AdobeStock_351973475_Editorial_Use_Only.jpeg
VoxEU Column COVID-19

Mask usage reduces COVID-19 deaths: A US county-level analysis with a unique identification approach

The science behind mask usage and its ability to reduce airborne particles seems clear. Despite this, many individuals are sceptical that wearing masks can reduce the spread of COVID-19 and many refuse to wear one even when required. This column examines the effect of mask usage using county-level data from the US, employing an instrumental variable approach. The findings show that increasing the amount of individuals who frequently or always wear a mask when within six feet of people by 1% could reduce COVID-19 deaths by 10.5%, which translates into approximately six deaths in the average county. 

Several scientific studies have shown that mask usage reduces respiratory droplets, which is now believed to be the main transmission mechanism of  COVID-19 (Bahl et al. 2020, Lindsley et al. 2020, Verman et al. 2020). However, despite this evidence, individuals are sceptical that wearing a mask will affect the number of deaths related to COVID-19, and many refuse to wear one even when businesses or municipalities require it. In addition, many police departments seem reluctant to enforce mask mandates, often due to their own personal scepticism. 

Controlled laboratory studies that examine how masks work to reduce respiratory droplets and other airborne particles are useful, but they are unable to fully determine the effectiveness of mask usage in the real world. A few studies have now examined the effect of mask usage or mask mandates on COVID-19 cases or deaths in practice (Chernozhukov et al. 2020, Karaivanov et al. 2020, Mitze et al. 2020, Yilmazkuday 2020, Zhang et al. 2020). These for the most part have found that an increase in mask usage will reduce COVID-19 cases or deaths.

The complication of examining mask usage in an observational study is that frequency and amount of mask-wearing likely correlates with other factors related to COVID-19 deaths. If not accounted for, this will either make mask usage look more or less effective than it truly is. Areas that are more (‘naturally’) prone to COVID-19 deaths – such as, but not limited to, areas with an older population or where COVID-19 has previously spread – are also likely to have increased mask usage because of this increased threat. Techniques that don't account for this will underestimate the effect of mask usage. On the other hand, areas where the population has individuals who are, on average, more risk tolerant are likely to have less mask usage, but also at the same time the people there will be more likely to engage in other behaviours that would put them at increased risk for COVID-19. Empirical techniques that don't account for this will overestimate the effectiveness of mask usage. Thus, using methods that don't properly identify the model will have biased results, where even the direction of this bias is unclear. 

In a recent paper (Welsch 2020), I contribute to the growing literature on the real-world effectiveness of masks by examining county-level data in the US and, importantly, employing a unique identification approach to address the above complications. For my outcome measure, I use county-level COVID-19 deaths reported by CDC and for mask usage I utilise a survey conducted in July by Dynata and the New York Times. While in many cases surveys can have misleading elements, in this instance it has some advantages over some of the alternatives. Some of the studies above examine mask mandates. However, mandating mask usage does not mean that mask usage will occur. In addition, mandates are often enacted in conjunction with other measures aimed at reducing COVID-19, so it is difficult to disentangle the effect of mask usage from these other measures.

In my study, I start by examining a simple correlation between mask usage and COVID-19 deaths. This correlation shows the counterintuitive result of a positive and significant relationship between increased mask usage and COVID-19 deaths, highlighting the dangers of not accounting for the first complication mentioned above. Once I partially account for an area's natural predilection for COVID-19 by including deaths from COVID-19 prior to the survey and the population of the area as controls, the size of this counterintuitive effect is reduced. When I (partially) account for both the risk tolerance and health/fragility of the population by including all deaths in a county from 2016, this effect shrinks further in magnitude and loses statistical significance.

Next, I include a full set of controls, including but not limited to age of the population, education levels, average income, racial makeup, and population density. When included, the results indicate that increasing mask usage will reduce COVID-19 deaths, but this effect size is small. Specifically, a 1% increase in individuals who say they often or always wear masks is associated with a 0.4% reduction in COVID-19 deaths. However, it is unlikely even a full set of controls can fully account for the concerns mentioned above. 

If the controls do not fully account for the ‘natural proclivity’ for COVID-19 deaths in an area or there are remaining unobservables that correlated with both mask usage and COVID-19 deaths, the results will remain biased. To fully identify the model and account for these concerns, I employ an instrumental variable technique, where I instrument mask usage with percentage of votes for Donald Trump in 2016. 

My instrument is not weak and arguably passes the exclusion restriction and monotonicity assumption. The instrument passes all tests of strength. The first stage F-stat of the excluded instrument is 49.73 and the t-stat on the mask usage coefficient in the second stage is 6.24, which according to Lee et al. (2020) makes the coefficient significant at least at the 5% level. It seems probable that the monotonicity assumption is satisfied as well. It seems unlikely that if a county were to have increased (decreased) vote for Donald Trump that it would increase (decrease) mask usage. 

Finally, the exclusion restriction is likely satisfied. Recall the two most troubling unobservables in the analysis are the natural susceptibility of the population of the area to dying from COVID-19 and the risk tolerance of the population. It seems unlikely that areas with different levels of vote for Donald Trump are likely to have different levels of natural susceptibility to COVID-19 deaths or different levels of risk aversion, especially considering the controls that are included. In addition, I examine both of these possibilities, by informally using the historic level of fragility/risk aversion of the population of the counties – I examine the relationship between the percentage of vote for Donald Trump and the historic overall death rate, and death rate due to injuries. The results indicate very little evidence of an association between fragility and percentage of vote for Donald Trump and if anything, areas with a larger percentage of vote for Donald Trump are less risk tolerant. 

The main instrumental variable results find that if the amount of individuals that indicate that they often or always wear a mask when in public within six feet of others increases by 1% that it would reduce COVID-19 deaths by 10.5%. This equates to approximately six deaths in an averaged sized county. These results are robust to several different specifications. 

An additional concern with the instrument is that percentage of vote for Donald Trump might be negatively correlated with other COVID-19 prevention measures that also reduce deaths, such as social distancing, amenability to quarantining, or isolation, and so on. I find that a common measure of social distancing is a weak indicator of both percentage of vote for Donald Trump and COVID-19 deaths, and including it in the model does not qualitatively affect the results. However, even if this issue persists, it likely only leads to some attenuation bias, not a nullification of the results.

References

Bahl, P, S Bhattacharjee, C de Silva, A A Chughtai, C Doolan, and C R MacIntyre (2020), “Face coverings and mask to minimise droplet dispersion and aerosolisation: a video case study”, Thorax, 75 (11), 1024-1025.

Chernozhukov, V, H Kasaha, and P Schrimpf (2020), “Mask mandates and other lockdown policies reduce the spread of COVID-19 in the US”, VoxEU.org, 15 July.

Karaivanov, A, S E Lu, and H Shigeoka (2020), “Face mask mandates slowed the spread of COVID-19 in Canada”, VoxEU.org, 9 October. 

Lee, D S, J McCrary, M J Moreira, and J Porter (2020), “Valid t-ratio Inference for IV”, paper, arxiv.org.

Lindsley, W G, F M Blachere, B F Law, D H Beezhold, and J D Noti (2020), “Efficacy of face masks, neck gaiters and face shields for reducing the expulsion of simulated cough-generated aerosols”, article, medRxiv.org.

Mitze, G, R Kosfeld, J Rode, and K Wälde (2020), “Unmasked! The effect of face masks on the spread of COVID-19”, VoxEU.org, 22 June.

Verman, S, M Dhanak, and J Frankenfield (2020), “Visualizing the effectiveness of face masks in obstructing respiratory jets June 2020”, Physics of Fluids, 32 (6), 061708. 

Welsch D M (2020), “Do masks reduce COVID-19 deaths? A county-level analysis using IV”, Covid Economics 57: 20-45.

Yilmazkuday, H (2020), “Fighting Against COVID-19 Requires Wearing a Face Mask by Not Some but All”, Available at SSRN 3686283.

Zhang, J, J Li, T Wang, S Tian, J Lou, X Kang, H Lian, S Niu, W Zhang, B Jiang and Y Chen (2020), “Transmission of SARS-CoV-2 on Aircraft” Available at SSRN 3586695.

2,946 Reads