FE model relevance

[Angelo Pinzino]

Dear All,

I am new to STATA and would like to ask a few questions about my analysis. I am analyzing labor force survey data where individuals are participating 2 consecutive quarters and next 2 quarters, they are not participating. Thereby, in the data there are a lot of same individuals participating few times in the survey. I am analyzing the impact of educational mismatch on the wages. Firstly, I have done OLS regression without panel and by including all the respondents (because removing the same respondents could lead to biased results when I measured the required years of education necessary for jobs). My code is as following:

regress lnhourlygrosswage requirededucation overeducation undereducation industry_dummies year_dummies age agesq marital_dummy region_dummies firm_size_dummies if employed==1

Is this syntax correct? I have read that i. for time-invariant variables should be used, but i reckon in my case, all can vary by the time as the data covers 2010-2020.

Later, for controlling unobserved heterogeneity, I regressed FE model. I added panel identifier (unique household number + member number) and time identifier (yearquarter). The FE model syntax is as follows:

xtreg lnhourlygrosswage requirededucation overeducation undereducation yearquarter industry_dummies year_dummies age agesq marital_dummy region_dummies firm_size_dummies if employed==1, fe

My questions are:
1) Do you think my OLS and FE is correct, any of the variables should be time-invariant?
2) Relating with bias, do you think it is good to keep all the respondents or they should be removed?
3) Any better idea how to estimate?

Thanks in advance!

[George Ford]

do these sum to 1? (requirededucation overeducation undereducation)

you need i.xvar for any categorical variables (industry_dummies, region_dummies are categorical).

you'd be better off with reghdfe, which allows you to absorb multiple fixed effects (anything labeled "dummies" is likely a fixed effect, and you can put dummies in there if you aren't interested in the coefficient).

(It appears the next update of Stata allows you to absorb multiple FE with areg/xtreg. I recently updated _all and it still doesn't work.)

It's not entirely clear what your data looks like and what's going on, but xtset on individuals/quarter caught my eye. You need something in the model (education, I presume) that's changing at the individual level over time to get a coefficient.

Is this a balanced panel?

[Carlo Lazzaro]

Angelo:
welcome to this forum.
Your -regress- code differs from the -xtreg- onem in that, with default standard errors, you are treating the observations as independent (as they are not, when belonging to the same panel).
As per George's wise recommendation, if you want to calculate/include multiple fixed effects, you should switch to the communitty-contributed module -reghdfe-.
The -fvvarlist- notation that you implicitly mentioned (-i.industry-) holds for all the categorical variable (those time-varying too).
It would be weird to have time-varing variables only during the timespan a panel stretches over. A case in point is -industry-: firms are not expected to change industry as across panel data waves.
Eventually, as far as you -if- qualifier (- if employed==1-) you have to consider the risk of bias in your resulting sample: why some observations are -employed==0- in your dataset? Is their missingness informative or not?