0
$\begingroup$

Can I estimate a diff in diff model to compare the effects of two different treatments that apply in different time periods in different countries?

I have 30 countries for an average time span of 34 years. All of these countries are subject to these two different shocks (treatments) that happen several times during the time span.

An example could be the following:

1) Treatment one: Austria 1972, 1989, 1990 Belgium 2000, 2002 Canada 1972, 1999, 2008, 2009, 2010

2) Treatment two: Austria 1990, 1992, 2014 Belgium 2005, 2014 Canada 1990, 1999, 2001, 2015

I have no un-treated countries.

Is it possible to estimate a diff in diff? Do you have alternative suggestions?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

For simplicity, suppose there are 3 countries, and you can extend the model into more countries.

Let $Y$ be your continuous response variable. $X_1 = 1$ for treatment 2 and =0 for treatment 1. $X_2 = 1$ for country 2, = 0 for country 1 and 3, $X_3 = 1$ for country 3 and =0 for contry 1 and 2. $X_4 = $year - 1972 (suppose 1972 is earliest year in dataset). Suppose the model is: $$Y=\beta_0 + \beta_1X_1 +\beta_2X_2+\beta_3X_3+\beta_4X_4 + \beta_5X_1X_2+\beta_6X_1X_3 +\epsilon$$

Suppose you want to test/estimate the difference between tow treatment on the difference between country 2 and country 3 at the given year $X_4=x_4$, i.e., $[E(Y|X_1=0, X_2=1, X_3=0,X_4=x_4) - E(Y|X_1=0, X_2=0, X_3=1,X_4=x_4)] - [E(Y|X_1=1, X_2=1, X_3=0,X_4=x_4) - E(Y|X_1=1, X_2=0, X_3=1,X_4=x_4)]$ .

$E(Y|X_1=0, X_2=1, X_3=0,X_4=x_4) = \beta_0 +\beta_2 +\beta_4x_4$

$E(Y|X_1=0, X_2=0, X_3=1,X_4=x_4) = \beta_0 +\beta_3 +\beta_4x_4$

$E(Y|X_1=1, X_2=1, X_3=0,X_4=x_4) = \beta_0 +\beta_1 +\beta_2 +\beta_4x_4 +\beta_5$

$E(Y|X_1=1, X_2=0, X_3=1,X_4=x_4) = \beta_0 +\beta_1 +\beta_3 +\beta_4x_4 + \beta_6$

Plugging them in to the diff in diff,

$E(Y|X_1=0, X_2=1, X_3=0,X_4=x_4) - E(Y|X_1=0, X_2=0, X_3=1,X_4=x_4)] - [E(Y|X_1=1, X_2=1, X_3=0,X_4=x_4) - E(Y|X_1=1, X_2=0, X_3=1,X_4=x_4)] =[\beta_2 - \beta_3] - [\beta_2-\beta_3+\beta_5-\beta_6]= \beta_6-\beta_5$

So $\beta_6- \beta_5$ is the diff in diff, and you can estimate and/or test it after your model fitting.

Improve this answer
$\endgroup$
6
  • $\begingroup$ Thank you very much for your useful answer. Can this works for multiple years? In addiction these two treatments are totally different from each other and there is no reason to suppose that in a given year we must have both, we actually have that sometimes you have only one of them $\endgroup$
    – Macrina
    Commented Nov 9, 2018 at 21:27
  • $\begingroup$ Yes, it works for multiple years, $X_4$ = year - 1972. Do not need to have two treatments at the same year. In summary, for each measurement of response variable, the country, treatment, and year are needed, no more other restriction. $\endgroup$
    – user158565
    Commented Nov 9, 2018 at 21:36
  • $\begingroup$ great. And what if the two effects are contemporaneous? $\endgroup$
    – Macrina
    Commented Nov 10, 2018 at 8:32
  • $\begingroup$ which two effects? country and treatment? $E(Y|X_1=1, X_2=1, X_3=0,X_4=x_4) = \beta_0 +\beta_1 +\beta_2 +\beta_4x_4 +\beta_5$, this is for treatment 2 in country 2. $\endgroup$
    – user158565
    Commented Nov 10, 2018 at 15:52
  • $\begingroup$ Clear. I mean that treatment 1 and treatment 2 can be present in the same country in the same year. What to do in that case? $\endgroup$
    – Macrina
    Commented Nov 10, 2018 at 20:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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