Ben Goldacre & colleagues just published a preprint in which they have screen electronic health records of 17 million NHS patients to explore deaths from COVID-19 by subgroup of co-morbidities etc.
While the work undertaken is impressive, I noticed that the hazard ratios reported in figure 3 is derived from a single Cox model including all covariates. This strikes me as odd if not inappropriate, especially considering the number of covariates and also that we don’t really have an overview on how these covariates relate to each other. This could have been illustrated in a directed acyclic graph which I think would have been useful when you have 21 covariates in the model, many of which are categorical.
I wonder if this potential overadjustment/collider bias may have induced the relationship observed for e.g. current smoker which appears “protective” and can be misinterpreted and misreported? Any insight and discussion regarding this would be very useful.