Dealing with NHIS (from IPUMS) imputed income variables

[Mike McDonald]

Hello everyone.

I need some help with the income variables that IPUMS makes available in the NHIS database.
My scope is to estimate (through a DiD estimation) the effect of a particular insurance program on the health status of individuals. This insurance works this way: an individual gets treated if his/her income is, say, x% of a specific threshold.

Now, IPUMS makes available 5 income point estimate variables (incimppoint1, ..., incimppoint5), which I downloaded. Also, it provides 5 flags (which are exactly the same) indiciating whether the specific figures in incimpoint1, ..., incimppoint5 are reported by the individual or imputed.

My idea was to obtain a single income variable to use for separating treated and controls AND in the estimation too.

By looking at the mi-command help of Stata I think I grasped that DiD/TWFE estimation is not supporte by the "mi estimate" command.

Therefore my questions are:
1) Is this last sentence correct, i.e. is it true that I cannot do any DiD estimation including a multiply-imputed variable?
2) If this is correct, how can I deal with these 5 imputed variable for my purpose?

I know there are a couple of posts about (almost) the same issue, but they only helped me to a certain extent (see: https://www.statalist.org/forums/for...y-imputed-data and https://www.statalist.org/forums/for...-data-in-stata)

Thank you in advance.

[George Ford]

This might help.

https://www.stata.com/support/faqs/statistics/cmdok-option/

You could just estimate the DID using regression rather than the canned commands.

[Mike McDonald]

Thank you George Ford I'll definitely check if I can estimate using cmdok.
Do you have any tips for getting a single income variable to separate treated from controls? Thanks for the suggestion.

[daniel klein]

Do you have any tips for getting a single income variable to separate treated from controls? Thanks for the suggestion.

If there is uncertainty about an individual's income then there should probably also be uncertainty about their treatment status. Multiple imputation reflects this uncertainty by not providing one but multiple (plausible) values. My guess is that you would be better off with treatment status that varies over imputed/completed datasets.

[George Ford]

given Daniel thoughts, maybe take the extremes where you know for certain that they are in or out. of course, this creates a problem if there are other things correlated with income, which nearly everything is.

also, make sure you've exhausted all income definitions in ipums. I think there is a hhincome variable that is continuous.

[Mike McDonald]

Thank you both daniel klein George Ford for the advice! Really appreciated.
If got Daniel's suggestion right, I could build a treamant/control dummy for each of the point estimate variable to see whether the treatment status changed across the 5 imputed variables. If an indivdual is treated/control no matter what income estimate, then I should consider him/her as treated/control, if the treatment status changes across the NHIS imputed income variables, then I'll figure out what to do.
As for the continuous income variable, what IPUNS-NHIS provides are just reported categories, imputed categories, imputed point estimates and imputed earnings.

[daniel klein]

If got Daniel's suggestion right, I could build a treamant/control dummy for each of the point estimate variable to see whether the treatment status changed across the 5 imputed variables. If an indivdual is treated/control no matter what income estimate, then I should consider him/her as treated/control, if the treatment status changes across the NHIS imputed income variables, then I'll figure out what to do.

Why would you want to drop the observations with varying treatment status? My guess is that those observations would not qualify as a random sub-sample, thus potentially introducing bias.

FWIW, I do not like using pre-imputed values from data providers. The problem is that you cannot know which variables went into the imputation model(s) and, thus, you have no idea whether the imputations are compatible (congenial) with your analysis model. My guess is that you should probably impute (missing) treatment status along with (missing) income (and whatever goes into your analysis model).

[George Ford]

I was thinking he was using income (crudely specified) to create the treatment variable. I now think that is not the case.