Yep! Different estimation framework entirely but similar in spirit to this approach. I haven't read the paper you linked but looks like it builds on this one which influenced my thinking on this problem academic.oup.com/restud/artic...

In principle yes. Data requirements are: individual level Y1, Y2, X (survey); ground-truth joint distribution of X (census); known marginals of Y2 (auxiliary data). Also need svy to include relevant geographic vars to link to ground-truth Y2 (though this can be relaxed, see sec 5 of the paper)

This package is a companion to my paper with Josh Clinton on MRP calibration (osf.io/preprints/so...)

Please let me know if you try it out the package and email me or file a github issue if you find bugs. Thanks!

This function also works to implement the "logit shift" calibration with a single outcome — e.g. to ensure your subgroup estimates of vote choice are consistent with known election outcomes.

The package supports both full Bayesian calibration (i.e. calibrating each draw from the posterior) as well as a plug-in estimator which works with posterior summaries.

calibratedMRP provides a high-level interface for doing this type of calibration. First, specify a `brms` multilevel regression model. Second, call the `calibrate_mrp()` function to calibrate cell-level estimates to known geographic margins. Third, poststratify to the target groups.

calibratedMRP implements a calibration procedure that accounts for known margins of Y2 to improve estimates of Y1.

Intuition: if your MRP model overestimates Biden vote share, it probably overestimates liberal policy attitudes too. The method implemented here accounts for this discrepancy.

Suppose you have a survey that measures policy attitudes (Y1) alongside demographics (X) and related behavior (Y2). You have population data on the joint distribution of X from the census, but only coarse geographic aggregates (e.g. election results) of Y2.

I have never seen Casablanca [embarrassing admission] but am resolved to watching it this week