Comorbidity Analysis for In-hospital Deaths+

Additional analyses for paid subscribers

In the primary post for this analysis our eyeballing of a colourful slide provided two conclusions:

1. That there is no difference in the distribution of comorbidities for COVID-19 deaths during 2020-21 and influenza deaths during 2017-19;

2. That there is no difference in the distribution of comorbidities for COVID-19 deaths pre and post vaccination rollout.

Eyes are one thing but rigorous hypothesis testing another, so let us run the data through generalised linear modelling (GLM) with the parametric distribution set to Tweedie with identity link and parameter estimation based on Fisher’s method using maximum likelihood as the scale parameter. This sounds complicated but this is just a fancy way of conducting plain old analysis of variance for a variable - count of diagnoses made - that is ostensibly non-Normal. Herewith key outputs…

Hypothesis #1

This GLM model run compares the diagnostic count for the 186 influenza deaths over the period 2017-19 with 3,436 COVID deaths for 2020-21 using a simple dummy variable (morbidity comparison) to differentiate the data records. Right over in the column headed ‘Sig.’ in the lower table we see a value of p=0.713, this confirming that there is no statistically significant difference in the distributions of diagnostic counts between influenza and COVID. In terms of case complexity prior to death COVID and influenza are interchangeable.

Hypothesis #2

This GLM model run compares the diagnostic count for the 1,687 COVID deaths over the pre-vacccine with 1,749 COVID deaths over the post-vaccine period using a simple dummy variable (Vaccine indicator) to differentiate the data records. Right over in the column headed ‘Sig.’ in the lower table we see a value of p<0.001, this confirming that there is a highly statistically significant difference in the distributions of diagnostic counts for in-hospital COVID death pre and post vaccination rollout. This difference is not large, with an estimate of +0.19 more diagnoses made per electronic patient record in the post-vaccine period whether or not patients had been vaccinated.

The next logical step here is to break the analysis down further into vaccinated and non-vaccinated COVID deaths post vaccination rollout, but this is where it gets murky - vaccination rollout was far from random and we should expect to see case complexity associated with early vaccine uptake. In all likelihood this is what the +0.19 figure represents, with the association effect diluted as it gets smeared over all 3,436 cases.