Question for Panel Regression Robustness Check

[Ferdinand Thies]

Dear Statalist,


i am currently running a fixed effects panel regression with indicators for each time-period in order to assess the effect of an event over several weeks. This is a large-panel and we use weekly observations. We observe the emotional state of sadness:

In a nutshell the model looks like this with the IV as dummy indicator of a treatment

xtreg sadness i.week##IV , fe

Assuming we have 10 weeks of data for each individual. (week 1-10)

In order to check for the robustness of our results, i figured we could run a simple regression for each week (1-10):

reg sadness IV if week==1

reg sadness IV if week==2

reg sadness IV if week==3

etc.

Is this approach feasible as i cannot find a good source for this approach. Maybe I am missing something here?


Kind regards

Ferdi

[Carlo Lazzaro]

Ferdinand:
welcome to this forum.
If you state that you're dealing with a large sample, I get that you mean a T>N panel dataset.
If that were the case, -xtreg- is nit the way to go, whereas -xtgls- and -xtregar- come in handy.
That said, assuming that -xtreg- is the way to go, there are more relevant issues than time horizon to test, such as endogeneity, heteroskedasticity, autocorrelation of the epsilon error term and the correct specification of the functinal form of the regressand.

[Ferdinand Thies]

Dear Carlo,


i rather meant that the sample is large. In fact the panel is N>>T. We got N= 35000 and T=104.

We try to assess possible endogeneity problems and tested for heteroskedacity already, but want to show a clearer picture of the time horizon.
Running it with xtgls might also be good way of showing some additional robustness. Thank you!

Still, is the mentioned approach with these multiple cross-section regressions feasible here to show that the effects lingers for a certain amount of weeks? We are using the parmby command for that and plot the regression coefficients for better interpretability.

[Carlo Lazzaro]

Ferdinand:
therefore you have a short panel dataset (i.e., N>T).
That said, the issue here is not switching to -xtgls-, rather, with N=35,000, to invoke -vce(cluster panelid)- standard errors.
In addition, I would investigate the correctness of the functional form of the regressand.
If yuo went -fe-, you should altready have included -i.year- in the right-hand side of your regression equation. However, exploring how the coefficients behave on different timespans might be interesting as a sort of sensitivity analysis, especially if you're experiencing an increasing attrition.

[Ferdinand Thies]

Ferdinand:

In addition, I would investigate the correctness of the functional form of the regressand.
If yuo went -fe-, you should altready have included -i.year- in the right-hand side of your regression equation. However, exploring how the coefficients behave on different timespans might be interesting as a sort of sensitivity analysis, especially if you're experiencing an increasing attrition.

Thank you so much.
Alright, so we will label it as a sensitivity analysis. Just to make sure, the general approach of estimating a single regression for each week around the treatment is a sensible thing? I just couldnt find a reputable source to base this analysis on. Or is it just such a common thing to do that i am overthinking this a little bit?

[Carlo Lazzaro]

Ferdinand:
I misread 14 instead of 104.
Sorry for that.
It makes more sense if you consider 1 year (instead of 1 week) at a time.