Is my svy, subpopulation results "omitted strata" connected to "stratum with only 1 PSU" issue?

[Sapna Varkey]

Hello,

I am using Stata 14 to analyze Medical Expenditure Panel Survey’s household component data. I’ve combined cross-sections from 2007-2015 and created a dataset in long format. I am running: svy, subpop(K6SMI): regress HAVEUS42 i.YEAR##i.LARGEGRP FAMINC i.MARRY PRIORITYCOND SMOKER AGE i.MALE i.LESSHSDEG i.HSDEGONLY i.SOMECOLLEGE i.COLLEGEGRAD i.WHITE i.BLACK i.OTHER i.HISPANIC

This command returns a note saying: 34 strata omitted because they contain no subpopulation members. Is this a problem? My subpopulation is those with moderate mental illness, so it’s understandable that there would be some strata without anyone with this condition.

I do know that I have strata with single sampling units. I realized this after running: svydes; svydes if e(sample); single, list VARSTR VARPSU if VARPSU==1. I got an extremely long list of observations that were in strata with single sampling units. If I’m not getting an error message about this issue when I run my svy regress command, do I need to deal with combining these strata with single PSU?

I’ve seen several examples of how to combine strata (Stata Forum posts from June 2007, Stata FAQ: What do I do when 1 of the survey estimators returns an error message, “stratum with only one PSU detected.”), but none that deal with a relatively long list of observations and strata to combine. If this is an issue I still need to deal with, can you help me with commands that would allow me to handle a long list of strata to combine?

Thank you in advance!

[Steve Samuels]

1. Strata without subpopulation members are not a problem. Stata will do the right thing when you include the subpop() option, as you do in your svy: regress statement will take care of the missing strata.
2. Standard errors depend on there being two or more PSUs in each stratum. When this isn't the case, the single-unit strata do not contribute to the standard error calculations. To work around this, Stata has the singleunit() option. This was discussed just recently here.

[Sapna Varkey]

Thank you Steve,

I did review those singleunit() options(https://www.stata.com/statalist/arch.../msg00245.html), but saw that it could also be useful to combine all strata with single PSUs together. I read about this option in a 2007 post (https://www.stata.com/statalist/arch...sg00043.html); what are your thoughts on this tactic?

Have you ever employed this tactic? If so, did you find it more appropriate than the singleunit() options?

I suppose this could be dependent on the dataset, but was hoping I could try all three singleunit() options and the combining option to see which adjustment would work best with my dataset. Unfortunately, I'm still having trouble writing commands that will combine a long list of strata with single PSUs to compare this 4th option to the other 3.

[Steve Samuels]

I don't recall that I've ever combined strata. The only basis for doing so is to combine "like" strata, such as geographical neighbors, or strata known in advance to be similar with respect to important demographic or study variables. Unfortunately, there's not enough information to combine strata in MEPS: The strata are taken from the National Health Interview Survey, but NHIS provides only pseudo-stratra (and pseudo-PSUs), not the original ones.