RChung

Actually, at the time I wrote that a month ago, we were lacking a great deal of detail on racial and ethnic disparities in mortality reports: the initial data showed some info on age and sex, but with a long reporting lag that depended in part on whether the death occurred in a hospital setting or a nursing home or somewhere else. Since then, we've been starting to get more detailed data, which means that the modeling I was doing probably won't be necessary. Basically, we have a pretty good idea about age, sex, and race/ethnicity for all 3100 counties in the US, so we could do a Bayesian analysis to indirectly estimate differentials based on county-level compositional characteristics. We know that's too simple, but this might've been a way to identify unusual outliers, potential hotspots, and would give us a way to think about reliabilitiy of hotspots. At the same time, compositional effects aren't just a number on a spreadsheet; those compositions are the result of a long string of constraints over time, and that's always thought-provoking.


The connection to cycling is that we have large aggregative measures of speed and power and we're trying to indirectly back-out plausible values for rolling and aerodynamic resistance. The speed-power model is pretty well understood, so there are parts of that model that constrain what those plausible values are. Cycling data, no matter how lousy, are cleaner than the data I normally work with, so applying analytic methods developed for crappy data to relatively clean data turns out to work pretty well.