1
$\begingroup$

I am trying to estimate the impact that the onset of a medical disease has on a number of outcomes, call them $O$. To do this, I am using an event study model with individual and time fixed effects... \begin{equation} O_{i,t} = \alpha + \sum_{\substack{k=S \\ k\neq -1}}^{F}{\mu_k} + Z_{i,t}\delta + \sigma_i + \epsilon_{i,t} \end{equation}

I have theoretical reason to believe that characteristics of the individual at the onset of the individual may moderate the effect of disease onset, and so I would like to explore heterogeneity in the treatment effect by this at-diagnosis characteristic. For instance, an individual's marital status at diagnosis may moderate his response following diagnosis.

I apologize if this is a trivial question but how can I incorporate this time invariant characteristic into a model with the individual effect? Should I simply run subsample analysis by group? And if so, what is a good strategy for dealing with continuous (but fixed) heterogeneity -- for instance, heterogeneity of income at diagnosis? Should I drop the individual fixed effect?

I appreciate all of your assistance and thoughts in advance.

$\endgroup$

1 Answer 1

1
$\begingroup$

If you have a time-constant covariate $X$, you could include it but drop the individual fixed effects $\sigma_i$. (There could be perfect multicollinearity if you kept them.) Since you mentioned that $X$ might have a moderating effect w.r.t. $Z$, you would include interaction terms between $X$ and $Z$. So you would have something like $$ O_{i,t} = \alpha + \beta_t + \gamma x_i + \delta Z_{i,t} + \theta (x_i\cdot Z_{i,t}) + \epsilon_{i,t} $$ with $\beta_t$ being time fixed effects. You could set $\beta_1=0$ for identification. (One more linear restriction might be needed.) From your verbal description of the problem, I am not sure what $\sum_{\substack{k=S \\ k\neq -1}}^{F}{\mu_k}$ is meant to represent, so I did not include it in the equation above.

(I might be completely wrong. It has been a long while since I did anything with panel data models.)

$\endgroup$
4
  • $\begingroup$ Without the individual fixed effect, it would not be a DiD anymore? I think a better approach would be to keep the individual fixed effects and run two separate estimations for the sub-samples. $\endgroup$
    – Papayapap
    2 days ago
  • $\begingroup$ @Papayapap, hm, I don't know. Is that possible if $X$ and the individual effects are perfectly multicollinear? Now regarding your suggestions, how would you define the subsamples and what estimations do you have in mind? $\endgroup$ 2 days ago
  • $\begingroup$ Marital status is binary, so we can use two sample for marital status=1 and marital status=0 $\endgroup$
    – Papayapap
    yesterday
  • $\begingroup$ @Papayapap, for a binary variable like that it sounds like a good idea. $\endgroup$ yesterday

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.