0
$\begingroup$

When it comes to causal inference, if I use the entire population data, is Inverse Probability Weighting (IPW) still effective?

I have access to the entire population data and need to conduct some regressions to adjust for some confounders. Based on my understanding, IPW simulates a pseudo-population to mimic the entire population data for exchangeability.

Therefore, I believe that adjusting for confounders is okay, but I am still unsure whether IPW is essential or not.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$

You may have the population values of the observed outcomes, but you don't have the population values of the potential outcomes, which is what you need to estimate a causal effect. That is, if your quantity of interest is $E[Y^1] - E[Y^0]$, i.e., the average treatment effect (ATE), having the values of $Y$ for the population doesn't mean you don't need to do any analysis to estimate the causal effect. If you want to estimate a causal effect, you still need to use the assumption $E[Y^a] = E[E[Y|A=a, X]]$ (i.e., conditional exchangeability) and use a method that allows you to validly estimate $E[E[Y|A=a, X]]$. IPW is one such method (among many others). With population data, you may need to do work to estimate standard errors correctly since most methods of doing so assume your sample is drawn with replacement from a super-population that is much larger than it.

Improve this answer
$\endgroup$
4
  • $\begingroup$ Thank you so much! $\endgroup$
    – user1190107
    Commented Apr 17 at 18:24
  • $\begingroup$ One more thing, is it okay to use usual IPW package in R or STATA for the entire population data? I am worried about the standard error thing what you mentioned. $\endgroup$
    – user1190107
    Commented Apr 17 at 18:29
  • $\begingroup$ To get valid estimates of the treatment effect, sure. To get standard error estimates that are valid, I cannot say. Maybe someone with more expertise in this area can chime in. We can known for quite certain that the conventional standard errors will be too big at worst. $\endgroup$
    – Noah
    Commented Apr 17 at 18:48
  • $\begingroup$ Thanks. I will dig deeper. $\endgroup$
    – user1190107
    Commented Apr 17 at 19:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.