There are more and more laudable efforts aimed at assessing statistically the evolution of Covid-19 epidemics in various countries and regions. For instance, Yinon Weiss analyzed the data from US states and surprisingly found that there is no obvious relationship between the speed with which states adopted lockdowns and the mortality outcomes.
Here, I would like to do something similar with testing and mortality. The mortality performance of a number of high-testing countries, such as Iceland, Austria, South Korea, Norway, Taiwan, Germany suggests that testing may be a key factor explaining Covid-19 mortality. However, there may be significant differences in data reporting among countries and the epidemics there may vary greatly in terms of their starting date, potential initial size and so on.
Regional comparisons in countries that provide relatively reliable statistics may be an alternative subject of study. Here, I’m looking into the data from Italian regions and the US states, in particular the relationship between the number of tests per confirmed case and the number of deaths per million.
An amazing picture emerges if this analysis is made for Italian regions (see below).
Testing intensity seems to explain 57% of the mortality differences among regions. One could of course object that this may be confounded by two major factors. First, regions that test more may be ones that were hit later than the first regions, thus they may have had time to learn the lesson. However, this is not the case for Italy. One of the highest-testing regions is Veneto and it was the Italian region with the first reported Covid-19 death on February 21.
Secondly, in some regions (Lombardy, Emilia-Romagna, Veneto, etc.) epidemics started early, while others may have had most of their outbreaks under lockdown conditions. But this probably does not affect the conclusions on testing since there are a number of Italian regions (Aosta Valley, Piedmont, Marche, Liguria) that had their first deaths relatively late but have high mortality rates. In fact, the Aosta Valley region had its first death on March 11, yet its mortality rate is higher than in Lombardy.
Interestingly, if one more variable is added to the analysis, namely, total hospitalizations after 2 weeks from the first reported case, testing (x1)still retains its effect and adding hospitalizations (x2) improves the fit, with the regression explaining more than two thirds of variation.
What could total hospitalizations after two weeks signify? They may be related to the first-mover disadvantage but they may also reflect the potential hospital superspread. Interestingly, part of Veneto’s clear success may be explained not just by testing but also the decision to not hospitalize the majority of confirmed cases.
I suspect that if weather and pollution are taken into account, the bulk of variation will be explained, and even the outlier of Aosta Valley will cease to be so much of an outlier.
If we perform the same exercise for the US states and the District of Columbia, we get a similar relationship but with a significantly lower R-squared (0.21), even though it remains statistically significant.
One needs to remember, however, that testing in the US had been very heavily curtailed by the failed CDC test and FDA’s resistance to allowing other labs to test. As late as March 10, the whole of the US had done only 5,177 tests. For testing to matter isolation and contact tracking and testing efforts should also accompany testing. And that seems to have been essentially absent at least in New York:
A week later, on March 1, she tested positive for the coronavirus, the first confirmed case in New York City of an outbreak that had already devastated China and parts of Europe. The next day, Gov. Andrew M. Cuomo, appearing with Mayor Bill de Blasio at a news conference, promised that health investigators would track down every person on the woman’s flight. But no one did.
In contrast, even in Lombardy, there had been initially a significant test-and-trace effort that was tragically curtailed on February 25 by the Italian government afraid that it could create too much fear. It is that effort that allowed a group of Lombardian researchers to reconstruct the early outbreak in the region in a groundbreaking paper that should have been noticed long ago.
It is clear that the low R-squared is driven by the Northeastern states with high mortality, especially New York (557 per million) and New Jersey (314) that are the biggest outliers but also Connecticut (188) and Michigan (176). The only other state with mortality higher than 150 per million is Louisiana (218).
This seems to suggest that the climate, especially powerful superspreader events (like, potentially, Mardi Gras in Louisiana) and perhaps air pollution, added to testing would go a long way towards explaining the differences in mortality among US states. I call upon statisticians and econometricians who are better at it than me to do just that.
In sum, testing intensity appears to be an important factor in reducing Covid-19 mortality both in Italy and the US. The explanatory power is weaker for the US, probably because of the delays in testing, lack of contact tracing and high heterogeneity of conditions.