Yesterday Rhode Island reported its highest number of new positive tests for coronavirus so far. But is this a cause for worry that things are unexpectedly going sideways, or is it something else?
“Dr. Drang”, an engineer and data expert, has noticed a consistent bump in the national numbers every Tuesday for almost a month now:
These plots are for the daily figures, and you can see by the arrows in the top subplot that there’s been a distinct jump every Tuesday since late March. Why is this?
My strong suspicion is that this is due to the weekly work schedule of the people who report the figures the COVID Tracking Project compiles. Deaths over the weekend probably don’t get reported in full and the backlog of paperwork doesn’t get finished until Tuesday. And it may be due specifically to the weekly schedule of health officials in New York, as it’s New York’s numbers that dominate the country’s totals.
You can go see the plots at his site linked above. What’s striking to me is that the shape of the weekly curve segment is similar from week to week. To me that’s a sign that there’s a systematic counting artifact present. (Anyone who’s ever plotted experimental data knows what this is like.)
Like political polling or weather modeling, it’s important to look at the trend in the data rather than focusing too much on a signal data point.
The RI Department of Health site data is hard to read in this context because the state’s testing capacity has changed (for the better) again and again. The number of tests administered has been going up and so have the number of positive results. You’d really need to normalize the data, but to do that you’d really need random population testing and we don’t have that either – we’re testing the people either presenting symptoms or most at risk (congregate living residents and staff).
I think the take-away is keep an eye on the positive test numbers, but you’ll probably get a better sense of what’s going on by looking at hospitalizations or ventilator use numbers. Neither number is maxed out yet (thanks be to God) so it’s more likely to have fewer confounding variables mixed into the data. (If you’re looking for a peak, it’s probably going to show up more clearly in one of those reports than in the “positive case” dataset.)