implement claims-only hospitalization indicator

[mariajahja]

Hospitalization indicator based solely on claims based data. EMR data from HSP will be deprecated by Oct 1.

I have done checks in covidcast-indicators/claims_hosp/REVIEW.md, but of course will need a third-party code review. Much of the code is the same as the original delphi_emr_hosp package. In addition to logic changes to remove the EMR code, there has been some minor clean-up and naming changes (e.g. "sensor"->"indicator", "backfill"->"backwards_pad") as several terms have changed meaning over time. Pylint checks and unit tests are available.

Summary checks on differences between a claims+EMR signal and a claims-only signal:

1. The claims' denominator stream is 4x larger than EMR (except in the most recent timepoints). Despite this, we will lose estimates in 5-7 states. Historically, those states were Alaska, Maine, Montana, North/South Dakota, Vermont, Wyoming. In counties, we will lose ~40-100 (out of 400-1000).
2. The Spearman correlation over time (within each county individually) is comparable between a claims+EMR signal and an claims-only signal.
3. Spearman correlation by ranked geographies for a claims-only signal is slightly lower, but still has reasonable performance.

Additional details are available in the HSP slack channel (#covid-19-hsp).

[nmdefries]

Aside from the handful of minor errors (they don't appear to impact results), looks good. Nice documentation and use of classes.

[nmdefries]

Looks good!