Reader Anonymous was skeptical that previous correlations between deaths, hospitalization and ICU shown in this post use would continue to hold, specifically, in a first differences specification:
Δldeathst = 0.00045 + 0.276Δlhospt-21 + 0.270Δ licut-14
Adj R2 = 0.07, SER = 0.048, DW = 1.69, N=529, bold denotes significance at 5% msl, using HAC robust standard errors.
As critique, he wrote:
1. Omicron is a very different variant than Delta (or alpha, beta), especially in terms of per case impact. So training the model on earlier Covid has some real uncertainty here.
2. In particular, the model is trained only with US observed lags (fine, since we’re looking at US to come, but then you’re not capturing insights from other countries where Omicron has swept through. Seemingly this info ought to affect a Bayesian prediction somehow.
3. Recent (DEC) data may be heavily contaminated with Delta, but that is changing to Omicron very rapidly. So DEC may not drive JAN.
4. Even if it were exactly same Covid, population is different.
As it turns out, actual fatalities (7 day trailing moving average) exceeds the forecast from the first differences specification.
Figure 1: Covid deaths, 7 day moving average (black), and prediction based on log-levels OLS regression 7 day moving average of deaths on lagged hospitalization and ICU (red), and based on first differences of logged values (green) (see text above), and actual 7 day moving average on 1/28 as reported in NYT 1/29/2022. Source: Our World in Data accessed 1/16/2022, NYT accessed 1/29/2022, author’s calculations. [Update of Figure 4 in this post]
The average 2529.2 deaths/day exceeds the point estimate of 2342 by 8%, but is well within the 95% prediction interval.