The question nobody else has been able to answer is "What is a satisfactory analysis for these purposes?" The closest I've seen to an answer is to remove accid

The question nobody else has been able to answer is "What is a satisfactory analysis for these purposes?" The closest I've seen to an answer is to remove accidental deaths (eg, hit by a car, firearms fatality, etc) but there's a huge range of arguable cases like stroke which is plausibly covid-affected. Second, you have the problem of comparing stats between locations that used different definitions, so it's harder to do post-hoc correlations. The decisions seems to have been avoiding false-negative mistakes by increasing false positives in the first wave of analysis: "Let's include everything under a simple rule for covid stats so that we at least have some kind of worst-case baseline modelling with similar datasets, and we can figure out afterwards which are real and which aren't" Anyway, just wanted to see if you actually have a better answer to the problem than the standard that was used.