NHIS: Singleton PSU's and sampling weights


I am using stata to work with IPUMS NHIS data from 2002, 2009 and 2016. When I tried to get weighted odds ratios of my outcomes, I got estimates without confidence intervals and a message saying “missing standard errors because of stratum with single sampling unit.” I was able to find the 10 strata w singleton PSU’s. Given the sampling strategy, do you recommend combining the singleton PSU’s with the strata before or after, deleting them, or something else? I’ve read elsewhere that deciding what to do with singleton PSU’s depends on how the data is collected.

A second, separate question: If I am using data from the 3 years to compare certain outcomes between years, do I need to divide the sampling weight by 3 (as suggested in other posts)?

Thanks for your time.


Regarding your first question, each of the methods you have proposed sound reasonable. I would recommend that you try each and see how they affect your analysis. If there is a strong preference for one option in your field, you may wish to follow that convention.

We provide some guidance on variance estimation using IPUMS NHIS here.

As for your second question: yes, you would want to divide your sample weight by the number of years of data pooled (in your case, 3). Information on the use of sampling weights with IPUMS NHIS can be found here.