I have data from a survey, stratified by country with a roughly equal number of participants in each country. Also, I have calibration totals, three dimensions (say, age, sex and work status) plus the country dimensions. A few of the category combinations of the calibration data have not been covered by the survey.
I have merged categories using two ad-hoc strategies to allow for poststratification. For the first strategy, I combined "neighboring" categories (e.g., age 50 and age 55 of male unemployed); the second strategy consisted of collapsing entire categories (e.g., treating the unemployed and the "working at home" as one stratum). In addition, I calibrated against marginal totals using GREG estimation.
My question is: Is there some general guidance which weights would be preferable, and why?
The client provided us with indicators for which the weights should be checked. The point estimates vary slightly after the reweighting, but the variance by country increases for some countries and decreases for others. Can I decide from the data which weights to prefer?
