How to handle missing data in xtreg, fe

[Tongtong Li]

I have unbalanced panel data where the dependent variable (DV) is only recorded if it stays above a "entry barrier." If the DV falls below this barrier, the entire observation for that time period is missing (not zero, but a null entry). Crucially, this barrier's value changes over time.

I want to estimate how a specific treatment affects this DV using Fixed Effects. If I only analyze the "surviving" observations that stay above the barrier, I am worried about survivorship bias.

What is the best practice in Stata to correct for this?

[Carlo Lazzaro]

Tongtong:
welcome to this forum.
My guess is that you can only specify that yout -xtreg,fe- refers to obs whose dependent variable is above the threshold.
Obviously, this is a biased subsample of the potential study population.
How this kind on analysis are acceptable in your research field, I cannot say.

[Felix Bittmann]

Just a few ideas:
1. You can try to impute the missing data using mi impute chained. However, for this to work, at least some observations need observed data. Given your data, this might not apply.
2. xtheckman
3. If you cannot look at the actual value, you can try to at least understand the missingness mechanism. Recode the DV (0 if observed, 1 if missing) and work with this binary DV. This might help you understand what is the cause of the missingness.

[Jeff Wooldridge]

This is a difficult problem, falling into the "truncated regression'' category. Just applying FE to the complete cases doesn't solve the problem -- certainly not completely -- because the DV has to cross a (known, I assume) threshold. If you always observed the covariates x(i,t) then relatively simple solutions exist because you could use the Mundlak device and then apply pooled truncated regression with clustered standard errors. If the covariates are independent of the unobserved effect then this will work without observing the covariates, but the whole idea of using fixed effects means you think c(i) is correlated with x(i,t).

You might approximate the Mundlak solution by averaging the covariates across those time periods for which you observed x(i,t), but I doubt one can establish good statistical properties.