Last week my analysis of the Covid-19 epidemic in Italy yielded a very depressing result: despite all the measures taken by the government, there was no sign that they were making a difference. The epidemic was still growing exponentially with dire implications for the overall number of deaths it could cause.
My approach to tracking Covid-19 dynamics in various countries is explained in a technical publication and, less technically, in my previous blog post, How Effective Are Public Health Measures in Stopping Covid-19?
Rerunning the analysis on the more recent data (up to March 31) results in a much more optimistic conclusion. There are now clear signs that Italy has turned the corner. Here are the results:
The most visible sign of change is the decline of New Cases. Less visibly, other curves also have started to bend down. The change in dynamics is driven primarily by declining transmission rate of the disease:
At the beginning of the epidemic (in early February) the transmission rate, beta, was almost 0.4, which means that the number of infected grew at the rate of nearly 40% per day (but this explosive rate of growth was invisible to the public or the authorities, see below). The decline in beta was very slow and late (the inflection point of the curve is at March 13). This suggests that it took a while for the exhortation of the Italian government to change the behavior of its citizens. We know this also anecdotally. As late as on March 21, the Chinese Red Cross vice president, Sun Shuopeng, was reported saying “Here in Milan … public transportation is still working and people are still moving around, you’re still having dinners and parties in the hotels and you’re not wearing masks. I don’t know what everyone is thinking.” More amusingly (and we need some humor in these trying times), the latest run-away hit on YouTube has been videos of small-town mayors in Italy raging at people flouting the lock-down. But clearly the gravity of the situation — over 13,000 deaths as of today, has finally impressed itself on the Italians. The transmission rate has been declining — slowly, but is heading in the right direction.
Note the second graph (on the right). It shows quantitatively what we all know: the initial detection rate (the probability that an infected person would become known) started low and then increased. What this means is that we cannot simply use the reported numbers to accurately trace the dynamics of the epidemic. As the detection rate increases (first due to the general awareness that we are in an epidemic, and later as a result of massive testing of asymptomatic people) it would inflate the number of new cases, artificially increasing the transmission rate. To see what is actually happening we need to factor out this change in the detection rate, which is what my model does.
I’ll continue reporting my analysis results as more data come in. I was actually, hoping to run the data from New York, but yesterday they were not available as the team at Johns Hopkins (many thanks to them for putting together and updating the data) changed the way they report USA data.
Great news! I am curious, according to your model, what percentage of Italy has been exposed by this point?
I should have said that I ask because a recent report suggested that Iceland’s estimate of its own population was around 6%. Since most countries other than Iceland have not done any random sampling to really know (or at least haven’t shared it with us), I’m curious if your model would be able to estimate this.
The problem is that while I can estimate a change in the detection rate, I can’t estimate the overall level — only a relative change. It’s because the data are not informative on this aspect of the epidemic. I need a direct measure of the ratio of detected/actual (including asymptomatic) cases for at least one point in time.
This is not a big problem for my main question, estimating the effect of interventions on the transmission rate. But it also means I can’t answer your question.
I live in Italy: the turn Is probably three days ago (there Is 3-5 days from test to data) also death numbers are probably understimated: in Lombardia region sistem was near to collapse so a lot of old people death at home without test and are not counted. estimate are x2 at best to x4 in the worst, and i suspect the same for Madrid and NY and so on
Hi Paolo, actually you need to add the delay of 3-5 days, rather than subtract it. So the turning point was probably 3-5 days before the estimated inflection point (March 13). A more important aspect is how gradually the government interventions took effect. This is very different from China, where the turning points were much sharper (you can see the results for Hubei in my posted technical document).
Hi Peter,
yes add my mistake tanks.. about death rate in Italy: it’ s now sure, confirmed from ISTAT (national statistics agency) that dead are at least double: in the first 3 weeks of march 2020 we have on 1084 municipality (7904 total) 16.216 dead, the average in the 5 years 2015-2019 was around 8.000. at 21 of march know coronavirus dead for overall italy (all 7904 municipality) was 4825.
yes agree that are more important the effect of different government interventions. i read older coronavirus post and technical document tomorrow: i work a lot more with lockdown than before….
This is great progress, Peter. I like how you have aimed at trying to solve the initial transmission-detection issue. How soon could these models be used for policy-decisions? I see this transmission-detection issue being a massive problem as Africa joins the pandemic in the coming months…
The plan is to estimate the effect of interventions in each country and them do a statistical analyses relating this effect to the nature of various interventions (and how strictly they were implemented). So we are far from making definitive recommendations yet.
You should check this site, https://healthweather.us, based upon a cellphone app, that measures temperatures. Like your model, it shows a distinct drop in the US starting 3/20. Some of that drop is from reduced spread of flu. Fewer people are ill now that in a typical year.
Quite interesting. And it makes complete sense that the flu and common cold are down — thanks to the severe social distancing measures.
Roger: looking at the US as a whole doesn’t make much sense because there are huge differences in policies and public attitudes. Generally, blue states might be approaching the peak while red states will keep growing exponentially for weeks to come.
I wonder if all that’s happening will lead to extinctions of some flu strains and some common cold viruses; all of them have lower R0 than the new coronavirus. Would work better if all countries went into lockdowns at the same time, of course.