Neil D writes:

I’m starting to work on a project which I think might benefit from some multilevel modeling, but I’m not sure. Essentially it is a multiple regression model with two explanatory variables. The intercept is expected to be close to zero.

Over the time the data has been collected there have been six changes to the relevant tax rules in my country, and what has been done so far is to fit a model where the regression coefficients are different in the seven tax regimes. I’m thinking that some partial pooling might be helpful, and might improve the estimate of the the regression coefficients in the last tax regime, which really is of major interest.

I haven’t done the analysis yet, but I’m assuming that such an analysis would be worthwhile and relatively straigtforward assuming that the regression coefficients in the different tax regimes are bivariate normal. What worries me a little, particularly as the analysis is sure to be scrutinized and criticized by various interested parties, is the assumption of exchangability, since as the different tax regimes are introduced there was an expectation that the regression coefficients would both go up in some cases, or both go down, or one would go up but the other would likely be unaffected. I’m not sure if it is possible to incorporate this information.

My reply: I’d start with the secret weapon.

P.S. Whenever I put “multilevel” in the title of a blog entry, I get spam from multilevel marketing companies. Could you guys just cut it out, please?