Non-deterministic standard errors

[parkerwhitfill]

I am running into a problem using feols where the standard errors are not the same every time you run a regression! You can run a regression, then immediately run the same regression and get different standard errors. The point estimates do not change. Let me be precise about the situation I am running into this problem since it only appears in this specialized case.

I have a panel dataset. I have a time-varying Y, time-invariant D and a time-invariant Z. I am trying to run (Y ~ 1|year| year:D ~ year:Z) The coefficients on year:D have the changing standard errors. These standard errors move around and are sometimes all NAs.

I've attached the dataset where this error appears
error_df.csv

Load this dataset into R, then run
df$year <- as.factor(df$year)

results <- feols(data = df, y ~ 1|year| D:year ~ Z:year)
summary(results)

about 10 times. You can then verify that the standard errors are not the same in all 10 tries.

[lrberge]

Oooo, I confirm the very very odd bug.

Thanks a lot for the reproducible example! I'll try to take care of it soon.

[lrberge]

Hi, finally got it. There was a race condition that I didn't spot in the parallel loop computing the standard-errors. Now fixed.

Thanks again for the clear report!