Modelling Policy Measures in Fixed Effects Panel Regression?

[Nikolaus Schueler]

Dear all,

I have a general question regarding the modeling of policy measures:

I am investigating expenditures of municipalities in a german state between 2009 and 2020 (N = 395, T = 11 years) by using a fixed effect panel data regression with a balanced panel. One central point of interest is the investigation of a policy measure that took place in 2015 and lasted until 2020. This is the reason why I am using a fixed effects model as I am intrested in intra-municipal changes over time especially due to this external shock (hausmann test recommended it as well).

This policy measure affected all municipalities, so there is no differentiation between a treatment and a controll group. Therefore, Difference-in-difference estimation does not apply in this setting. Since it is a before/after investigation, maybe a regression discontinuity design could be an option, but I am not familiar with this design yet...

Do you know in which way such policy measures could be modelled?

I woul be glad to hear from you!

Best regards
Niko

[George Ford]

Discontinuity is also a treatment effects type model with a running variable. Probably stuck with a means difference model, unless the treatment varied in some way by municipality, or there's something about the municipality you know that would cause a differential response to the same stimulus.

[Nikolaus Schueler]

Dear George,
thank you for your response! There is no variation between the municipalities regarding the treatment.

So you suggest to run a test for difference in means before and after the treatment, right? Since I am using a panel structure, the municipalities in both "groups" are not independent. So a paired sample tests wouldn't work, right? Would your rather recommend using a Wilcoxon signed rank test?

Best regards
Niko

[Carlo Lazzaro]

Nikolaus:
why not sticking with -xtreg,fe. and posting the outcome table? Thanks.

[Nikolaus Schueler]

Dear Carlo,

this is the output of xtreg,fe with cluster robust standard errors.

I implemented the policy measure ("MOD") as a dummy variable. It switches to 1 when year_>2014.

The appendix "_pc" means "per capita", "Q" means "Quota".

The dependent variable has been logarithmized.

I added year dummies in order to controll for time fixed effects, but I have some doubts about it because I am afraid the MOD variable might interfere with them.

If you need any further information please tell me!

Thanks a lot!
[ATTACH=CONFIG]n1688576[/ATTACH]

Version Stata 16.1 SE

[Carlo Lazzaro]

Nikolaus:
1) you can test the joint singnifcance of -i.year- via -testparm-;
2) cluster-robust standard error is OK given the number of your panels;
3) the within-Rsq is good;
4) I'd double-check the evidence of a panel-wise effect.

[Nikolaus Schueler]

Dear Carlo,
thanks for your response!

Nr. 1) -testparm- is telling me that all year dummies are significantly different from zero, so this means there are time fixed effects, right?
Nr. 4) What do you mean by that?

Best regards
Niko

[Carlo Lazzaro]

Nikolaus:
1) correct (even thugh you actually you cannot retrieve time fixed effect from -xtreg-);
4) you can repeat the regression with default standard errors and see what the F-test appearing as a footnote of the outcome table tells you.

[Nikolaus Schueler]

Dear Carlo,

thank you for your hint!

I repeated the -xtreg, fe- without the use of any vce. My interpretation of the F test is that the observed and unobserved fixed effects are significantly different from zero, so there is a need for taking panel effects into account, correct?

Best regards
Niko

[Carlo Lazzaro]

Nikolaus:
yes, there's evidence of a pane-wise effect. Go -xtreg,fe- then.

[Nikolaus Schueler]

Thank you, Carlo!

[Nikolaus Schueler]

Hi everyone,

I have an additional question regarding the question of modelling police measures: I found several papers on the subject of implementing panel events, e.g. by Damian Clarke (https://www.iza.org/publications/dp/...el-event-study). Even though it especially deals with estimating Diff-in-Diff and lagged periods, I want to ask you if anyone of you has any experience with the "eventdd" command in Stata. Do you know if it can be applied for a treatment with no control group as well?