The effect of masks on cognitive performance

Significance

Does wearing a face mask reduce cognitive performance? In a study of international chess players who played millions of chess moves before and during the COVID-19 pandemic, masks substantially reduced the average quality of cognitive decisions. However, the negative effect is short-lived and primarily operates on elite players and in high-incentive contexts, with minimal impact on the broader population of players.

Abstract

The use of face masks has been a key response to the COVID-19 pandemic in almost every country. However, despite widespread use of masks in classrooms and offices around the world, almost nothing is known about their effects on cognitive performance. Using a natural experiment, I show that mandatory mask wearing has a negative causal effect on the cognitive performance of competitive chess players. I analyzed the quality of almost 3 million chess moves played by 8,531 individuals (ages 5–98 y) in 18 countries before and during the pandemic. Wearing a mask decreased the quality of players’ decisions—a measure of their cognitive performance—by approximately one-third of an SD. However, the disruptive effect of masks is relatively short-lived, gradually weakening such that there is no measurable disadvantage from wearing a mask after roughly 4 h of play. The mask effect is driven by a large, negative effect for experts, with minimal change in performance at lower levels, and is stronger in high-incentive competitions. I provide support for a distraction mechanism whereby masks interfere with performance when working memory load is high.

In 2022, 178 of 186 countries have required the wearing of facial coverings in some or all public spaces, with a further 7 countries recommending their use (1). Despite the ubiquity of mask mandates in work and educational settings around the world, almost nothing is known about their impact on cognitive performance. Given the established benefits of masks for containing the spread of COVID-19 (2–6), in combination with widespread social media claims about negative side effects (7), it is vital to understand whether and to what extent there are legitimate costs to mask policies.

The primary mechanism cited in social media messaging is the well-established association between cognitive performance and cerebral oxygenation (8–10). However, there is little evidence that masks reduce oxygenation or increase the risk of hypoxia for the average adult (11–13). As an exception to these findings, a recent study reported that wearing a mask during a 150-min university lesson slightly increased the heart rate and lowered blood oxygenation but did not affect reaction times or perceived mental fatigue (14). No large study to date has directly tested the cognitive effects of mask wearing on the general population, but two recent studies involving small samples are directly related to this paper. Wearing a mask for two class lessons did not significantly lower the cognitive performance of children (N = 133) (15), while wearing an N95 face respirator led to small or negligible differences in cognitive performance in a low-powered experiment with adults (N = 45) (16).

In this study, I used a large dataset from a natural experiment to quantify the magnitude and persistence of the effect of wearing a mask on cognitive performance. The setting is tournament chess, a cognitive domain in which individuals compete in person. Since March 2020, the wearing of face masks in chess tournaments has been mandatory in several countries at different times. I created a dataset of 8,531 individuals (ages 5–98 y) who played in 217 chess tournaments in Europe and North America before and during the pandemic. I used the best available computer software to evaluate the quality of a total of 2.9 million moves, a method that can detect the errors of even the world’s best players (17–19). Using a variety of specifications, including within-player estimation and controls for the local effects of the pandemic, I then compared the quality of decisions in games with and without mask-wearing requirements.

As a demanding task played competitively by millions of people around the world, chess has been used extensively by researchers in psychology, neuroscience, and economics as an instrument for measuring changes in cognitive performance (17–22). High performance in chess requires proficiency in calculation, memory, problem solving, and pattern recognition and is incentivized via both monetary prizes and rating points that quantify a player’s ability relative to their peers. The exceptional quality of longitudinal chess data affords the researcher the use of different methods for causal estimation and control over (unobserved) confounding factors. The conditions for a tournament chess game are standardized across countries. Numerical ratings capture a player’s underlying strength, and powerful chess computers objectively evaluate an individual’s chess moves to provide accurate within-game measures of the quality of decisions (18, 19). These features allow for individual performance to be reliably compared across games, players, and even time periods.

There are at least two pathways by which wearing a face mask might conceivably reduce cognitive performance. A physiological explanation is supported by studies that report a positive relationship between cerebral oxygenation and cognitive performance, although, as noted above, there is scant evidence to date linking wearing masks with lower oxygen saturation. If masks disrupt cognitive performance via decreased oxygenation, the cost in performance in a tournament chess game should increase over its duration. Alternatively, wearing a mask may be a distractor, reducing concentration and attention to the task at hand. Studies of cognitive control show that distractors interfere with cognitive performance when there is a high working memory load (23–26) as is often the case when playing chess (27–29). If masks decrease chess performance because of distraction, the effect should be strongest in the early stages of a chess game and then weaken over its duration as individuals adapt (30, 31).

I generated four measures of cognitive performance, calculated per player and per game: (a) the share of optimal moves, (b) the share of nearly optimal moves, (c) the average move quality, and (d) the share of large errors (see Materials and Methods for details). Measure (a), given by the proportion of a player’s moves in a game that were evaluated by computer software as optimal, is the outcome variable used in most chess studies to date. I confirmed compliance with a tournament’s mask policy through its official regulations, media coverage of events, and directly with tournament organizers and players. The primary sample included adults who played ‘classical’ chess games (typically 2–6 h in duration) and consisted of 6,505 individuals and 66,641 player-game observations (N = 13,368 with mandatory mask wearing and N = 53,273 with no mandatory mask wearing), of which most observations contain additional data about age and sex (see Materials and Methods for details on the data and procedure and SI Appendix, Table S1 for descriptive statistics of the sample).

Results

Overall Effect of Masks on Cognitive Performance.

As a baseline estimate of the effect of masks on performance, I first ran simple ordinary least squares (OLS) regressions of move quality on an indicator for mandatory mask wearing, controlling for chess strength. Wearing a mask has a significant negative effect on the share of optimal moves (β = ‒0.330, se = .126, p = .009, 95% CI = [‒0.578, ‒0.082], N = 66,081; all estimates are in SD). The negative effect of masks is robust to including controls for age, sex, and time fixed effects, and across these and all specifications in the study, the results are broadly consistent with the estimated effects for either the share of nearly optimal moves or the average move quality (Fig. 1). Because of potential bias from unobserved confounding effects at the individual level and over time, I next ran regressions with individual fixed effects and half-yearly time dummy variables. This preferred specification estimates the effect of masks on performance by using individuals who played games both with and without mask policies and for whom there are data on age (see ‘Main’ in Fig. 2). The estimates are similar to the baseline regressions (share of optimal moves: β = ‒0.348, se = 0.11, p = .002, 95% CI = [‒0.567, ‒0.129], N = 54,273).

Fig. 1.

Multiverse plot of estimates. The y-axis displays the standardized estimated effect of masks on one of the four outcome variables used to measure cognitive performance in a chess game. Gray bars indicate 95% CIs for each estimate. There is a consistent, negative effect of masks on performance as measured by the share of (nearly) optimal moves and overall move quality, while the effect on avoiding large errors is weaker and generally insignificant. All models control for (quadratic) strength and cluster standard errors by tournament. Specifications with Age include quadratic terms for player age. COVID controls include country per-capita new cases and deaths (‘spread’) or Google search trends (‘sentiment’) about COVID-19 recorded on the game day in that country (see Procedure); adding total or lagged indicators do not materially change the results. Time trends are controlled linearly, with half-year fixed effects, or by restricting the sample to tournaments during the pandemic period. Juniors adds players aged under 18 to the main sample. Main is the preferred fixed-effects model for estimating the effect on the share of optimal moves (see Table 1, column 2).

Fig. 2.

The effect of masks over the course of a game. Game duration is measured in 10-move bins, starting from move 10 and ending at move 60. Regressions control for individual and half-year time fixed effects, strength, and age, and standard errors are clustered at the tournament level. Plotted marginal effects are standardized onto a common y-axis that increases in performance for all variables. Displayed are 95% CIs. A typical tournament chess game completes the third bin after 4 h of play.

Because the mask treatment was determined by country-level mandates, it is possible that the main estimates for the mask effect also capture the influence of other factors related to COVID-19, such as fear or anxiety related to catching the virus, on performance. Using panel data about the country-level incidence of COVID-19, I repeated the main analyses with included controls for the local severity at the time of the game, such as including indicators for per-capita new and total cases and deaths, and their 1- and 2-wk lags, and using Google search trends as a proxy for COVID-19 sentiment (SI Appendix, Tables S2 and S3). The results were similar to the main specifications in every case, suggesting that the negative effect on performance is indeed due to wearing masks; for example, including per-capita new cases and deaths, coded as zero for prepandemic periods, produced an estimated mask effect on the share of optimal moves of β = ‒0.291 (se = 0.110, p = .009, 95% CI = [‒0.509, ‒0.073], N = 54,205; see also Procedure). Additional robustness checks that restrict the sample to games played since the beginning of the pandemic produce even stronger estimates for the mask effect, with or without local COVID-19 spread and sentiment controls.

Overall, there is strong evidence that the estimates reflect a negative, causal effect of masks on chess performance. Across the board, masks reduce the frequency of optimal moves by roughly a third of an SD. This is equivalent to a decrease in the share of optimal moves of 6 percentage points from a baseline average share of 29%, a relative drop in performance of roughly 21%. A similar standardized effect holds for the drop in performance for the share of nearly optimal moves and average move quality, while the effect on the frequency of making large errors is much smaller and statistically weaker (Table 1).

Table 1.

Main results by outcome measure

	Optimal moves		Nearly optimal moves		Average quality		Large errors
	OLS	Fixed effects	OLS	Fixed effects	OLS	Fixed effects	OLS	Fixed effects
Masks	−0.450** (0.145)	−0.348** (0.111)	−0.352** (0.121)	−0.265** (0.088)	−0.464** (0.154)	−0.355** (0.115)	0.103* (0.042)	0.077* (0.032)
Player strength	0.347*** (0.049)	−0.025 (0.088)	0.458*** (0.037)	−0.038 (0.072)	0.398*** (0.050)	−0.055 (0.091)	−0.465*** (0.011)	0.015 (0.049)
Player strength2	0.034 (0.018)	0.032 (0.040)	0.037** (0.014)	0.042 (0.036)	0.026 (0.019)	0.028 (0.041)	−0.019*** (0.005)	−0.008 (0.028)
Age	−0.005 (0.020)	−3.462 (1.791)	0.002 (0.016)	−2.809 (1.557)	−0.001 (0.021)	−3.673* (1.836)	0.006 (0.008)	0.632 (0.626)
Age2	0.031* (0.013)	0.013 (0.067)	0.027* (0.011)	−0.026 (0.056)	0.037** (0.013)	0.025 (0.069)	−0.015* (0.006)	0.105** (0.033)
Female	0.236* (0.104)		0.183* (0.083)		0.240* (0.109)		−0.081** (0.028)
constant	0.098 (0.192)	−0.465 (0.349)	0.052 (0.141)	−0.357 (0.297)	0.120 (0.201)	−0.479 (0.354)	0.030 (0.035)	−0.044 (0.125)
R 2	0.174	0.412	0.227	0.419	0.220	0.483	0.197	0.302
Num. Players	4,801	976	4,801	976	4,801	976	4,801	976
Observations	54,923	54,273	54,923	54,273	54,923	54,273	54,923	54,273

Notes: A negative effect of masks on performance is indicated by a negative coefficient on Masks in columns 1–6 and a positive coefficient in columns 7–8. Regressions include half-year dummy variables. Variables have been standardized. Standard errors (in parentheses) clustered by tournament. Num. Players is the number of individuals used to estimate the coefficient on Masks.

^* p < 0.05, ^** p < 0.01, ^*** p < 0.001.

Persistence of the Mask Effect.

Next, I examined the persistence of the effect over time. I separated the player-game evaluations of move quality into five sequential bins, each consisting of 10 moves, and interacted wearing a mask with each of the bin dummy variables to see how the effect changes over the course of a game. There is a clear and consistent pattern (Fig. 2). The negative effect of masks on optimal moves and average move quality is substantial in the early stages, on the order of 0.5 SDs, but the effect lessens considerably over the course of the game. By the end of the third bin of observations—typically reached after 4 h of play in tournament chess—the performances of a player with and without a mask are statistically indistinguishable (SI Appendix, Table S4). This pattern suggests that individuals adapt to the effect of masks on cognitive performance, which is not consistent with a physiological explanation. The mask effect also weakens in the second game of a two-game day, but there is no detectable weakening of the effect by day of play across the course of a multiday chess tournament (SI Appendix, Tables S5 and S6).

Cognitive Control and Working Memory.

An explanation consistent with the diminishing effect of masks over time is that masks are distractors that interfere with an individual’s working memory processes to which individuals can adapt. Working memory is an important component of chess expertise (20, 27–28, 29), and studies in cognitive control show that a person is more vulnerable to interference by distractors when working memory load is high (23–26). In two studies with small samples of chess players, blocking components of working memory via distracting secondary tasks decreased players’ performance on chess tasks (32, 33). If masks act as cognitive distractors, then the negative effect on performance should be stronger in contexts where working memory load is plausibly higher, such as in games with higher incentives. Because chess strength is a function of both skill and effort, and because working memory load is affected by individual and tournament characteristics, the distraction explanation makes three testable predictions in these data. First, a distraction explanation predicts that the mask effect increases with a player’s strength, under the assumption that stronger chess players have a higher working memory load during a game. This assumption is supported by a body of research on chess players that suggests that stronger players have a similar working memory capacity to weaker players but search through more chess possibilities when selecting a move (see Materials and Methods). Second, the mask effect should increase with the opponent’s strength, assuming that playing against a stronger opponent increases a player’s working memory demands. Third, conditional on player and opponent strength, the mask effect should be larger for tournaments in which there are higher incentives to effort.

To test the first prediction, I divided the sample into quintiles based on an individual’s chess rating. While there is no discernible difference in performance from wearing a mask for lower levels, the effect grows significantly with player strength, reaching a difference in the share of optimal moves of a full SD for the highest quintile of players (Fig. 3). This effect is consistent for the other outcome variables and when player strength is operationalized in other ways (SI Appendix, Table S9, column 1; SI Appendix, Tables S10–S13, columns 13-14). To test the second prediction, I interacted the standardized chess rating of a player’s opponent with Mask. The interaction effect supports the distraction explanation: a one-SD increase in the opponent’s rating increases the (negative) mask effect by 0.123 SDs (SI Appendix, Table S9, column 2). To test the third prediction, I created an indicator for whether the game was played in a high-incentive (‘elite’) tournament. In support of the assumption that a player’s working memory load increases in high-incentive tournaments, the share of optimal moves is substantially greater in elite tournaments, conditional on player and opponent strength (SI Appendix, Table S9, column 3). As predicted by a distraction/working-memory explanation, there is a significant interaction effect: Masks reduce performance by almost half an SD more in high-incentive tournaments than low-incentive tournaments (SI Appendix, Table S9, column 4). The interaction effect holds across subsamples of different player strengths, suggesting that nonexperts also experience a performance drop from masks in high working memory contexts. Taken together, the results are consistent with a distraction explanation whereby wearing a mask decreases cognitive performance when working memory load is high.

Fig. 3.

The mask effect and expertise. Expertise is measured in quintiles of chess rating. Regressions control for age, strength, sex, and half-year time fixed effects, and standard errors are clustered at the tournament level. Plotted marginal effects have been standardized onto a common y-axis that increases in performance for all variables. Displayed are 95% CIs. Cut-offs for each quintile are chess ratings of 2,093, 2,276, 2,410, and 2,540 rating points, with the highest quintile having a floor of roughly the top 500 in the world rankings.

Heterogeneity of the Mask Effect.

Finally, I investigated potential heterogeneous effects of masks along three other dimensions: age, sex, and time pressure. I repeated the main regressions with controls for strength, age, and time, adding interaction terms or splitting the sample where appropriate. Contrary to the main results, no significant mask effect was identified for juniors (age < 18) or seniors (age > 50, the seniors' threshold used in international chess). Similarly, there was no effect in fast (‘blitz’) games, in which the total game duration is much shorter. Replicating the main analysis with a sex interaction suggests a substantially stronger effect of masks for females (OLS with controls and half-year fixed effects: β = ‒0.354, se = 0.140, p = .012, 95% CI = [‒0.631, −0.078], N = 51,857). However, the sex effect is less robust than other results because it is potentially confounded by the large proportion of high-incentive games for females in the sample (90%, vs. 57% for males), and including an interaction between masks and incentives returns no significant sex differences in the mask effect (β = ‒0.221, se = 0.137, p = .109, 95% CI = [‒0.491, .050], N = 54,923).

Discussion

The results of this study present evidence that wearing a mask reduces cognitive performance in tournament chess. The plausibly exogenous treatment assignment, high statistical power, and the consistency of the effect across multiple specifications suggest that this conclusion is not a function of unobserved confounding variables, sample selection, or other potential challenges to a causal interpretation (Fig. 1). The magnitude of the decrease in performance is substantial, with an estimated effect size of roughly one-third of an SD for the share of (nearly) optimal moves and the average move quality in the preferred specification with fixed effects. By way of comparison, this effect of masks is roughly 10 times the estimated effect of a one-SD increase in daily air pollution on cognitive performance (34, 35).

Three other results shed light on the mechanism behind the mask effect. First, the effect weakens over the course of a chess game, suggesting an adaptive response. Second, the effect increases with a player’s chess strength and with the strength of a player’s opponent. Third, the effect is stronger in high-incentive tournaments. Taken together, these three results are consistent with a distraction explanation, whereby masks interfere with cognitive performance when working memory load is high.

In contrast to the strong effects on optimal moves and move quality, there is a smaller and statistically weaker effect on the share of large errors. While the statistical (in)significance can be explained by the low variability of this measure, the point estimates for the standardized effect are also consistently smaller than for other outcomes. It is not clear why the effect of masks might differ for making optimal moves compared to avoiding large errors. Some studies involving problem-solving tasks have found that reduced cognitive control causes individuals to use simpler, less cognitively demanding shortcuts (36, 37). In one experiment with chess players, search depth and breadth decreased when subjects were required to perform a demanding secondary task during move selection (33). If finding the best chess move requires complex search algorithms, whereas avoiding large errors is achieved through simpler heuristics^*, then it is possible that masks cause players to rely more on pattern recognition and heuristic reasoning relative to complex visualization of long sequences of moves.

With regard to the implications for mask policies in workplaces and educational settings, the results suggest that the effect of masks may depend on the type of cognitive task, the duration of the task, and working memory load. In particular, it is noteworthy that performance deficits fade over time, such that there is no detectable effects 4–6 h into a chess game and that masks do not decrease performance in low memory load contexts. These data indicate that the negligible and/or temporary cognitive performance reductions caused by masks are likely to be outweighed by the public health benefits in many contexts. In addition, and potentially relevant to policies related to masks in schools, I found no evidence for an effect of wearing masks for individuals under the age of 18 y, nor for those above 50 y. Combining these findings, the results suggest that masks are most likely to decrease performance in environments in which individuals face a cognitively demanding task with a high working memory load and in which optimal decision-making is crucial. Investigation of the precise mechanism for the effect and the extent to which masks affect different types of cognitive tasks would seem to be promising avenues for future research.

Materials and Methods

Data.

I used ChessBase Mega Database (2022 edition), the most popular commercial dataset of tournament chess games from around the world, consisting of approximately 9 million chess games. I collected move-level data from 217 international chess tournaments, of which 71 tournaments required players to wear face masks for the duration of every game. Internationally rated chess tournaments must follow strict, standardized rules and regulations regarding the tournament and game conditions as prescribed by the World Chess Federation (FIDE). The data search prioritized tournaments in seven countries that introduced mandatory mask restrictions for chess events in periods since March 2020: England, Greece, Italy, Portugal, Spain, Switzerland, and the United States. For balancing purposes and to reduce the risk of unobserved event-specific effects, the preregistered sampling strategy involved prioritizing National Championships, National leagues, and large ‘Open’ competitions that were run for multiple editions before and during the COVID-19 pandemic. In the second stage, tournaments from 11 countries that never introduced chess mask policies were selected on the basis of the extent of player cross-over between mask and nonmask events in order to maximize statistical power in the individual fixed-effects regressions. This process led to the inclusion of major chess events that were held in Belarus, the Czech Republic, Georgia, Germany, Iceland, Latvia, North Macedonia, Romania, Slovakia, Slovenia, and Turkey.

Confirmation of mask-wearing regulations and compliance at each event was conducted in three ways: (a) examination of tournament rules and regulations, (b) contact with tournament organizers, officials, and participating players, and (c) viewing photo and video footage. The latter was used to check both the degree of compliance at tournaments with mask regulations and the degree of voluntary mask wearing at tournaments without mask regulations. Over 50 h of video footage was used to ascertain that average mask wearing compliance was above 99% while average voluntary mask wearing was approximately 2% and always below 5%. Note that both noncompliance and voluntary mask wearing would bias an estimated mask effect toward zero.

The sample consists of 8,531 individuals who played a total of 45,272 games. I used the strongest commercially available chess engine, Stockfish 14.1^†, at a minimum search depth of 20 moves (‘plies’), to analyze the moves of every game and measure their quality relative to the best move in each position. In chess, players strive to make the objectively best move based on the (full) information contained in a position and, contrary to asymmetric information games like poker, there is typically no value to be extracted from an opponent’s facial expressions. Chess computer evaluations are measured in ‘centipawns’, which roughly equate to one-hundredths of the value of a pawn in chess.

Following the convention in the literature, I excluded the first nine moves for each player from the analysis, as the early stage of a chess game typically involves players executing preprepared, rote-learned moves before the main battle begins (18). I also excluded moves beyond the 60th move for each player, as the deep endgame stages of a game can be fully solved by engines, leading to extreme evaluations that are not comparable to typical centipawn evaluations. Of note, 89% of games in the sample did not go beyond move 60 (median = 40.0). I dropped games in which an individual played fewer than 10 evaluated moves. This left an average of 32.09 evaluations per player per game and a total of 2.90 million evaluations used in the analysis.

From the move-level dataset, I generated four outcome variables for each player-game observation: (a) share of optimal moves (defined as a move with an evaluation not worse than the evaluation of the computer’s suggested best move, measured in centipawns), (b) share of nearly optimal moves (defined as being within 50 centipawns of the best move), (c) logged average centipawn difference in evaluation between the played and best move, and (d) share of large errors (defined as decreasing the evaluation by more than 100 centipawns). The correlations between the outcome variables were: ρ_ab = .73, ρ_ac = .94, ρ_ad = ‒.46, ρ_bc = .84, ρ_bd =‒.78, ρ_cd = ‒.60.

The processed sample consists of 90,322 player-game observations (N = 70,189 with mandatory mask wearing and N = 20,133 with no mandatory mask wearing) from 8,531 individuals (ages 5–98, 106 nationalities) who played chess games in Europe and North America before and during the pandemic. Descriptive statistics and balancing are present in SI Appendix, Table S1. Without adjusting for clustered standard errors, there are significant differences between the mask and nonmask samples in terms of age, strength, and the share of females. The magnitudes of the differences for age and strength are within the bounds specified in the preanalysis plan for a balanced sample, and equivalence tests using the preregistered bounds of 1.0 y and 100 rating points support balancing on these factors (p = 0.003 for age and p < 0.001 for strength).^‡ These variables, along with sex and time, are controlled for in the OLS regressions. However, the sample differences and nonlinear time trends suggest the inclusion of individual and time fixed effects, which are the specifications reported in the even columns of Table 1. The preanalysis plan only proposed controlling for a linear time trend, but after collection of the data, the revealed nonlinear time trends and a strong seasonality of tournaments suggested including half-yearly calendar variables with cut-offs in March and September. The qualitative conclusions are similar regardless of the method for controlling for time trends (Fig. 1). Specifications with a linear time trend can be found in SI Appendix, Tables S10–S13, which list all 56 preregistered regressions. The preanalysis plan and other materials are available at https://osf.io/tkex7/.

Procedure.

I restricted the sample to players aged 18 y or above due to the unreliability of chess ratings as a measure of strength for youths and to games played at ‘classical’ speeds (typically 2–6 h), excluding ‘fast chess’ games (typically 10 min – 1 h). I first ran OLS regressions, progressively adding controls (quadratic forms of age and chess strength, sex, and time) and clustering standard errors by tournament, for each of the four dependent variables. Because players were not randomly assigned to the mask-wearing group, the simple linear estimates may be biased if there are unobserved confounding effects at the individual level or over time. For example, the types of tournaments that require masks may differ in composition, the types of players who choose to play mandatory-mask tournaments may be correlated with performance, or there may be unobserved factors related to the pandemic that influence the quality of play, such as changes to training regimes, or a fear of getting infected in a group setting. I repeated the analysis with individual fixed effects to eliminate these potential sources of bias; games from 1,070 players in the sample who played both with and without masks were used to estimate the within-individual mask effect, and age data are available for 976 of these individuals. I added time fixed effects or removed prepandemic observations to account for changes that can be attributed to the pandemic (per se) and to compositional changes of players who choose whether to compete during the pandemic. The results from the OLS and fixed-effects regressions are largely consistent (Table 1 and Fig. 1).

For a final set of robustness checks, I repeated the OLS and fixed-effects regressions with additional controls for the local extent of COVID-19 on the day of each chess game. The decision to impose mask policies in chess tournaments may have been motivated by factors that also affected player performance. For example, a sudden increase in infections or deaths in a country might increase the likelihood of mask policies being introduced in tournaments and simultaneously decrease player focus and motivation. Because the local incidence of COVID-19 may simultaneously affect treatment assignment and performance, I reran the analysis after including the daily country per-capita numbers of new COVID-19 cases and new deaths as control variables, coding prepandemic levels as zeroes (SI Appendix, Table S2). Additional specifications restricted the sample to pandemic periods and included total cases, total deaths, or 1- and 2-wk lags for each COVID-19 measure. As an alternative method, I used Google search trends as a measure of COVID-19 sentiment, which may be a better proxy for confounding factors of fear and anxiety related to the virus. I included control variables for the search volumes of the terms “coronavirus” and “coronavirus deaths” in the week preceding a chess game, relative to the country’s maximum search volumes for these terms over the pandemic period (SI Appendix, Table S3). In all estimations, adding COVID-19 spread or sentiment controls did not materially change the estimated coefficient on Mask. In addition, neither COVID-19 spread nor sentiment was predictive of move quality in nonmask games. Taken together, these robustness analyses support a causal claim that masks reduce cognitive performance.

Next, I investigated whether the effect of masks on cognitive performance changed over the course of a game. SI Appendix, Table S4 presents the results of the within-game persistence analysis, whereby the outcome variables are derived for each successive bin of 10 moves during a game. The interaction coefficient with mask progressively increases with each bin, suggesting a weakening of the mask effect (see also Fig. 2). This effect is robust to selection on game length: the even columns in SI Appendix, Table S4 restrict the sample to long games with observations for all bins.

To rule out the alternative hypothesis that the interaction effect increases because it becomes easier to play optimal moves as a chess game progresses (due to the reduction in the number of pieces, and therefore the number of possible moves, on the board), I tested whether the effect of masks differs between games over the course of a day. While tournament conditions are standardized by the World Chess Federation, there is a small degree of flexibility for the tournament organizers to schedule two ‘classical’ chess games in a single day. This requires a shorter maximum duration of 4 h per game, which explains why the move quality is slightly lower for this sample. SI Appendix, Table S5 looks at the interaction of the mask effect with whether the observation was derived from the second game in a day. The strong, positive coefficient suggests that players continue to adapt to wearing a mask over the course of a day as well as over the course of a game. Interacting the effect of masks with a variable for the round (the sequential number of the game in a tournament, which typically lasts 1–2 wk) showed no significant effect in the fixed effects regressions and a small, negative effect in the OLS regressions (SI Appendix, Table S6). However, these estimates may be confounded because a weakening of the mask effect due to within-tournament adaptation is likely to be moderated by strengthening of the mask effect due to higher working memory loads in the later, higher-incentive tournament rounds. Taken together, this pattern of results is potentially more consistent with the explanation that masks work as cognitive distractors that an individual adapts to over time than with a physiological explanation.

As an alternative means of controlling for any other effects of comparing the pandemic to a nonpandemic period, I restricted the sample to tournaments played since March 1, 2020. The masks and nonmasks groups in this sample are significantly more balanced in terms of age and strength, though not in terms of the share of females (SI Appendix, Table S7). As before, I included fixed effects and quadratic controls in age and strength as well as sex in the OLS regressions. The estimates of the mask effect are robust and are also substantially larger, with average effect sizes on the share of optimal moves of roughly one-half of an SD (SI Appendix, Table S8).

I next tested three predictions related to the distraction/working memory explanation of the mask effect. The first prediction is that the mask effect should be larger for stronger players, under the assumption that working memory load is positively correlated with chess strength. While a long history of research using chess experts has shown that working memory is an important component of chess expertise (27–28, 29), more detailed models of memory in chess have been the subject of a rigorous and still unsettled debate (see refs. (38) and (39) for reviews), and so this assumption is not self-evident. Chess experts might have a higher baseline working memory capacity, perhaps because of higher general cognitive abilities (20, 22), meaning that they can perform the same task under a lower working memory load. Alternatively, stronger players might search to the same depth as nonexperts but use their experience or intuition to select better moves, therefore experiencing a similar working memory load.^§ In contrast to these possibilities, the weight of evidence from studies of chess players suggests that they do not have stronger baseline working memory capacities, but they do engage in greater search depth. While experts are significantly better at recalling chess game positions than nonexperts, there is typically a much smaller or no association found between chess strength and recall of randomly populated chess positions (29, 38, 40–45), and chess expertise does not predict superior performance in other nonchess backward induction tasks (46). Yet, despite similar baseline working memory capacities, stronger chess players search deeper and are more likely to consider an opponent’s possible replies during the process of move selection than weaker players, both of which increase working memory load (39, 41, 47–49). These studies are consistent with an incentives argument—stronger chess players, especially professionals, face higher financial and reputational incentives to expend cognitive effort—and a small literature using functional neuroimaging on chess players that suggests that chess strength is related to enhanced activity in brain areas associated with working memory activation when performing chess-related tasks (50–(51), see also refs. (52) and (53)).

Taken together, the body of evidence on chess strength and search processes supports the assumption of a higher working memory load for stronger players. Based on this assumption, I tested the first prediction—that the mask effect should be larger for stronger players—in three ways. First, I investigated the effect on chess Grandmasters^**, who make up 21% of the observations in the sample and whose level typically indicates that the individual is a chess professional, and for those at or above Master level (a rating of 2,000), who make up 79% of the observations (SI Appendix, Tables S10–S13, columns 13–14). The negative effect of masks on performance is significantly larger in the high-level samples in all eight regressions, though the effect is again weakest for the share of large errors. Second, I divided the sample into quintiles based on chess rating and ran linear regressions with controls for age, strength, sex, and half-year periods (Fig. 3). The cut-off for the highest quintile was a chess rating of 2,542, roughly the level of the top 500 chess players in the world. The effect of masks grows with the quintiles, reaching approximately one SD in magnitude in for the highest rating quintile. Finally, I operationalized strength by including a variable for the interaction between masks and a player’s standardized chess rating (share of optimal moves: SI Appendix, Table S9, column 1; results are similar for the other outcome variables). In all cases, the estimates support the hypothesis that the (negative) effect of masks on performance is larger for stronger players.

The second prediction is that, conditional on a player’s strength, the mask effect should be larger when playing against a stronger opponent. In chess, it is widely believed that stronger opponents create situations in which it is (more) difficult for the player to play good moves. As above, I operationalized opponent strength by including a variable for the interaction between masks and the opponent’s standardized chess rating. The results support the hypothesis that the (negative) effect of masks on performance is larger when playing against a stronger opponent (SI Appendix, Table S9, column 2).

The third prediction is that the mask effect should be larger in high-incentive games, conditional on player and opponent strength. I split the sample based on whether the tournament had high incentives to effort. High-incentive tournaments included national championships, continental championships, and the Grand Swiss Tournament, which forms part of the World Championship qualification cycle. The high-incentive sample comprised 64% of the total player-game observations. The remaining tournaments included national club competitions and nonelite ‘Open’ events. In the sample without masks, conditional on player and opponent strength, the share of optimal moves is substantially greater in high-incentive tournaments (SI Appendix, Table S9, column 3; note that the within-subject estimation controls for self-selection into tournaments). This finding suggests that a player’s working memory load increases in high-incentive tournaments. Next, I interacted the dummy variables for mandatory masks and high-incentive tournaments. As predicted by a distraction/working-memory explanation, there is a significant interaction effect: masks reduce performance by 0.483 SDs more in high-incentive than low-incentive tournaments (SI Appendix, Table S9, column 4). Moreover, although there is no detectable effect of masks for nonexperts overall, the negative interaction effect of masks and high incentives is also significant and substantial for nonexperts (0.390 SDs), suggesting that this group also experiences a performance drop from masks in high working memory contexts (SI Appendix, Table S9, columns 5–6).

The methods for testing other heterogenous effects can be summarized as follows:

Age: Age data or classification of junior or senior exists for 85.8% of the observations in the sample. I ran separate regressions for juniors (age < 18) and seniors (age > 50) and also interacted the quadratic Age terms with the mask effect (SI Appendix, Tables S10–S13, columns 8–10). There was no negative effect of masks on juniors or seniors.

Sex: I interacted Female with the mask effect (SI Appendix, Tables S10–S13, column 11). The negative cognitive effect of masks was substantially larger for females than for males. As discussed in the main text, this effect is potentially confounded by the large proportion of high-incentive games for females.

Time pressure: Roughly 3.4% of the sample (3,084 player-game observations) consists of fast-chess games, in which the maximum duration of the game is between 10 min (‘blitz’) and 1 h (‘rapid’). I interacted Time pressure with the mask effect (SI Appendix, Tables S10–S13, column 12). There was no interaction effect.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Code and reproduction instructions have been deposited in [OSF] (https://osf.io/tkex7).