In brief: this is the largest data-driven study trying to disentangle the effect of individual countermeasures, and one of the most thoroughly validated ones. The authors ran multiple predictions with heldout data and various sensitivity analyses. The results seem robust. It's still unclear how generalisable they are, since it's essentially an observational study. (I managed the project who wrote the paper, but I wasn't involved as a co-author.)

The effectiveness and perceived burden of nonpharmaceutical interventions against COVID-19 transmission: a modelling study with 41 countries

Background: Existing analyses of nonpharmaceutical interventions (NPIs) against COVID19 transmission have concentrated on the joint effectiveness of large-scale NPIs. With increasing data, we can move beyond estimating joint effects towards disentangling individual effects. In addition to effectiveness, policy decisions ought to account for the burden placed by different NPIs on the population. 

Methods: To our knowledge, this is the largest data-driven study of NPI effectiveness to date. We collected chronological data on 9 NPIs in 41 countries between January and April 2020, using extensive fact-checking to ensure high data quality. We infer NPI effectiveness with a novel semi-mechanistic Bayesian hierarchical model, modelling both confirmed cases and deaths to increase the signal from which NPI effects can be inferred. Finally, we study how much perceived burden different NPIs impose on the population with an online survey of preferences using the MaxDiff method. 

Results: Eight NPIs have a >95% posterior probability of being effective: closing schools (mean reduction in R: 50%; 95% credible interval: 39%-59%), closing nonessential businesses (34%; 16%-49%), closing high-risk businesses (26%; 8%-42%), and limiting gatherings to 10 people or less (28%; 8%-45%), to 100 people or less (17%; -3%-35%), to 1000 people or less (16%; -2%-31%), issuing stay-at-home orders (14%; -2%-29%), and testing patients with respiratory symptoms (13%; -1%-26%). As validation is crucial for NPI models, we performed 15 sensitivity analyses and evaluated predictions on unseen data, finding strong support for our results. We combine the effectiveness and preference results to estimate effectiveness-to-burden ratios. 

Conclusions: Our results suggest a surprisingly large role for schools in COVID-19 transmission, a contribution to the ongoing debate about the relevance of asymptomatic carriers in disease spreading. We identify additional interventions with good effectiveness-burden tradeoffs, namely symptomatic testing, closing high-risk businesses, and limiting gathering size. Closing most nonessential businesses and issuing stay-at-home orders impose a high burden while having a limited additional effect.

The team who produced this work is also hiring for a new project manager.


New Comment
5 comments, sorted by Click to highlight new comments since: Today at 11:41 AM
closing schools (mean reduction in R: 50%)

I'd like to know how large part of this is universities, high schools, elementary schools, and kindergartens. In other words, how dangerous it would be to open kindergartens but keep universities closed, etc.

We have no info on that, sorry. That's because we have a single feature which is switched on when most schools are closed. Universities were closed 75% of the time when that happened IIRC.

Now there is a study from Germany, comparing different cities within Germany, that arrives to completely different results. They conclude that wearing face masks reduces the number of daily new infections by 40%.

So I guess I am lacking the necessary knowledge to understand " >95% posterior probability of being effective", and I have to ask: What are "credible intervals"? With a frequentist-econometrics background, my reaction is: What? Zero is in the intervals?

Think of it like one-sided vs two-sided. You can have a 95% CI that overlaps with zero, like [-2, 30], because 2.5% of the probability mass is on >30 and 2.5% on <-2, but still the probability of >0 effect can be >95%. This can also happen with Frequentist CIs.

A credible interval is the Bayesian analog to a confidence interval.