[burger]

From the original paper, "We adjust by population size, hospital beds, number of individuals tested, weather, and socioeconomic and behavioral variables including, but not limited to obesity and smoking. We include a random intercept by state to account for potential correlation in counties within the same state."

The possibility of residual confounding cannot be discounted, however.

BC

Just dug in the supplementary materials.  When they excluded counties with <10 cases (those should mostly be rural, but i haven't downloaded the data to look), they got non-significant results for 3 of 4 tests.  The population size adjustment did not change the results, which is contrary to what I expected.  I'd really want to dig into the data to understand the interaction of population size and death rate.

Anyway, I still smell a rat here.  Overall effect size is small and P values are ~0.01 despite a brobdingnagian sample size of counties.  Also, they didn't adjust for spatial autocorrelation.  Autocorrelation in pollution is obvious from their maps, so I would bet they have spatial autocorrelation in their residuals.  If so, that would violate model assumptions and call the results into question.  If I get some time (unlikely), I'll download the repo and look at the residuals.

If I was a reviewer on this paper, I would send it back.

[burger]

From the original paper, "We adjust by population size, hospital beds, number of individuals tested, weather, and socioeconomic and behavioral variables including, but not limited to obesity and smoking. We include a random intercept by state to account for potential correlation in counties within the same state."

The possibility of residual confounding cannot be discounted, however.

BC

Just dug in the supplementary materials.  When they excluded counties with <10 cases (those should mostly be rural, but i haven't downloaded the data to look), they got non-significant results for 3 of 4 tests.  The population size adjustment did not change the results, which is contrary to what I expected.  I'd really want to dig into the data to understand the interaction of population size and death rate.

Anyway, I still smell a rat here.  Overall effect size is small and P values are ~0.01 despite a brobdingnagian sample size of counties.  Also, they didn't adjust for spatial autocorrelation.  Autocorrelation in pollution is obvious from their maps, so I would bet they have spatial autocorrelation in their residuals.  If so, that would violate model assumptions and call the results into question.  If I get some time (unlikely), I'll download the repo and look at the residuals.

If I was a reviewer on this paper, I would send it back.