14 Comments
Oct 7Liked by John Dee

Really interesting article - thank you.

I find the graphical representation of the data particularly helpful.

I eagerly await the next exciting installment of vax roulette!

I hope you will have sufficient data to show the age banded unbiassed effect.

Expand full comment
author

Ta muchly! There’s nowt like a colourful graph to clear the murk of staged multivariate logistic regression. Lots more to come…

Expand full comment
Oct 8·edited Oct 8Liked by John Dee

- I second the desire to see bias magnitude grouped by age, is it looks much bigger in the ages that matter most.

- Can you clarify in what department(s) the "future" recorded vaccinations occurred? Or do you not distinguish? Mind that vaccination in the ED could be a proximal cause of hospitalization.

- I'm not clear on why you limited to before vaccine rollout. Couldn't it be better to instead limit to after vaccine rollout? I would think that unvaccinated and "admission before dose 1" would better reflect different base health status in the time period we actually need to know what the level of bias was.

- When the goal is to quantify bias, my first impression is it would be best to only look at those who are still unvaccinated after hospitalized, and after rollout, and then make two groups. Though a problem with that and maybe other setups/questions is that getting vaccination in the system would indicate that person presently didn't have covid or probably any respiratory virus.

- Not sure if you have baked in a type of survival bias as of yet. The longer you stay in the system (i.e. get hospitalized), the more chance you have to get a vaccination. Are you in essence comparing people with longer stays to people with shorter stays?

- Would it be possible to do a formal test-negative study on this data against covid specifically (comparing unvaccinated to "admission before dose 1")? This would in effect report the effectiveness of a placebo vaccine that test-negative studies are likely to regurgitate.

Expand full comment
author
Oct 8·edited Oct 8Author

Age banding, whilst easy on the mind, is problematic in terms of sample size reduction which is why I derive propensity scores (risk scores) that do the job seamlessly.

We have no idea where vaccination took place, only when. Since we don't have dates of discharge we can't infer vaccination during the current ED admission or subsequent hospitalisation.

A study looking at ED admissions before an individual's decision to vaccinate is in the pipeline.

Since there is no survival data in terms of length of stay or status at discharge (alive/dead) then we can't adjust for survivorship bias, though health status at point of admission is not affected by this.

Looking at COVID only and non-COVID cases is a great idea but the CDS 010 data doesn't code for COVID!

Expand full comment

"We have no idea where vaccination took place, only when."

Right. This is key.

Expand full comment
author

Work continues using the ECDS enhanced dataset (4 more reports in preparation). At some point I’ll flip to survival analysis and start looking at time frames.

Expand full comment

Thanks.

Re: "At this point in time nobody had got jabbed..."

True, if you're only talking about the COVID shot... :)

Expand full comment
author

Well now, that’s a whole new packet of spaghetti!

Expand full comment

with meatballs!!

Expand full comment

as a paid subscriber to this substack i would very much like to get the free admission to the climate substack.

Expand full comment
author
Oct 11·edited Oct 11Author

Your wish is my command! Comp on its way... Thank you for your support.

Expand full comment

thank you!

Expand full comment