Apologies if you've already thought of this, but some quick points:
Thanks for the comments!
I agree with all of these points. In fact, with respect to (4) it is even plausible that some "once infected" people never go on to develop the kind of antibodies that are being tested for. Point (3) is why I control for age/sex, but of course there are a number of further complexities.
These further complexities, along with (1) and (2) are currently "un-modelable complexities" for me. There are just so many selection effects in play that it isn't clear if you gain anything from trying to take them into account. Given that there are a number of papers that try to calculate some kind of IFR making a number of basic mistakes, I wanted to set out to see what the data + simple models gets if you do the maths correctly, as opposed to doing ridiculous things like caring about the median study like the Ioannadis meta-analysis does. After all, it seems like the IFR is what everyone cares about, so it would be nice if we were doing things "less wrong" here.
This is mostly cross-posted from here , but edited slightly for the crowd here. Given that this is grand Bayesian Central, I'm surprised to see the lack of good Bayesian analysis as applied to covid-19. This seems to be a major failure of science, lesswrong should pick up the slack!
In my previous blog post, I had some comments on the statistical analysis present in some covid-19 studies. It is easy to criticize, but can I do better? Let's find out what the data tells us about the fatality of SARS-coV-2.
Data by itself tells us nothing
Well, almost nothing - remember that we are trying to find something that should be more generally true than just for the individual data points that we have. For this we need a model, a mathematical description of the process that created the data points including the unknown parameters (in this case, the IFR) that we want to discover.
Deciding on the model in advance
By tailoring a model to the data, one can strongly influence the results it shows. To avoid suspicions of this happening, and to avoid it happening inadvertently, one must/should always decide on the model before looking at the data. I have in fact seen some of the studies I'll be getting data from (see previous blog post). But I have not done any statististical analysis that looks at how age/sex affects the IFR. I will use this blog post to set a fixed methodology before I look further.
Garbage in, garbage out
No matter how good you are at statistics, your conclusions can only be as good as your data. I will therefore limit myself to studies of a certain type and quality. In particular, I will only look at antibody studies in Western countries that
As I only have so much time, I will not look for studies not covered by the two meta-analyses in the previous blog post.
The model
I start by binning ages into 10 year brackets, i.e. 0-10, 11-20, and so on, with a bin for each sex. I will label these by . For each age bracket and sex, an IFR will be estimated which I denote by . We assume that these IFR parameters are shared between studies, and do not depend on population characteristics other than age or sex.
In each study, the samples are assumed to come from some large total population (e.g. "people living in NYC"). I assume that within each age/sex bracket, the people sampled are random sample with respect to SARS-coV-2 infections. This is obviously not exactly true (for example, we can't sample people who have already died), but seems like a reasonable approximation.
For each study, I split the total population into subpopulations per age/sex group: "has antibodies" and "does not have antibodies". Some studies do both a PCR test and an antibody test. I do not know how to sensibly deal with this (the problem is that it's hard to decide which date to use for death statistics if we allow for active cases), so I ignore the PCR test.
I assume that in each age bracket, the participants in a study randomly either have antibodies or do not at the rate in the total population. In each case, they test positive based on the sensitivity and specificity of the test, which does not vary by age/sex. The positive rate for each subpopulation (i.e. true and false positive rates) is a random variable, about which we obtain information from the assay validation.
In each age/sex category of the total population, the proportion of those who who die of covid-19 after being infected is given by . I treat "has antibodies" the same as "was once infected".
If age/sex data is not binned the same way as here, I will either try to increase the number of bins, or treat "20 people aged 20-40" as "10 people aged 10-20 and 10 people aged 30-40". If a study uses intervals that are off by one from mine, I will pretend they are the same. If I don't see a way to salvage things, or think that combinations aren't sensible, I will exclude the study and make a note of having done so.
I will use the deaths counted by the end of the time of data collection of the study. This may miss some deaths. As it takes some time to develop antibodies we also miss some infections.
Priors
For each test, the priors for the true positive and false positive rates are uniform random variables over [0,1].
Likewise, the priors for the proportion of people who have antibodies are uniform random variables over [0,1].
IFR priors are uniform within each age/sex group over [0, upper limit of age group/100].
Things I might change
I might approximate some binomial distributions by normal ones. I might change the number of age bins if having too many becomes computationally intractable. I might use different priors for the IFR and/or proportions of antibodes, as it doesn't really make sense to assume these are completely independent random variables accross age/sex.
Request for comments
It would be great to have collaborators on something like this. I feel like something like this is potentially publishable in a journal, but then again peer review and writing an introduction is a pain.