Australia 0 - 0 Austria

Luke 3:5

Welcome to the 2024 Covid Vaccine Effectiveness Paper (CoVEP) games. Today, Australia is playing Austria. Kick-off Australia!

Roughly 4 million Australians aged 65+ are included in the study, which covers January-May 2022 (3250 Covid deaths, or maybe 3235, the paper actually contradicts itself) and June-November 2022 (3185 Covid deaths). Vaccine effectiveness is studied for nine categories, distinguishing by number of doses (0-4) and time since last vaccination. Overall Covid death rates per 100 person-years (PY) are 0.206 = 3235 * 100 / 1,578,000 for January-May, and 0.167 = 3185 * 100 / 1,906,000 for June-November.

The main results of the paper are depicted in Figure 1, which I reproduce with my supplements in red:

More doses and shorter time since last dose seem to result in higher vaccine effectiveness (VE). However, strange things are going on. In the “raw” column, I computed raw VE from rates (one minus quotient of rates). For example, for January-May, the death rates for “unvaccinated” and “Dose2 >180 days” were almost the same, resulting in zero raw VE. How do the authors arrive at 34% VE? Discrepancies like this one seem to be correlated with small sample size (as indicated by the size of the blue squares).

Moreover, the death rates in the single categories differ vastly between the two periods. And it is exactly those with more recent vaccination (number of doses does not matter) who are doing worse (red boxes) in June-November than in January-May.

Now, I would love to dig into the data and find explanations (for example, I would investigate the stories behind the much larger proportion of people with no comorbidities, and fewer GP visits, among the unvaccinated). But, unfortunately, what I mentioned so far (total numbers of deaths in the two periods, and death rates per category) is all that can be found about deaths. There are four giant tables, but they only slice populations (according to age, sex, income, and everything but favorite color), not deaths. And that information is useless anyway because it only refers to the starting points (1 January, 1 June). Figure 3, allegedly even showing benefit of vaccination to all-cause mortality, only lists VE, not death rates. There are three files with supplementary data, a useless one with decision trees and some more figures (VE against everything, hallelujah), a useless one with a table of types of vaccines received, and (guess what) a useless one with methodological remarks and source code.

Why is that last one useless? Because there may be eight pages of code, but the hidden action is taking place inside the few lines like this one:

#fit the model for competing risk analysis
compRisk0 <- coxph(Surv(t1, t2, F_UCOVDTHINTV==1) ~ VAXSTATUSC + SEXCD + agecatc + GPviscatc + comorbcatc +STEUCPc + HIEDcatc + F_FLU21C   ,
                   ties="efron",
                   data=COVDS3)

Here they use Cox regression (Oh no! Not! Again!) to fit a proportional hazard rates model to the death events data. And that black-box model is what produced the splendid vaccine effectiveness results. Yes, they might have stared at the Schoenfeld residuals while chewing on their Vegemite sandwiches, but they did not share them (the residuals, that is) with us.

To conclude: a bunch of government-adjacent authors, another vast government database we do not get to analyze, a paper that displays lots of figures without any chance for readers to verify, a statistical method that has become the standard way to replace thinking. That’s what gets you into Lancet.

Counter attack Austria. With roughly 4 million subjects (of 9 million Austrians), total sample sizes are almost equal. The selection of subjects, however, is different: age does not play a role, but previous Covid infection is required. The study covers November-December 2022 (69 Covid deaths), and then January-June 2023 (225 Covid deaths). In general, vaccine effectiveness against Covid death (and Covid infection, but I do not trust that endpoint) is found to be indistinguishable from zero.

Now, as regards transparency, I am much happier with Austria than with Australia (maybe because they strengthened their defense by signing John PA Ioannidis). The tables are informative, and the “supporting information” is actually that (P.S.: next time, please do not only claim to have stared at Schoenfeld residuals of your Cox regression, but show them).

But if you select for previous infection, and your database is so wonderfully complete, why on God’s green earth don’t you compare the results to those for the previously uninfected?

And why are we even discussing “vaccination” against a virus that kills (or is present at death of) around one person of 4 million, per day? Seriously, this is only around twice the number of traffic deaths in Austria. If I secretly swapped the Covid deaths for traffic deaths in the raw data, would the results be different, and would anyone even notice?

Let’s call it a draw. A boring, pathetic-vs-miserable, goalless draw.