for you to make this case with two time periods.

If the earnings difference between men and women is only due to unobservable factors that remain constant over time, then you will be able to identify the causal effect of your policy. The problem is: you cannot test this. What people usually do is to show that the outcomes of the treatment and the control group moved parallel (or at least very similar) before the treatment occurred. Then you can make the case that the outcomes trends would have continued in the same way in the absence of the treatment.
The difference between this hypothetical continuation in the pre-treatment trend in the outcome for the treatment group after the treatment and the new trend in the outcome for the treatment group after the treatment is your difference in differences estimate, i.e. your treatment effect.

If there are time-varying factors which are unobserved and that affect your gender gap in earnings, difference in differences can't help you. For instance, if employers cut female wages in recessions by more than male wages, this is a problem. In order to make an argument that this gap is only affected due to time-invariant factors besides the female dummy, you would need to show this common pre-treatment trend in earnings between men and women.

share|improve this answer
    
Yes, you are right the females are the treatment group. As far as I am aware, the difference-in-difference estimator can be computed both in the way you suggested and my original formulation (Wooldridge Chap. 13, 4th ed) –  ajohnrobertson May 17 at 17:18
    
Oh okay, so I will be unable to estimate the causal effect of the policy if there are unobservable factors that affect wage which are not constant over time? –  ajohnrobertson May 17 at 17:28
5  
Exactly. And to argue that there aren't any, you would need to see the pre-treatment trends in the outcome for the treatment and control groups. –  Andy May 17 at 17:31
    
You are right about the equivalence of the two expressions, I edited the answer accordingly. I hope the post in general was useful for you nonetheless :-) –  Andy May 17 at 20:32
1  
Generally yes since $\widehat{\gamma}_3$ should be the same magnitude as $\widehat{\beta}_3$ but of opposite sign. If in the first regression women earned 5% more than men due to the policy, then men cannot also earn 5% more if you reverse treatment and control group. –  Andy May 18 at 0:06

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.