I have two questions related to having fixed effects in the DD model.

I have a treatment that occurs at different times (e.g., 2001, 2005, etc.). I want to fit a DD model, so I standardize the treatment years to year "0" as the the treatment time. To control for treatment year heterogeneity, I included the true year fixed effects.

$y_{it} = \beta_0 + \beta_1 \text{Treat} + \beta_2 \text{After} + \beta_3 (\text{Treat $\cdot$ After}) + \eta (\text{Year Fixed Effects})+ \gamma C_{it} + \epsilon_{it}$

Question 1: Is there anything wrong with this model?

Question 2: Is there an issue with including time-constant fixed effects to this DD model? For example, what if I include i-level fixed effects ($\alpha_i$) and/or group indicators of i fixed effects (e.g. male/female or race)? I realize that DD cancels out time-constant i-lvl FE, but what if I include it here again?


1 Answer 1


The model is fine but instead of standardizing the treatment years there is an easier way to incorporate different treatment times in difference in differences (DiD) models which would be to regress, $$y_{it} = \beta_0 + \beta_1 \text{treat}_i + \sum^T_{t=2} \beta_t \text{year}_t + \delta \text{policy}_{it} + \gamma C_{it} + \epsilon_{it}$$ where $\text{treat}$ is a dummy for being in the treatment group, $\text{policy}$ is a dummy for each individual that equals 1 if the individual is in the treatment group after the policy intervention/treatment, $C$ are individual characteristics and $\text{year}$ are a full set of year dummies. This is a different version of the DiD model that you stated above but it does not require standardization of treatment years as it allows for multiple treatment periods (for an explanation see page 8/9 in these slides).

With regards to the second question you can include time-invariant variables at the individual level. You cannot add them at the group level (treatment vs control) because these will be absorbed by the $\text{treat}$ dummy. You can still include individual control variables like gender but note that they do not play a mayor role in DiD analyses. Their only benefit is that they may reduce the residual variance and hence increase the power of your statistical tests (see slide 8 here).

  • $\begingroup$ does this apply to cross-sectional data set. And how do we determine the treatment effec from the model you stated. $\endgroup$
    – user53983
    Aug 13, 2014 at 12:18
  • 5
    $\begingroup$ You need panel data for difference in differences because you require a pre- and post-treatment period. The treatment effect is $\delta$, the implicit assumption is that the treatment effect is constant over time but this can be relaxed if needed. $\endgroup$
    – Andy
    Aug 13, 2014 at 14:35
  • 1
    $\begingroup$ Andy, can you give some ideas on how to relax the implicit assumption that the effect is consistent over time? $\endgroup$
    – user001
    Jan 31, 2017 at 4:38
  • 2
    $\begingroup$ @user001 you can interact your treatment variable with the time fixed effects (leaving out one interaction as the baseline). The time interactions for periods before the treatments happen should be insignificant (the treatment can't have an effect before it even happens, otherwise sth is wrong) and the post-treatment time indicator interactions will estimate the fade-out time of the treatment. Somewhere I have given a similar answer which shows the regression specification for this. $\endgroup$
    – Andy
    Feb 5, 2017 at 23:57
  • $\begingroup$ Andy, I have a related question, what if I don't observe each individual in each time period. Let's say, I start with 10.000 individuals but number of individuals in the sample increases over time ? $\endgroup$
    – edyvedy13
    Mar 16, 2019 at 19:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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