I am looking at linked CPS respondents between November 2019 and November 2020 based on household ID, age, sex, and race. My understanding is that I can use the IPUMS weight (LNKFW1YWT) to make point estimates about this population (means, proportions) but I’m not sure about variance estimates. For instance what information should I use to declare a survey dataset in Stata and then perform a regression analysis?
I have thought of 2 approaches -
Under a taylor series linearization approach I could set cluster/psu to the household ID and strata to the State FIPS.
Under a replicate weight approach I could adjust each cross-sectional replicate weight using iterative proportional fitting as described here (IPUMS CPS).
I’m no survey statistician. Am I on the right track, or what am I missing?
I don’t know if there’s a best practice for this particular problem. Because the CPS microdata doesn’t contain any design variables apart from the weights, researchers have pretty limited options for calculating standard errors. Generally for basic monthly data, researchers just treat the data as a simple random sample, assuming the clustering and stratification effects on variance more or less cancel out. If you decide to do this, just use the longitudinal weight as an inverse probability weight. Note this calculation will then be representative of the population of interest in Nov 2019, specifically those who are eligible to link to Nov 2020.
For more recent supplements (including ASEC), replicate weights do exist for more accurate variance calculations, though currently IPUMS only provides replicate weights for ASEC. It seems to me that if you’re using replicate weights, using IPF separately on each set of weights, and then using those adjusted replicate weights in your regression, would be appropriate, but I’m no survey statistician either. Your best bet might be to consult with a survey statistician (preferably at the Census Bureau) about this question.