jhm
Well-Known Member
I was reluctant to model cases in the first case because of the differences in test levels between states. Now we've got testing going on in all the states. So perhaps the differences wash out. It is just an indicator of the intensity of the epidemic. So are deaths. What I try to keep an eye on is the consistency between different types of measurements. The rolling 28-day regression will keep adjusting to changes in the quality of the data as well as the correlations. Abrupt revisions are a bit hard to take, but gradual changes are fine.Interesting, so it might still serve as an indicator of the current trend.
So how will it respond if known cases grow partially because of increased testing? Will it adapt? Could testing be another "leading indicator", or is the data too limited to derive an improvement from that?
Because of many states reopening to various levels, there may be/will be another sub-wave soon. Wave 1.5. So we will see. (Probably.)
You can play around with testing data. An increase can make a small increase in cases, but in most places the mostly likely cases are given priority for testing. So increases in testing can have "diminishing returns" as the marginal test has lower chances of being a positive.
The problem in using test in a forecast is that you also would want to forecast the tests too. It's questionable whether ones attempt to forecast future test availability really adds much to the forecast of endpoints you care about. So you can try it, but may find that it is not so helpful.
As a statistician I'm always happy to work with lousy data. People are always dreaming about the data they wish they had. But good statistical practice makes the most of the crappy data that you do have. Bringing cases and deaths together is good because there is a common underlying process generating both and the correlations between them.