Version 3 of the Minnesota Model was released a few days ago. The model immediately drew skeptical reactions for its prediction that Minnesota would have roughly 1,440 deaths by the end of May:
[S]ome say the numbers just aren’t plausible… To hit the estimated deaths by the end of the month would require, on average, between 41 and 44 deaths per day for the rest of the month. Minnesota has averaged 22 deaths per day in the last week. (Source)
I grabbed the model source code from GitHub and played around with it. Here is how Version 3 of the Minnesota Model projects the next couple weeks will play out:
| Date | Projected Deaths | Actual Deaths (Daily updates in green) |
| Before 13 May | 557 | 614 |
| 13 May | 29.0 | 24 |
| 14 May | 30.3 | 25 |
| 15 May | 31.7 | 20 |
| 16 May | 33.1 | 17 |
| 17 May | 34.6 | 22 |
| 18 May | 36.1 | 9 |
| 19 May | 37.7 | 17 |
| 20 May | 39.3 | 29 |
| 21 May | 41.1 | 32 |
| 22 May | 43 | 33 |
| 23 May | 45.2 | 10 |
| 24 May | 47.6 | 17 |
| 25 May | 50.5 | 12 |
| 26 May | 53.7 | 18 |
| 27 May | 57.4 | 33 |
| 28 May | 61.4 | 35 |
| 29 May | 65.9 | 29 |
| 30 May | 70.7 | 30 |
| 31 May | 75.9 | 14 |
| TOTAL | 1441.1 | 1040 |
Feel free to grab the model yourself and play around with it. (Technical notes: my numbers are from a reconstructed Scenario 5, with parms$end_time_social_distancing <- 1 + 5 + 51 + 21. [UPDATE: repo here, branch “Scenarios5and6.”])
The uncertainty on this seems to be in the neighborhood of +/- 300, based on Slide 12 here. That means anything from 1100 to 1700 deaths should be considered a success for this model.
Back in March, I wrote a very simplistic projection of how quickly covid would spread through the state if we did nothing. I did this to build a case for immediate cancellation of all large gatherings and the imposition of a temporary lockdown.
Today, I’m not writing to advocate anything in particular. These deaths can no longer be stopped. The people who will die of covid-19 over the next couple weeks have, mostly, already been infected, and will die regardless.
…Well, okay, not quite. According to the model, ending the stay-at-home order on May 18th (instead of continuing it into June) does mean an additional 53 deaths, with the first “extra” death most likely occurring on May 24th. But 53 extra deaths on a model with a margin of error of +/- 300 is invisible statistical noise. Besides, even if Gov. Walz reversed course today by extending the stay-at-home order, expectations of re-opening are now so high that it seems unlikely we could actually get back to a strong enough stay-at-home order to save those lives. (I’ll write about Minnesota’s declaration of surrender as soon as I can.)
So why make this post at all?
After I did the March post, I found it useful to have a daily “mile marker” around to check whether covid was beating my expectations or falling short. I checked back regularly to help confirm whether I was overreacting or underreacting. (In retrospect, I appear to have been just-right-reacting.) By the same token, having these “mile markers” for May will help tell us how much stock we should place in the Minnesota model.
If we end up with a death toll significantly above these numbers (above 1700), then it means the model is missing something. Perhaps it underestimated the lethality of the virus, or perhaps it overestimated the extent to which people will continue to socially distance after Sunday’s official end to stay-at-home orders.
On the other hand, if we end up significantly below this death toll (under 1100), then it means the model is missing something in the other direction. Perhaps the new treatments we’ve started receiving are really effective, or perhaps summer weather is having a substantial effect on slowing the virus’s spread. (The Minnesota model does not include weather in any way.) Or, more cynically, perhaps it means people are dying at home and not getting discovered. I’ve become a lot more pessimistic about these things since Minnesota gave up on test/trace/isolate and Virginia got caught cooking its books.
To be frank, I’m optimistic that the reality will indeed end up on the low end of this model. Looking at places like Florida and Singapore, I can’t help thinking that there may be something to this idea that the summer offers a respite from covid — a respite we can use to get on top of this thing after all. It doesn’t hurt my mood that, while I’ve been working out the details of the model, the weather has warmed up and the daily deaths in our state have started going down for the first time in weeks.
But I’ll be watching. And now you can watch with me.
See this blog’s previous analysis of the covid epidemic in Minnesota here.
EDIT: Now that I’m exploring the released version of the Minnesota model for myself, let me know in the comments if there’s anything specific you’d like me to take a look at, and I’ll see whether I can make the time during the next few days.
Minnesota released Scenarios 1, 2, 3, and 4 (described here). I was able to exactly reconstruct Scenarios 5 and 6. I think I can build out the rules to reconstruct Scenario 7. Scenarios 5a, 6b, and 8 are beyond my capabilities right now, because the Minnesota modeling team did not include sufficient (any?) documentation about them.