Thread by @connolly_s

Thanks to spotting this retweet on @RanIsraeli 's feed I was able to update my basic Z-score analysis

https://twitter.com/rzioni/status/1423901461679849474

Here's the aggregate for almost the full month. A Z-score of 3.77 is, single tailed P<0.00008, double tailed P<0.00004. The Z-score of 5.28 and 7.00 is where the NCEST tables give up and say it's effectively beyond doubt. These are +ve Z-scores ie saying unvaccinated LESS likely

Caveat: This says there is a difference between the two groups... and it says for the 20-29, 40-49 and 60-69 age groups there is a very significant difference between vaccinated and unvaccinated rates... but it doesn't tell us what the difference is

Is the difference the fact that one group has been vaccinated or is the difference something to do with the probability of each group getting tested or is it something else entirely?

The stats will never tell you what is the cause of the difference. Only the process: hypothesis -> design experiment to test -> do experiment -> analyse results, can hope to reveal

BUT... a Z-score of 7!!! You can no longer sit idly by on the sidelines and pretend that there is no difference when the Z-score is 7... heck a Z score of 3 (i.e. P<0.007) would wet the pants of most scientists, especially in a biological field... you NEVER get 7 by chance

Note: if I see people making claims based on this thread of tweets and I do not believe are supported by the actual data, I will reserve the right to state what I believe can and cannot be supported by this data analysis

Note: I am relying on the accuracy of the source data. I do not speak or read Hebrew so my ability to verify from the Israeli data dashboard is restricted. I believe the source data to be correct but Google translate may have lead me astray

Note: I am human, so if I have made a mistake in aggregating the four weeks then I’ll update, however that’s double checked aggregation, so I believe it to be correct… even if 2200 twice seems unlikely (which is why I double checked 🤣)

Note: if you are not familiar with my previous analysis. I assume case numbers are Poisson so std dev = sqrt(N). Then I scale both N and sqrt(N) to case rate by division by population %. Finally we compute Z=difference between means/sum of std deviation