Results turned insignificant after applying fixed effects

[Kristin Wener]

Dear community,

I am sorry if my questions are stupid, I really struggle with econometrics.

I'm trying to run a regression to analyze the effects of childcare (enrolment and public spending) on maternal employment with 6 control variables, unbalanced panel. Initially, nearly all coefficients were significant and all signs made sense. After applying the country- and time-fixed effects, all variables except for GDP unfortunately turned insignificant and the signs are almost all counterintuitive. Instead of using the xtreg,fe command, I tried the areg command where my main independent variable suddenly turned significant, however, all other values were identical to the xtreg regression.

My first question is if there is anything I can do to find out why the results became insignificant or if there are any tests to check whether I have have any measurement error.
My second question is what exactly the difference between xtreg and areg is or if they can be used interchangably (In that case I would of course prefer the areg command as at least one my main independent variable is significant then).


Thank you very much in advance!

[George Ford]

That's not a surprising result, really. Often FE make variables insignificant, on account of the FE being a fix for some types of omitted variables. It could be a problem with the data, or it may simply be true. Measurement error won't the be cause. It's simply that the regression without FE is not accounting for mean differences across countries and time.

You also have the issue of GDP possibly being determined by maternal employment.

https://blogs.worldbank.org/en/developmenttalk/how-much-would-gdp-capita-increase-if-gender-employment-gaps-were-closed-developing

What does the prior literature do for fixed effects?

https://aspe.hhs.gov/sites/default/files/migrated_legacy_files//171046/EffectsCCSubsidiesMaternalLFPTechnical.pdf

https://labourmarketresearch.springeropen.com/articles/10.1186/s12651-023-00344-9

You can run with just country or just year FE to see which is causing the change, if not both. I suspect country is the killer.

What version of Stata are you using? Only the newest version (the updated v18) allows for multiple FE.

reghdfe does so and is the command of choice for multiple FE (though areg and xtreg in newest version may be useful for such models).

ssc install reghdfe

[Kristin Wener]

Hi George,

Thank you so much for your reply! I highly appreciate it!


Is there maybe a way to find out whether the results are real or whether there is a problem with the data?
Also thank you for your hint about reverse causality between GDP and maternal employment, is there any test I could run to see if that is true?
I tried running the regression with the reghdfe command, but sadly I get the same results (I have version 18)

I also tried removing the i.year from the xtreg command (I attached two pictures, with and without i.year) and the only big difference is that now in baseline regression my main indepedent variables are significant without the year-fixed effects.


I have the feeling that I made a mistake somewhere or didn't correct for things (e.g. outliers), but I am so unsure. Therefore, I would love to run some tests and see if I can do any better.

Thank you so much for your help!

[George Ford]

GDP gets you in the country FE only model, as it probably is capturing year effects. See what happens in the first 2 columns of the country/year FE model? GDP and Year are highly correlated. I'd drop GDP.

Still, the FE lead to nothing being significant.

What's the definition of the DV?