Face masks don’t work. Not!
Latest update: 30 July 2021; see the last two paragraphs and the tweets that follow.
A recent randomized controlled trial by Bundgaard et al in the Annals of Internal Medicine has captured considerable attention by the likes of Sebastian Rushworth who summarized the article in a tweet as saying, “Face masks don’t work.” As I pointed out within minutes, “That’s really simplifying the results section in questionable ways.”
Let me be blunt, Rushworth’s tweet is deeply misleading science communication by someone who should know better had he read the piece before he tweeted. Sure enough his tweet was retweeted more than 100 times as of the writing of the present blog entry and it was liked many times more contributing to the silly debate we currently have in some countries about the efficacy of masks.
A day later the not-so-good doctor doubled up with a somewhat more differentiated piece on his blog where he back-peddled a bit but not much.
Rushworth is the guy who in early August told us on that very same blog that “covid is over in Sweden. People have gone back to their normal lives and barely anyone is getting infected any more.” Really? I am on record of reserving my judgement about the success of the Swedish strategy and am actually convinced, based on current evidence, that it will emerge as one of the better ones although I am also convinced that we do not look where we really should look for good strategies: Taiwan, Hongkong, Singapore, Japan, South Korea, Vietnam, and maybe Iceland. Surely the present surge in cases (and to lesser extent in deaths) in Sweden contradicts what Rushworth suggested happened and would happen. I am not going to hold my breath for a correction coming from him any time soon. He seems, I have to say, not a reliable narrator.
Back to the research study by Bundgaard et al and what it really says. Spoiler: it does not say what the not-so-good doctor claims it says.
The study was conducted in Denmark 3 April — 2 June 2020, i.e., it spanned the period of re-opening on 18 May 2020. The study involved about 6,000 adults (not yet afflicted by covid-19/SARS-CoV-2 which took about a month to establish) that were randomly and about equally assigned to the two treatment arms. The baseline condition (“the control group”) were people that followed the standard social distancing measures outside their home than in place in Denmark. The treatment (“the mask group”) were people that followed the standard social distancing measures and that in addition agreed to wear a mask outside their home.
Less than 5,000 participants completed the study. Among those in the baseline condition 53 (2.1%) contracted the virus and among those in the treatment 42 (1.8%) contracted it, leading the authors to conclude that the difference is not statistically significant. This seems to be the key results that Rushworth managed to read before he tweeted away. (More on the issue of statistical significance below.)
The problem is, this study happened in a very specific environment “where others are not wearing masks”. And, “the findings should not be used to conclude that a recommendation for everyone to wear in the community would not be effective in reducing SARS-Cov-2 infections, because the trial did not test the role of masks in source control of SARS-CoV-2 infection. During the study period, authorities did not recommend face mask use outside hospital settings and mask use was overwhelmingly to persons not wearing masks.”, as Bundgaard as his colleagues were careful to point out.
So, to claim that face masks do not work is … really irresponsible coming from someone like the not-so-good doctor with a few thousand followers.
This is not the place for a long and comprehensive discussion of the evidence on the overall effects of masks. Only this much: By now it is well established that the virus is transmitted typically person-to-person, primarily through mouth, nose, or eyes via respiratory droplets, aerosols, or fomites (surfaces), with the latter being a very-small probability event as we know at this point. I summarized the evidence on transmission on 25 April 2020 here. Since then Erin Bromage has written another excellent piece on the issue that deservedly was widely read.
A key insight is that the probability, and the severity, of infection is a function of the intensity of dosage times the length of exposure to the dosage. Masks, even simple ones, impede dramatically, and very intuitively, the ability of droplets and aerosols to reach other people. This has been known since early in the game (see this video from mid-April), or look at the successful prevention strategies of Asian countries such as Taiwan, Hongkong, Singapore, Japan, or South Korea, or for that matter Czechia. (It is notable that this posterchild of successful anti-SARS-CoV-2 strategies has now fallen behind Sweden in the covid-19 related deaths-per-million-of-population sweep-stakes after it let it guard down and got overwhelmed by a massive second wave.
That masks are effective has been shown convincingly in many other settings. A clever experiment with hamsters reported in mid-May, for example, demonstrated the basic intuition. As did the involuntary experiment involving two symptomatic hair stylists (wearing masks) and their 140 customers.
There should be no question that wearing a mask in places that are crowded, closed (with poor ventilation), and where social distancing is not possible (close contact situations) is a good and responsible thing to do. Hence the recommendation to wear a mask on (crowded) public transportation is quite sensible, while the requirement to wear one in (not-so-crowded) public parks seems silly and counterproductive.
Before I conclude, briefly back to the issue of statistical significance in what is now becoming known as the DANMASK-19 study. For starters it is noteworthy that the confidence interval suggests that a positive outcome (reduction in infection) was twice that of a negative outcome. Remember though that the environment where the study took place was characterized by a) a very low infection rate (about two percent for the one month it was effectively running and b) very low usage of masks in the population. Had the number of incidences scaled up linearly across the two treatments conditions, statistical significance might have been reached. (For other questions about this DANMASK-19 study, and it turns out there are many pertaining to design and implementation, see the expert opinions in the Science Media Centre.)
Again, the key thing is that this DANMASK-19 study establishes, for a very idiosyncratic environment, the effect of masks for the wearer when the real issue is how masks prevent the transmission from symptomatic and non-symptomatic carriers of the virus to those around them.
So boys and girls, if the bus gets crowded, put on your damn mask, not necessarily for you but for others. Face masks do work!
Update: NN Taleb on the same study a couple of days later: https://fooledbyrandomnessdotcom.wordpress.com/2020/11/25/hypothesis-testing-in-the-presence-of-false-positives-the-flaws-in-the-danish-mask-study/
Another update: This article was published in PNAS last year already but remains of relevance. Exploiting one of the latest identification fads in empirical research — synthetic controls — the authors take advantage of regional variation in Germany of the point in time when wearing a face mask became mandatory in public transport and shops. Based on data from 401 municipal districts, the authors conclude that face masks reduce the daily growth rate of reported infections by almost 50%. Synthetic controls have been for a number of pandemic outcome studies. Their very nature is backward looking which makes them problematic to assess fully intertemporal trade-offs. That said, for specific and narrow questions like the one discussed above the approach seems a useful tool in the research shed.
Yet another update: Just out in PLOS ONE, an article by a bunch of people from Melbourne and a(notoriously alarmist) colleague from Sydney. They study whether “the introduction of a mandatory mask policy was associated with significantly reduced COVID-19 cases in a major metropolitan city” and answer the question in the affirmative. The metropolitan city is Melbourne and the outbreak is the yellow one below.
The outbreak was driven by an earlier, less infectious variant of the virus (not the delta variant currently afflicting parts of Sydney) that it took Victoria more than three months to get under control. The face coverings order was imposed 20 July 2020 and started to be implemented 22 July 2020, with existing Stage 3 restrictions remaining in place. (Stage 4 restrictions were implemented August 5.) When the face coverings order was announced, the number of daily new cases had increased for a couple of weeks exponentially and was already half of what turned out to be the peak (about a week later). There is an obvious confound here in that the authors do not control for precautionary behavior of people for which there is plenty of evidence. They try to control for that through mobility data but I don’t think they succeed. (There may have been be an indirect effect in that the order re-inforced people’s precautionary sensibilities.) The NSW Health authorities have consistently stressed that the chance of catching the virus in outdoor settings is minimal and I cannot think of even one well-documented case here.
So, masks indoors in crowded and poorly ventilated spaces such as public transportation where social distancing is difficult — by all means. Even in elevators etc. Outside, in contrast, masks are unlikely to make much of a difference.
Just came across this:
Consider following me on twitter: @aortmannphd