Thread by @DevanSinha

This is good news.

The longer it takes to statistically tell difference between vax effectiveness against variants the smaller the actual dip (if any) in protection will be.

1/

https://twitter.com/jburnmurdoch/status/1396015155872509953

In randomised control trials we can be confident in vax effect after only 100+ infections because selection bias and confounding variables between the vaxxed and unvaxxed comparison populations are (e)limited by the randomisation process.

2/

This is not the case when we do retrospective observational studies - like vax effectiveness against variants in the field.

The statistical analysis is more challenging. There are biases in who is vaxxed or infected, living, working, mixing patterns, medical health/immunity.

3/

To account for these need to do matching/case-control or logistic regression to find true odds ratio between infxn risk of vaxxed vs unvaxxed.

cf RCT where Chi-squared or T test test could do.

4/

NB I still do simple short cuts for twitter analysis with point estimates of field vax effectiveness...

But if we're being proper clinical scientists trying to get robust publishable/usable data we should have higher standards!

5/

For proper analysis the retrospective cohorts have to be much larger with many events (infections). ~10X or more vs RCT.

So far there have only been 3.4k B1.617.2 cases identified. Predominantly in unvaxxed young people - so the vaxxed v unvaxxed groups will be v unbalanced.

6/

To get to a point when confidence intervals no longer overlap it will take many more weeks if there is a small difference.

Applies to 1 dose only effectiveness too which is even more challenging due to differing time windows of population in that state/community infxn rate.