Behind closed doors: Protective social behavior during the COVID-19 pandemic

The success of personal non-pharmaceutical interventions as a public health strategy requires a high level of compliance from individuals in private social settings. Strategies to increase compliance in these hard-to-reach settings depend upon a comprehensive understanding of the patterns and predictors of protective social behavior. Social cognitive models of protective behavior emphasize the contribution of individual-level factors while social-ecological models emphasize the contribution of environmental factors. This study draws on 28 waves of survey data from the Understanding Coronavirus in America survey to measure patterns of adherence to two protective social behaviors–private social-distancing behavior and private masking behavior–during the COVID-19 pandemic and to assess the role individual and environmental factors play in predicting adherence. Results show that patterns of adherence fall into three categories marked by high, moderate, and low levels of adherence, with just under half of respondents exhibiting a high level of adherence. Health beliefs emerge as the single strongest predictor of adherence. All other environmental and individual-level predictors have relatively poor predictive power or primarily indirect effects.

the submission is accepted. Please make sure it is accurate. The authors have declared that no competing interests exist. NO  The success of personal non-pharmaceutical interventions as a public health strategy requires a 50 high level of compliance from individuals in private social settings. Strategies to increase 51 compliance in these hard-to-reach settings depend upon a comprehensive understanding of the 52 patterns and predictors of protective social behavior. Social cognitive models of protective 53 behavior emphasize the contribution of individual-level factors while social-ecological models 54 emphasize the contribution of environmental factors. This study draws on 28 waves of survey 55 data from the Understanding Coronavirus in America survey to measure patterns of adherence 56 to two protective social behaviors -private social-distancing behavior and private masking 57 behavior -during the COVID-19 pandemic and to assess the role individual and environmental 58 factors play in predicting adherence. Results show that patterns of adherence fall into three 59 categories marked by high, moderate, and low levels of protective behavior, with just under half 60 of respondents exhibiting a high level of adherence. Health beliefs emerge as the single strongest 61 predictor of adherence. All other environmental and individual-level predictors have relatively 62 poor predictive power or primarily indirect effects. Introduction 93 Mitigating risk for future pandemics will depend upon the immediate and widespread 94 adoption of personal non-pharmaceutical interventions (NPIs) including social-distancing and 95 masking. The challenge of personal NPIs as a public health strategy is that they require 96  In practice, however, attitudes and beliefs tend to predict behavioral intention better 111 than they do behavior, and intention tends to explain only a portion of the variance observed in 112 health-related behavior. 16 activities listed, we use responses to the following four to measure adherence to private social-165 distancing behavior: (1) "Gone to a friend, neighbor, or relative's residence (that is not your 166 own)", (2) "Had visitors such as friends, neighbors or relatives are your residence," (3) 167 "Attended a gathering with more than 10 people, such as a reunion, wedding, funeral, birthday 168 party, or religious service," and (4) "Had close contact (within 6 feet) with people who do not 169 live with you." We combined the first two items into a single item: "household visits." We refer 170 to the third item as "social gatherings" and the fourth as "close contact." 171 A follow-up to the above was added to waves 7-28 to assess adherence to private 172 masking behavior. For each activity that received a participation response of "Yes," respondents 173 were asked "…how often, if ever, you wore a mask or face covering," with the following as 174 response options: always, most of the time, rarely, never, unsure. 175 We applied the following coding scheme to measure adherence to private social-176 distancing and masking behavior. In waves 1-6, our coding scheme was dichotomous and based 177 on social-distancing behavior. For each activity of interest (household visits, social gatherings, 7 close contact), respondents received a code of "1" (denoting adherence) if their participation 179 response was "No" or "Unsure" and a code of "0" (denoting non-adherence) if their 180 participation response was "Yes." 181 In waves 7-28, our coding scheme was ordinal and based on both social-distancing and 182 masking behavior. Respondents received a code of "2" (denoting high adherence) if their 183 participation response was "No" or "Unsure"; they received a code of "1" (denoting moderate 184 adherence) if their participation response was "Yes" AND their mask frequency response was 185 "Always" or "Most of the time"; they received a code of "0" (denoting non-adherence) if their 186 participation was "Yes" AND their mask frequency response was "Sometimes," "Rarely," 187 "Never," or "Unsure." adherence"), one consistently having the lowest ("low adherence"), and one consistently in 194 between ("moderate adherence"). 195

Individual-Level Predictors:
The conceptual framework for this study is presented in 196

Patterns of Protective Social Behavior 284
The results of our cluster analysis reveal three distinct patterns of adherence to 285 protective behavior in private settings: high adherence (44% of respondents), moderate 286 adherence (38%), and low adherence (18%) behavior. Mean scores for each activity and cluster 287 can be found in S2

Predictors of Protective Social Behavior 297
The results of our ordinal logistic regression models are presented in Figure 2

314
The statistically significant correlates of adherence within each domain are presented in 315 There are a few significant correlates that move in unexpected directions. For example, 323 there is a significant negative association between perceived risk of COVID infection and 324 adherence and a significant positive association between inability to afford a mask and 325 adherence. These associations may be due to demographic or other confounders not included 326 in the health beliefs domain. We account for confounders in our meta-regression analysis. 327 Meta-regression results can be found in Figure 3, which plots the percentage of each 328 domain's predictive value that is retained after controlling for all other domains. On the x-axis, 329 each domain is plotted in order of its C-statistic to denote the domain's predictive value prior to 330 controls. See S3 Appendix for odds ratios and standard errors from the regression model.

344
Results show that health beliefs retain 94% (p<0.001) of their predictive value, which 345 means only 6% is explained away by other domains. This percentage-retained is much higher 346 than the percentage retained by other individual and contextual domains. After health beliefs, 347 15 interpersonal context retains the highest percentage of its predictive value, at 87% (p<0.001), 348 although its predictive performance was quite poor, as indicated by its C-statistic of 0.59. 349 Community context, work context, and socio-demographic characteristics retain 52% (p<0.001), 350 34% (p<0.001), and 23% (p<0.001) of their predictive value after adjustment, respectively. Most 351 notably, political affiliation retains 6% (p=0.166) of its predictive value and informational trust 352 retains 0% (p=0.992), meaning essentially all of the predictive value of political affiliation and 353 informational trust is explained by their relationship to other domains, mainly health beliefs. Two key findings emerge from this analysis of protective behavior in private social settings. 361 First, patterns of adherence fall into three categories marked by high, moderate, or low levels 362 of social-distancing and masking behavior, with under half (44%) of respondents exhibiting 363 consistently high levels of adherence. Second, in support of prior research on the HBM, 24 health 364 beliefs are the single strongest predictor of adherence to protective social behavior. Alone, they 365 predict the largest share of observed adherence and they have the strongest direct, or 366 unmediated, relationship to adherence. 367 Compared to health beliefs, all other contextual and individual-level predictors have 368 either poor predictive power or poor primarily indirect associations. Political affiliation and 369 16 informational trust, for example, are moderately predictive of behavioral adherence but their 370 relationship to behavior is entirely explained by their relationship to other predictor domains, 371 mainly health beliefs. In line with prior research, 25,26 these results suggest political affiliation 372 and informational trust do influence protective behavior insofar as they influence beliefs. 373 A few qualifications to this analysis should be acknowledged. First, our results describe 374 protective behavior in private interpersonal settings. We focus on these settings because they 375 have been hotspots for the community spread of COVID-19 27,28 but the private nature of these 376 settings may explain why the effect of community context in our model is primarily mediated 377 by individual-level factors. In public settings, we might expect a different pattern than this to 378 unfold, where communities directly influence behavior without necessarily influencing beliefs. 379 Second, this analysis includes a limited set of measures of organizational and 380 interpersonal context. While we were able to account for the size of respondents' social 381 networks in this study, we could not account for network composition, although we expect this 382 to be correlated with community and socio-demographic characteristics. Furthermore, while 383 we were able to account for work conditions, we could not account for other organizational 384 contexts in which respondents are embedded. Third, our results reflect average patterns and 385 predictors of adherence to protective behavior during the COVID-19 pandemic, which may or 386 may not mirror patterns and predictors at a single point in time or in other public health 387 emergencies. 388