3
$\begingroup$

I am interested in determining the covariate balance of a PS model with high-order interaction terms. How can assess this? I have explored the use of standardized differences but I am running into problems because of the interaction terms in the model.

$\endgroup$
4
  • $\begingroup$ Have you tried the package MatchIt with its balance assessment functionality? $\endgroup$
    – frank
    Jul 16, 2022 at 11:04
  • $\begingroup$ Curious what made PS a good choice for your problem. $\endgroup$ Jul 16, 2022 at 11:57
  • $\begingroup$ @FrankHarrell My study is observational and I have many confounders to adjust for. Thus, to avoid the finite sample bias problem, Ps was a good choice because of my small sample $\endgroup$
    – Magut
    Jul 16, 2022 at 12:28
  • $\begingroup$ Sounds good. You might consider interacting the logit of propensity score with selected individual predictors. $\endgroup$ Jul 16, 2022 at 14:53

1 Answer 1

2
$\begingroup$

The form of the propensity score model has nothing to do with how you should assess balance. You should always assess balance on as many terms as possible, including interactions and other features of the covariate distribution. Different software make this easier or harder. If you are using cobalt to assess balance, you can just set int = TRUE in the call to bal.tab() to get standardized mean differences on all pairwise interactions. If you are using MatchIt, you can do the same in a call to summary(). If you want to test balance on specific interactions or higher-order terms, you can manually add them using the addl arguments to both functions.

There are some balance statistics that attempt to assess balance in the full joint covariate distributions (i.e., all possible interactions), like the L1 statistic and energy distance. These are good for comparing matching specifications but not necessarily for deciding whether a sample is "balanced" or not.

$\endgroup$
1
  • $\begingroup$ ,Thank you for your answer $\endgroup$
    – Magut
    Jul 17, 2022 at 10:19

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.