Thread on MB hotspot vaccinations.

I've been thinking about yesterday's announcement. At first glance, it seemed like they were targeting health districts that have had the most cases since the Fall wave, but there was more to it than that. (1/n)

High case counts were given 30% weight while 70% weight was given to characteristics that are predictive of outbreaks, such as housing density and higher vulnerability of ethnic populations. (2/n)

Yesterday I questioned the first criterion (and being unaware of the 2nd) as it seemed to me that the present is a better predictor of the future than the more distant past. The Gretzky metaphor "go to where the puck is going" seems applicable here. Why chase the past? (3/n)

Now, if they figure out a way to chase where the puck is going, that is even better than chasing where the puck is now. Is this what is happening? (4/n)

The first criterion is testable through a statistical process called backtesting. There is tons of global data to do this with. Backtesting fits a model being blind to the eventual outcome. Then you take the blinders off and see how well the model performs. (5/n).

Backtesting can solve for the question of whether where cases will be e.g. 4-8 weeks from now. Why 4+ weeks? ~2 weeks to schedule an appointment + ~2 weeks for the vaccine to provide a decent level of immunity. And then you want a window in time to measure outcomes. (6/n)

How fast do the hotspots move over time? I think the predictive modelers are asserting that there are better indicators to use for prediction than where are the current hotspots. (7/n)

These modelers think the 2nd criterion is more substantive in predicting where these hot spots are. I think there is little question that the measures they are using work quite well worldwide. They also help predict severity of outcomes, and not just case counts. (8/n)

So, why are 2/3 districts not current hotspots? The current is just too short a sample size. This 3rd wave is being driven by variants, and I would suppose once the variants hit these vulnerable regions, they will light up. We've seen this happen before. (9/n)

So, the point of this thread is to encourage an open mind, and to point to the potential features of predictive modeling that may be in play with these decisions. Keep in mind that there may be well qualified experts behind these sorts of determinations. (10/n)

Predictive modeling can be useful for other possible criteria. E.g. targeting essential workers, teachers, grocery store employees, etc. My intuition suggests these are reasonable criteria. However, I have more questions than answers right now. (11/11)