Everybody (here, here, here, here, here) is taking a look at the leaked New Zealand data, so why shouldn’t I? Be warned that this has been quick and dirty.
The data refer to a subset V of the population. All members of this subset are vaccinated, therefore V. Let M be the rest of the population. It is a mixed bag (some vaccinated, some not), therefore M.
According to my aggregation, 30,734 deaths have been observed in V in the 30-month period from January, 2021 until June, 2023. Here they are by age group:
I am mostly interested in comparison to all deaths, i.e., those in V+M (I am using the + for union of sets), which are published by Stats NZ:
They sum up to 92,139, but there are caveats. All numbers are “randomly rounded to multiples of 3 to protect confidentiality”. Whatever “randomly rounded” means, it applies at the level of sex (male and female deaths are given separately), so there are two errors adding up. Just forget about the age groups below 50 where deaths are rare. No signal can survive that amount of noise.
For age groups 50+, the monthly rates (deaths in V / deaths in V+M) computed from the above tables are thus:
Still, too much noise, so let’s drop the youngest and oldest age groups (argument regarding the latter to follow):
Now, these are looking rather similar, but what else can be said? There are two problems:
In each age group, V contains only a subset of V+M
Over time, people transfer from M to V
At the end of the period considered, there are around 2.2 million people in V, of around 5.5 New Zealanders in total. By age group, and neglecting the youngest, the proportions at the end are (again, according to my aggregation):
As you can see, I am making all kinds of stupid mistakes. How can V contain more than 100% of the population for the 95+ age group? The (main) answer is neglected turnover. We are looking at a 2.5 year period, and turnover (transfer to the next group, or death, which is the only option for 95+) can be expected to linger around 20%. That serves as my reason (announced above) to drop the oldest age groups.
For my pet age groups (70-84), I divided by the above proportions to address problem 1. Regarding problem 2, the only weapon in my arsenal is population vaccination rate (as available from OWID). I scaled this rate to 100% at the end of June, 2023, and divided by the result. That’s the picture I got (vaccination rate solid red, scaled vaccination rate dashed red):
Still, the correction seems to be insufficient. Maybe this is healthy user bias, maybe total vaccination rate is a bad proxy (for example, vaccinations in big government-sponsored vaccination centers seem to be missing from the data, and they might have been responsible for most of the turbo-charged vaccinations from August, 2021 to October, 2021), maybe it’s the turnover effect as well.
Anyway, even in 2023, and after all this scaling, I can’t push deaths in V to 100% of V+M. The vaccines still look beneficial, and even if they are not (and note that I strongly suspect that they are not), these data, though abundant, are of limited value.
"of limited value"... business as usual. Your article encourages me not to invest too much time in this data set. Thanks!