As a few of you know, I’ve been drafted into modeling the spread and consequences of COVID.

We’ve learned some important modeling lessons along the way.

I know a few of you are working on modeling too, so I’ll offer some sporadic thoughts and code snippets on this thread. https://twitter.com/ameliatd/status/1245040737764478980

First of all, I’ll echo my comments of the last few days: calibration is key.

If you construct a model and it can’t project what we know today (observed # cases, # hospitalizations, etc.), then you need to refine your model.

Second, given tremendous uncertainty, for any given model there are often multiple scenarios consistent with the observed data.

Embrace this, don’t ignore it — if you find that the policy recommendations are consistent with multiple scenarios, then you can feel more confident.

3) Since we didn’t have the luxury of being able to generate a detailed microsimulation, we’re using the chassis of classic epidemiology models and tailoring to COVID from there.

We’re finding that SEIR (as opposed to SIR) is more consitent with observed data patterns to date.

4) Preliminary analyses indicate that that ZIP-hospital market shares in the FFS Medicare Hospital Service Files roughly match overall market shares (across full population) in (retrospective) steady state …

…so you can use those to predict out hospital service use if you have ZIP-level case data.

BUT a lingering Q is whether “elasticity” of tertiary hospital use is higher with COVID — suggeting greater burdens on those facilities than steady-state market shares would indicate

(so if someone has bandwidth and access to claims data, looking at differential H1N1 care use patterns by hosptial type would be useful)

5) A really nice way to plot case series is not do to it over time, but to plot log10(total cases) vs. log(10 rolling average of total case growth over last XX days)

See here, for example: https://aatishb.com/covidtrends/ 

This traces out a case trajectory — and when a county/state/country turns the corner, it will show up as a precipitous fall in the trajectory. Makes it easier on the eyes to see when exponential growth moderates and falls.