Skip to content

How much granularity do you need in your Mister P?

Matt Kosko writes:

I had a question for you about the appropriate number of groups in an MRP model. I’m currently working on streamlining some of the code we use to estimate state-level political opinions from our surveys. I have state-level predictors and Census data for poststratification (i.e., population totals in each age-sex-state-education cell), but I’ve also found some information about Congressional districts.

My question is, is there anything gained by including Congressional districts as another group if we just want to get state-level estimates? I know having cell populations with Congressional districts as another group is necessary for poststratifying and getting district-level opinions (which don’t seem to be very interesting for presidential elections), but does it do anything to increase the efficiency of any of the parameter estimates or is there another benefit?

My reply: I guess that including information at the congressional district level won’t really help you with state-level inferences—but it could. The way that CD-level info could make a difference is if there is nonresponse that varies by CD and is correlated with political outcomes, beyond whatever variables you’re already adjusting for in your MRP analysis. For example, suppose your survey undersamples rural whites, and you did not adjust for urban/rural/suburban in your model. Then it could be that including CD will fix some of this. In such a case I think a better solution would be to include urban/rural/suburban, as adjusting for CD is kind of a crude tool. So my recommendation is to start by thinking carefully about including relevant variables in your MRP model, and also to include relevant state-level predictors.

Leave a Reply