Epistemic Status: As with all my Covid-19 posts I’m not any kind of expert and am doubtless making a lot of mistakes. Yet it seems worthwhile to persist.
In previous posts I’ve been focused my direct modeling on New York. That was for various reasons. It’s the epicenter, it’s where I live, it is the place where herd immunity potentially matters already, and data is better than trying to reconcile 50 different states.
Now it’s time to start fully expanding that to the country. Wikipedia’s data is slightly different than my main source, and has the more generally fatal weakness that it doesn’t provide the negative test counts, but it provides it in easier-to-access form for deaths and infections, so I’m going to use it for the following charts I created (see calculations in my spreadsheet here).
Deaths by Week in the Big 5 Regions:
|Mar 19 – Mar 25||164||450||182||143||364|
|Mar 26 – Apr 1||424||1894||667||856||1988|
Positive tests by Week in the Big 5 Regions:
|Mar 19-Mar 25||7176||7812||9927||11923||33106|
|Mar 26 – Apr 1||16665||22121||28412||37339||55123|
Tests each week:
|USA Pos%||USA ex-NY Pos %||NY Pos %||USA Tests||USA ex-NY Tests||NY Tests|
|Mar 26-Apr 1||20.2%||15.4%||45.1%||728474||611073||117401|
For completeness, the total positive tests:
|USA Positives||USA Ex-NY Positives|
|Mar 26-Apr 1||56198||27769|
|Mar 19-Mar 25||146888||93987|
I decided to go by week because reporting on different days of the week is different. Tuesdays and Wednesdays have a bigger share, while weekends have lower shares. There’s also a lot of day-to-day noise in general. Number of tests fluctuates wildly. By combining weeks, we get a much easier to understand picture, and there’s now enough data that we can do this.
New York: We Can Make It There
The picture for New York is clear and excellent. The new R0 calculation comes in as R0 ~ 0.78, if we assume a five day serial interval. Using deaths and using infections each give us the same answer, so it seems very solid.
It does raise one important note about timing.
New York’s rate of infections detected peaked in the first week of April as did its positive test percentage. They then decline a little in the second week of April and fall dramatically in the third week.
Deaths peaked the second week of April then fall off dramatically. This dates the infection peak to about 21 days earlier, which would be on or right after the lock down date. That makes sense.
Infections peaking only a few days before deaths implies a very delayed testing schedule. Perhaps more so than is even plausible, but certainly more than I was previously taking into account. I think we have to presume that if deaths lag by 21 days, a positive diagnosis has to be at least 10 days delayed, under the conditions at the time. It should be somewhat less now.
That’s excellent news. What about other places?
Other Regions: Can We Make It Anywhere (Else)?
We see the positive test percentage falling around the country. That’s very good news. It does reflect increased testing, and may reflect increased focus of testing on places that need it the least. That worries me.
If look at the West region, we see a slow but steady increase in deaths. California is steady from this week to last week, but increasing from before that. It has a very low baseline, which is great, but things are not improving.
We also see a similar pattern of things slowly getting worse in the Midwest.
The non-NY Northeast region looks like it is finally getting somewhat better. But it’s happening later and slower than New York. The timing is weird, and there are some funky artifacts in the data contributing to that. Pennsylvania’s data is a real mess and New Jersey’s isn’t much better, and since we’re excluding New York that’s a huge percentage of the whole region.
The South was also starting to show slow improvement, but now it’s starting to reopen. As I’ve noted before, the ‘reopen’ criteria being used makes no sense. So it seems unlikely that the improvements will be sustained. There is extreme difference between states here. Virginia’s cases are suddenly going way up, whereas there seem to be clear previous peaks for some other states such as Louisiana. Louisiana, and New Orleans in particular, seems important to understanding the general picture.
The overall positive test percentage does continue to slowly decline. That is good news, but the decline is small, comes along with modestly increased testing, and likely it partly reflects changes in the distribution of who is doing the testing. I’ll learn more about that if/when I get to dive further into the states and cities one at a time and see if that’s true. It would be good to try normalizing for test locations, but I haven’t seen anyone do that.
Getting Worse Before It Gets Better: Louisiana
What’s going on with Louisiana?
Louisiana confirmed positive tests and deaths by week:
|Mar 26-Apr 1||1913||73|
|Mar 19-Mar 25||6845||227|
Louisiana’s data is a royal mess. I’m not listing positive test percentages, because they’re nonsense. From April 18 to April 25, the number of cumulative reported negative tests went down. There’s no later day where it looks like their reporting caught up. We realistically have no idea how much testing is taking place. Despite this they claim an overall positive rate of only 21.4%, and below 20% each day since April 11 (not counting days when they reported zero or less than zero negatives).
The drop-off in deaths we’ve seen so far seems nowhere near dramatic enough to be compatible with the drop in positive test results, unless testing collapsed, which the negative test counts claim didn’t happen.
Still, this seems like a clear explosion and fast peak, followed by a clear negative trend.
We can do a similar calculation to the one I previously did for New York. We take the deaths, backdate them three weeks, assume IFR of 1%, a serial interval of 5 and an R0 that matches the data, and see how many infections we get. In this case, we can use the decline in deaths on a weekly basis to estimate the R0. We get about 0.8. If we check that against the positive test counts from April 9 onward, we 0.775. Close enough. Let’s split the difference and say 0.79. Correcting for 50% under-count of deaths we get 328,344 infected out of a population of 4.65 million. That’s only 7.1% of the state infected.
A big difference is that only 391,000, or less than 10% of that amount, live in New Orleans, versus almost half of New York state being in New York City.
Redoing the calculation for only Orleans Perish, which has 27.4% of the deaths (found by Google giving me stats directly, and me dividing Orleans’ count by the total from all perishes) and only 8.4% of the state’s population, I get 22.5% infected in New Orleans. That’s lower than my estimate for NYC, but it’s more than halfway there.
If we refer back to the log of On R0, even the 7.1% infection rate is a big game. The difference between 7.1% for Louisiana, and much smaller numbers for most of the United States, explains how California can be in limbo around R0=1 while Louisiana has turned the corner.
Is there another explanation for why the epicenters seem to be improving, while places that are doing ‘everything right’ but were not epicenters are not seeing that? I don’t see a case available that Louisiana is taking things especially seriously, or implementing better or stricter policies than the West.
The explanation that makes sense to me is that lock downs in the American style, by default, create R0 very close to 1.
We see improvements in places with some herd immunity, or which are especially well locked down or otherwise situated. While others, where we’re falling short, get worse.
Which is all, in some sense, the worst possible situation.
You can’t squash. You’re forced into a holding pattern that devastates your economy without building the herd immunity that would let you come out of it. Eventually, you’re forced out, and you end up with the same wave you would have had earlier, except now you’ve used up your ability to lock down your population. You don’t have the capacity to squash.
However, once you have substantial amounts of infection, that pushes things over the top. The people infected are not random, they reduce risk more than proportionally, and thus are able to drag places with earlier higher infection rates into a stable place.
Thus, we need to do some combination of improving the effectiveness of our lock downs, and allowing people to become infected fast enough to matter.
Unless we think we can afford to sustain things until a vaccine or very effective treatment comes along. That doesn’t seem that likely at this point.
If we can’t come up with something better and can’t experiment, the logical response to this situation is variolation. We could infect the young and healthy on purpose, with low viral loads, to create enough immunity to turn the corner. Alas, we are highly unlikely to do this.
Thus, second best solutions prove necessary.
Which means that a partial reopening soon, as crazy as it seems and as misguided as the reasoning being used might be, starts to make some sense.
As does opening tattoo parlors and gyms in the first wave, as in Georgia. At first I thought, what the hell, that doesn’t make any sense. Lobbyists at their worst. That perhaps makes sense for gyms, but I doubt anyone is especially in the pocket of Big Tattoo.
Opening tattoo parlors, where social distancing seems obviously impossible, and which are obviously completely inessential, is a way of selecting for people most likely to go around getting themselves infected, and also most likely to be relatively young and healthy, and gives them somewhere to expose each other. Gyms can serve a similar function.
There’s also the fact that given what we know, a lot of the things being closed make things worse rather than better. Or at least, a full closing of them is worse than a partial closing.
If you open restaurants at 25% capacity, it’s not obvious at all that this increases risk. I think there’s a strong case that it decreases it, instead. Grocery stores are overloaded. By going to the restaurant instead, you reduce density at the grocery store, reducing wait times for others and the number of interactions while gathering items. You reduce strain on the supply chain, which reduces how often people need to make trips to get what they need. You also enable places to get enough business to stay open that would otherwise have to close, allowing them to offer takeout and delivery.
You might already be winning at that point. If you could also convince everyone to eat in silence, and order off of apps on their phones? I’m guessing you’re winning. If you can do outside seating, you’re definitely winning.
It really could go either way. It potentially helps a lot in increasing our ability to sustain other measures for a longer period.
Due to the speed premium, I’m getting what I have out there now rather than waiting for next steps. It’s possible all of this is a little selective, yesterday (4/29) was an aberration of some kind, and things are better than they look.
The next step would be to look at the data state by state, area by area, and try to make better sense of it. I worry I’m mostly talking into the wind with these analyses. I hope not. I do know I’ve made enough of a difference, in a few small places, to justify continuing to try.
If I know you, or I should know you, and you’d be interested in chatting about what’s going on and helping figure things out, and perhaps even helping get better decisions made, or just want to let me know I’m not shouting into the wind, give me a buzz however you typically reach me. I’m at the point where it starts to make sense for me to start writing code. If you know me at all, me writing code is a sign that division of labor has probably gone horribly wrong.