The title image shows the USA and Florida new COVID cases curves on top of each other. I chose the scale to illustrate that for a long period of time Florida was contributing 20% of new cases nationally. That has dropped this month. Now we contribute merely about 15%.
The national curve is very close to an ideal logistic distribution, and that function arises as the solution to a basic differential equation often used to model population growth. The shape of the curve indicates that we should be seeing a drop in cases very soon.
This has been suggested by my calculations for at least since the beginning of the month — I looked over my models from 8/1 and 8/4 yesterday and was very surprised. The 8/1 model had an average error of 2.6% for 17 days. My goal had been to get 7-10 of predictive ability…
Obviously this model would be looking worse had I included yesterday. The 8/4 model still looks like I might be the best one I came up with:
Now, an interesting aspect to my work was that I was modelling CA, TX, FL, NY, and the remainder of the USA using a simple logistic curve model, and combining them. Until I saw how well these models did I was dubious if it was worth the effort. It seems that it is.
Not all is good — CA, TX, and FL might not quite as poised to peak. CA looked good for a few days before a bit of a jagged edge showed up yesterday.
TX has presented challenges — mainly that their data seeps in. For the 3 or 4 th week in a row I thought that maybe they were showing a drop in cases until laggards reported. The last model I calculated is below:
I’m a bit more confident that we are nearing a peak here in Florida — we’ve spent 7 days between 21-22K new cases a day (7 day moving average). This curve, like Texas, is not a great fit for the logistic curve, so we have that uncertainty.
New York has been a far easier state to model — and I’ve been using the 8/4 model in subsequent efforts. So we now have 14 days of data against this component model...but it appears that the model might have reached it’s end of usefulness…
The largest component of my national model is “The Rest of the USA”. This is the national number less my 4 sentinel states. This data is sensitive to missing states, so it took me a few tries before I felt I was getting a decent model.
A more close up view shows why I am hopeful that this has peaked:
So, I have been combining these 5 logistic models to get a national model, with decent results. When I look at the country as a whole I can get a simple logic cure that is very close to this composite model. So what is the benefit?
Well, I think the benefit is that I spend a fair amount of time actually working up a decent estimate for the equation parameter for “total population” — in this case the total number of people expected to be infected by this surge. By more carefully estimating this for 5 separate parts of the whole I was able to compare it to the estimates I was using before. And this step I think is what is giving me better accuracy for a longer period of time now. Without this extra effort I’m not sure I would be getting as good a simple model for the country:
Now, let’s reflect on last year. Cases started to rise mid September and then started to blow up in October. This surge started with national cases averaging roughly 40K a day.
We’ve already had Sturgis. A lot of other events with crowds. Labor Day. Schools are starting with the ridiculous anti-masking bullshit here and in Texas. If we are lucky we don’t see a rise in cases until mid September again, and a month of dropping at 4K cases per day per day might put us back in the 40K range where we were last fall.
But if cases don’t drop that fast, or for that long we could be starting a fall/winter surge at a worse place than last year. Hopefully, by then, more will be vaccinated and willing to change their behavior to slow the spread of this disease.
Stay safe!
ADDENDUM — ~5:30 PM EDT
I had to run out, so I took a look a todays numbers on Worldometers so far. 92K with 15 states including CA and FL not in. Assuming yesterdays numbers for FL/CA and 1K per state left other than these gets us up to about 142K for today. “Replacement” for the 7 day average is the number from 8/11 — 148,500. So my rough estimate seems to indicate that we’ll see a 7 day moving average tomorrow that is only a little more or less than today. More evidence that we are nearing a peak/plateau.