In spite of the title, the basic message of this diary is actually one of good news. It appears that almost everywhere, social distancing is working very well, and all signs are that the coronavirus is in rapid retreat all over the country. With one notable exception, that I will get to in a bit. First, though, I need to discuss how we can actually know this.
One of the most frustrating and difficult things about the coronavirus pandemic is the long lag time between the disease spreading in society and that spread becoming visible in the numbers that are publicly available at from your local Department of health, and, more conveniently, in places like the Johns Hopkins Coronavirus map or the Worldometer coronavirus page. Not only is the incubation time an average of five days, but after people get sick, they usually take a week or two before they seek help, and even if they are able to get tested, there is further delay in getting the test, the results coming back, and those results becoming part of official statistics.
All in all, this means that the data that gets reported in the news and on the web tell you about what happened in the community about two weeks prior. This is why it is so easy to act too late to stop an explosion of disease like what we are seeing right now in New York, and why it’s too early to tell from test results what the effect of social distancing and lockdowns really is. Basically, after you implement a policy like a stay-in-place order, you are left holding your breath for two or more weeks before you can really tell what it did. That is one thing that makes the epidemic very difficult to manage for public health officials, and very hard mentally for the rest of us.
The fact that testing capacity is right now increasing so rapidly makes the data even harder to interpret, since the raw numbers won’t tell what part of the increase is due to more people falling ill, and what part comes from a larger fraction of sick people actually being diagnosed. One way to tell the difference is to look at hospitalizations and deaths, since those should not depend that strongly on how many people get tested, but unfortunately, that means you have to deal with an additional one or two week lag in the data.
So is there anything we can do to get information more rapidly? Yes, there is. If you have been watching the news lately, you may be aware of Kinsa, a company that sells smart thermometers you can use to monitor your body temperature the same way a Fitbit allow you to keep track of have far you walk each day. They have sold over one million of these thermometers in the US, and this has produced a huge database of when and where people have fevers all across the country. The important point here is that this data shows you what is happening right now. Not two weeks ago. In past years, this real time view of fevers have allowed Kinsa to spot flu outbreaks up to two weeks before CDC was able to see them. This year, the company was able to spot rapid growth of additional fevers in some places that could not be attributed to simple flu. That’s because right now, the only two things that can produce fever in large numbers of people are the flu and the corona virus, so you can then be reasonably certain that if there are a lot of fevers beyond what could be expected from the flu (Kinsa calls this surplus “atypical” illness) is due to the COVID-19 outbreak. South Florida was one area where they were able to pinpoint a large outbreak long before hospitals there saw any increase.
This was all in the news a couple of weeks ago, so I think it’s worth revisiting some Kinsa’s data to see where things stand. As a first example, let’s take a look at the fever curve for Huntsville, AL and surrounds:
This is just an illustrative example. Kinsa has these traces for every county in the US, and if you want to find out what is going on in your area, you can go to healthweather.us to take a look.
There are several interesting things going on in this plot. At first, in February, the curve stays within the expected bound for the seasonal flu (in light blue). Then, around the beginning of March, the amount of fever starts to rise rapidly above the range we would expect for the flu. As we now know, this was when the coronavirus was starting to spread all across the country, and this is immediately visible in the Kinsa data. This increase is seen all over the country, and its size generally correlates well with the number of the COVID diagnoses made 2-4 weeks later. This is also the time when people started paying real attention to what was going on, and the first action most states took to stop the spread was to close the schools. The wisdom of starting with school closures is debatable, but that’s what happened. In the case of Alabama, schools were closed March 18th, so the 19th was the first regular school day without school in the state, indicated by a read arrow in the figure. The effect of this is an almost immediate plunge in fevers, and within ten days fevers fall by about the amount that can be attributed to the flu. At that point, the precipitous fall abruptly changes into a slow decrease over time.
If you think about it, this is pretty much what you would expect, because the seasonal flu is a very different disease from COVID-19. The flu first of all has a more rapid progression. The average incubation time is about two days, compared to five days for COVID. And fever during flu usually last only a few days, while in COVID it tends to last at least a week and often much longer. Even more importantly, adults have significant immunity to the flu, which means that transmission is almost entirely sustained by children, who mostly lack that immunity. As a result, if you close the schools and separate the children from each other, the flu will drop out of circulation almost immediately. As for COVID, there is some debate as to how much kids actually contribute to the epidemic, but it is definitely the case that children are not the driving force. So closing the schools might slow the epidemic somewhat, but will not come close to ending it.
Therefore it is reasonable to assume that the fevers that persist after the end of the rapid plunge following schools closing is almost entirely due to coronavirus infections. (There are of course reasons people might have a fever that is not related to an infection, but that is rare enough that it will have little effect on the data.) This means we can still use Kinsa’s data to monitor the pandemic outbreak in real time, only we will no longer have to subtract the now extinct seasonal flu from the numbers.
To see how this plays out across the country, here are the fever traces from six more US localities:
The top two traces are for Brooklyn, NY and Miami, FL, where the coronavirus as we all know is running rampant. Sure enough, these place both have high levels of fever, even though we are more than ten days past schools closing. (This is indicated by the red arrows. There are two arrows for Miami, since the city closed its schools four days before the rest of the state, and the data for each county actually includes data from the surrounding area as well.) As you would expect, fevers are particularly high for Brooklyn, but the good thing is that it is declining rapidly, which should mean that we can expect a rapid slowdown in the number of hospitalizations in New York City over the next two-three weeks. That does not mean we will see a dramatic improvement in the next week, unfortunately, but it is clear that things are looking up.
The second row of the graph shows fevers in Seattle and Worcester, MA with smaller but still quite substantial COVID outbreak (3668 cases in Seattle (King County), and 1296 cases in Worcester). Accordingly, remaining fevers are lower than in NYC and Miami, but still have a ways to go before getting to zero. Seattle appears to be making particularly good progress, which is a testament to the importance of test and trace in getting the pandemic under control. Washington was of course the first state to have community spread of coronavirus in the US, but they had comparatively good local testing capacity from the start and they have doggedly kept up tracing down contacts of ill people and putting them in quarantine. This is without a doubt why Washington has not seen the explosion of cases New York has experienced, even though the governor didn’t issue a stay at home order until pretty late (March 25).
The last fever traces in the graph are for Riley County, Kansas (home of Kansas State University) and Blue Earth County, MN (which I picked because of its awesome name). These are both rural locations with comparably few COVID cases. In both places, fevers essentially ended a few days ago, something that currently holds for much of rural United States. This does not mean there was never any coronavirus in these locations—Riley County has seen 19 cases, and Blue Earth County 21. It also does not mean coronavirus has been eliminated from circulation. It is very likely there are a few people still sick with COVID-19 around, it’s just that none of them owns a Kinsa thermometer, or they have one but don’t use it. But we can safely assume that we will not see a large increase in COVID in these places over the next few weeks.
So it seems that the social distancing policies work. Is that true everywhere? To help answer that question, Kinsa conveniently provides a map that show what direction fevers have been headed over the past week. This is the map at the top of the story, but since that is kind of far to scroll, here it is again (blue is decreasing fevers, yellow and blue increasing).
The reason I’m using a map from Apr 5 is because a couple of days ago, Kinsa changed the way they plot this data. This is what they say on their website:
On April 7, we changed the calculation behind the “trend” map to capture the slope of the line of best fit (linear regression) of the previous 7 days of illness data. This map was previously driven by the mean percent change of the previous seven days— a measure that proved volatile when illness levels approached zero in some counties. We believe the new method better captures what’s actually happening with illness, and paints a more coherent picture in general.
They do have a point here, but unfortunately their new map uses too few colors for the trends we are interested in to show up. Until they fix that, this is the best we have.
In this map, most of the US reassuringly blue, indicating that the residual fever that is probably due to COVID is decreasing in most places. There are a number of exceptions, however. Most of these are due to the volatility Kinsa’s statement talk about, and therefore not concerning. For instance, the Intermountain West has a lot of red on the map, but this is because the fever readings in this area has been essentially zero for the past week. This means that if random fluctuations in the data, which are inevitable, happen to produce a small spike in over the past few days, this appears as a large relative increase in the amount of fever in that area. So this means there is no resurgence of COVID in the west, quite the opposite. The same is true in most other red patches on the map, including the red streak across Iowa, Minnesota, and Wisconsin and in several other places. I’m less sure about Northeast Arkansas, but that is likely also not significant.
That leaves us with three non-blue areas on the map that are of greater concern. The two I want to discuss first are Southwest Georgia and the greater Washington DC area, possibly with an extension to the northwest into central Pennsylvania. These are both areas that are currently seeing substantial COVID outbreaks. Southwest Georgia in particular is one of the hardest hit areas of the country. Here are two representative plots for both regions, from Terrell County in Georgia and Fairfax County in Virginia:
In Georgia, we can see that the fevers are actually decreasing over time, although perhaps not as fast as in much of the rest of the country. In Virginia, it appears that fevers are holding steady, neither increasing nor decreasing over time. This is a bit disconcerting, but perhaps not too surprising, given that the DC region has only been under lockdown for a week, which is shorter than most other outbreak areas. So it may be that we just have to wait a bit longer for the corona-associated fever to begin to subside in the DC area in the way it already is elsewhere. On the other hand, there are some indications the social distancing in the DC area may not be as large as one might wish for. In any case, things are at least not getting worse in this region, so I don’t think there is any cause for immediate alarm.
That cannot be said for the last non-blue area on the map, which is centered on western Kentucky, with spillover into adjacent areas of Indiana and Tennessee. In this region, fever measured by Kinsa has actually increased over the past week, as can be seen in the plot from McLean County, KY
This area has until now seen a particularly large number of COVID cases, so the increase is not the tail of much higher earlier infection rate that was hidden by the flu. Instead, this is most likely a genuine increase caused by insufficient social distancing. The New York Times published an article yesterday about how a second wave of COVID-19 may be starting to hit rural America, which is particularly bad since these are places with a much weaker health infrastructure and a population in worse health than in the urban areas that have been hit so far. Kinsa’s data so far indicates that this second wave will probably be pretty manageable in most places, but western Kentucky might prove to different.
As for why this is, we can only speculate. This is coal country, and therefore also Trump country, so people there are probably more apt to believe that the coronavirus is not that big of a deal, except perhaps in liberal dens of sin like NYC. This attitude is of course not unique to western KY, so there must be some other factor. We know that the coronavirus tends to spread through events where a lot of people congregate in close quarters, ranging from funerals and weddings to Spring Break beach parties to Mardi Gras. For example, the large outbreak in SW Georgia I mentioned above can be traced back to a single funeral where apparently one of the guests was at peak infectiousness, producing dozens and then hundreds of cases within a few weeks. Perhaps something like that happened here sometime in early-mid March.