Post-stratification with missing subpopulations in the survey

Question

I have some survey data on a population described by age, gender, and weight. It’s quite skewed so I want to reweight it to a known target population (a larger study) using post-stratification.

Problem: some bins are absent in the smaller study (e.g. age 20-30, weight 120-150lbs, gender M). So when I try to map the population weights from the larger study using inverse proportional fitting, there's a chunk of the population that cannot be mapped. So you end up with an imperfect fit, and even the marginals don't look right (in that example, my overall size of subjects in the age 20-25 bin is off).

In addition there are some bins that have very low N so it's probably not advisable to map the full population weight of that cell onto, say, 1 person!

Is there some way around this? I could imagine some technique where you have a sort of adaptive re-weighting. If the N is too small to split into both gender M, age 20-25 and, say, weight 120-150 and weight 150-180, just group those two cells together into one. But of course there are multiple directions where you could merge cells! Are there some better techniques?

Answer 1

I think merging cells is a fairly common approach -- you need to merge cells that are similar a priori or based on the population information, not similar based on the sample.

McConville and Toth have a recursive partitioning approach that goes in the other direction: start with a single group and recursively split into smaller and smaller post-strata.