key: cord-0709010-ilopa76j authors: Bundorf, M. K.; DeMatteis, J.; Miller, G.; Polyakova, M.; Streeter, J.; Wivagg, J. title: Risk Perceptions and Private Protective Behaviors: Evidence from COVID-19 Pandemic date: 2022-03-10 journal: nan DOI: 10.1101/2022.03.08.22272111 sha: 45b37a24b119703d8ac92a440cba3907733e510b doc_id: 709010 cord_uid: ilopa76j We analyze data from a survey we administered during the COVID-19 pandemic to investigate the relationship between people's subjective beliefs about risks and their private protective behaviors. On average, people substantially overestimate the absolute level of risk associated with economic activity, but have correct signals about their relative risk. Subjective risk beliefs are predictive of changes in economic activities independently of government policies. Government mandates restricting economic behavior, in turn, attenuate the relationship between subjective risk beliefs and protective behaviors. Predicting responses to such policies requires understanding how policies influence perceived risks and, in turn, how risk beliefs influence behaviors (Manski 2004). We examine the relationship between individual perceptions of (a very salient) risk and economic behaviors, as well as how much people believe that policy interventions change this relationship. Our analysis is based on a nationally-representative survey that we fielded in the United States in May 2020, near the beginning of the COVID-19 pandemic. We collected information on individual beliefs about COVID-19 risks, including the likelihood of infection and the likely health implications of infection; how much people had changed their activities in response to the pandemic; and how much they believed that they would have changed their activities in a hypothetical scenario without government restrictions on activities. 1 We report three main findings. First, most people substantially overestimated the absolute level of risk. On average, participants reported that their risk of contracting COVID-19 while performing an economic activity early in the pandemic was 40% to 62%, depending on the activity. The actual prevalence of COVID-19 at the time, however, was much lower. By May 31, 2020, the U.S had reported 1,786,683 cumulative cases or a prevalence of 0.5%. 2 This is consistent with research demonstrating that, although elicited probabilistic expectations generally follow basic properties Finally, our data suggest that public mandates for activity restriction attenuate the relationship between subjective risk perceptions and behavior. As intended, by restricting population-level activity, mandates reduce the externality that lower risk individuals (who, as we find, would privately choose relatively less protective behavior) exert on higher risk individuals (McAdams 2021; Adda 2016). 5 At the same time, we also find that people who believe that they are at higher risk of an 3. We discuss the extent to which our findings on the relationship between demographics and perceived risk are consistent with recent research on this relationship during the COVID-19 pandemic in Section 3.2. 4. Our results are consistent with those of Heffetz and Ishai 2021 who document that individual risk perceptions are better predictors of behavior than case count beliefs. 5 . The fact that individuals substantially overestimated the risk of COVID-19 may in itself lead to a reduction in externalities because individuals would over-invest in protective behavior even in the absence of government mandates. At the same time, recent evidence suggests that behavioral biases may work the other direction as well, generating a "fatalism effect" -if individuals believe that COVID is widespread and inevitable, they may instead not take precautions (Akesson et al. 2020). We thank an anonymous referee for highlighting this point. In the very early stages of the COVID-19 pandemic, we developed a survey instrument aiming to evaluate knowledge about COVID-19 and daily behaviors among people in the United States. The survey was administered by Westat, a survey research firm, between May 7 and May 26. Researchers at Westat randomly selected 13,590 residential addresses across the U.S. and mailed them an invitation to participate in an on-line survey. The addresses were selected from the Delivery Sequence File maintained by the U.S. Postal Service through an authorized vendor. The invitation letter included a $1 cash incentive and told respondents they had the opportunity to contribute to policy development related to the COVID-19 pandemic and that they would receive $5 for completing the survey. To accommodate within-household sampling for addresses with multiple adult residents, the invitation letter randomly included instructions for the youngest male, oldest male, youngest female, or oldest female in the household to complete the survey. Alternate instructions asked either the oldest or youngest adult of the opposite sex to complete the survey if there were no males/females in the household. The invitation included a link to a website and a participant code that the respondents used to access the on-line survey. After completing the survey, the respondents were asked to provide their mailing address if they wanted to receive the $5 honorarium. Honoraria were mailed on June 5, 2020. 1,222 out of 13,590 individuals completed the survey-a response rate of 10% after accounting for non-deliverable addresses. The survey contained questions about the risk of contracting the virus and behavioral responses to the COVID-19 pandemic across several domains, which we elicited using visual aids (Delavande 2014). Appendix A.2 shows the key questions from the survey instrument. We asked respondents to indicate whether there were directions from their governor or officials to stay or home or shelter in place. For all respondents, we asked how the time they spent on activities such as grocery shopping and dining in restaurants had changed since before the pandemic. For those responding that there 5 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) were SIP restrictions active at the time of the survey, we asked how the time they spent on each activity would have changed if there had been no directions to stay at home. For those without SIP restrictions, we asked how their behaviors would have changed if restrictions had been in place. As 92% of survey respondents indicated that a SIP policy was active in their area of residence, we focus only on this group of individuals in our empirical analysis. 6 Heterogeneity in Subjective Risk Beliefs Our survey instrument directly asked individuals to evaluate their risk of contracting COVID-19 (on a scale from 0% to 100%) while performing the following economic activities: 1) seeing a movie in a theater, 2) eating in a restaurant (not including take-out or delivery), 3) using shared transportation such as commercial flights, trains, buses, or shared ride services, 4) personal services such as haircuts or manicures or going to the gym, 5) grocery shopping. The activities include some that are more discretionary (seeing a movie) and some that are less discretionary (grocery shopping). We also asked survey participants about their subjective beliefs regarding the risk of outcomes conditional on infection. This set of questions included the request to assess the likelihood that individuals would (1) have symptoms; (2) need medical care; (3) need to be hospitalized; (4) die if they caught the coronavirus; (5) have (insufficient) staff and supplies for treatment in a hospital. We convert the answers to these questions into a severity index by taking an unweighted mean of the five underlying indicator variables. We start our analysis by examining how the average level of subjective beliefs about risk compares to the level of COVID-19 prevalence reported by the Centers for Disease Control and Prevention, 7 and the extent to which risk beliefs vary across individuals. We next ask if individuals' subjective beliefs contain correct signals about differences in relative risk across either geographic locations or demographic groups. We use data on county-level COVID-19 prevalence, as reported 6 . Individuals' knowledge of whether there was a SIP order in their county of residence was nearly always accurate. 1,100 out of 1,219 respondents correctly reported that SIP was in place in their place of residence (we verified the county's SIP status as of April 27 2021 based on Lin 2020; Baker-McKenzie 2020; Mervosh et al. 2020; Semerad 2020; Sylte et al. 2020; Gross 2020). 22 individuals correctly reported that they were not subject to a SIP at the time of our survey. 70 individuals had an incorrect belief about having or not having a SIP order in their area, while 27 reported not knowing if they were subject to a SIP. 7. https://covid.cdc.gov/covid-data-tracker by usafacts.org in May 2020, as a measure of geographic risk exposure. We use demographic and socio-economic characteristics of individuals as self-reported in the survey. The regression equations take the following form: where Risk a i(c) is a continuous variable between 0 and 1 that measures the reported probability of contracting COVID-19 if the individual i residing in county c performed activity a. D i is a vector of demographic characteristics that includes: age (discretized into groups of (18-30 years, 30-44, 45-59, and 60+), sex, race (Hispanic, non-Hispanic Black, other), marital status (indicator for being married), education (indicator for having a Bachelor's degree or above), self-reported health status (indicator for having at least one of seven chronic health conditions, including high blood pressure, diabetes, depression/anxiety, heart disease, respiratory diseases, kidney disease, and autoimmune disorder). X c(i) is the number of days since the SIP order had been enacted. Equation 2 is an analogous specification that measures the relationship between perceived risks and the geographic prevalence of infection-GeoP revalence c(i) is the county-level number of COVID-19 cases per 1,000 in May 2020. The coefficients of interest are α a 1 and κ a 1 for each activity a that measure whether true variation in risk exposure across demographics or geography is predictive of individuals' subjective risk perceptions from engaging in activity a. All estimating equations are weighted with survey weights and replicate jackknife weights are used to compute standard errors. See Appendix Section A.3 for details on the survey weights. Private Beliefs and the Choice of Preventive Behaviors Our survey asked respondents to report how much more or less they went grocery shopping, used personal services, ate in a restaurant, saw movies in a theater, or used shared transportation since the beginning of the pandemic. The response categories for each activity included: 1) decreased a lot (by more than 50%), 2) decreased somewhat (by less than 50%), 3) has not changed, 4) increased somewhat (by less than 50%), 5) increased a lot (by more than 50%), 6) I didn't do this before the pandemic. For our baseline analysis of the relationship between beliefs and behaviors we create a simple 7 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 10, 2022. ; https://doi.org/10.1101/2022.03.08.22272111 doi: medRxiv preprint indicator variable for the respondent having reduced an economic activity by a lot (more than 50%). Then, for each activity a we estimate a linear probability model: Where, Y a i(c) is an indicator that takes a value of 1 of individual i living in county c reduced activity a by more than 50% ("by a lot"), Risk a i is a continuous variable between 0 and 1 that We report several variations of this baseline specification in the Appendix to test the sensitivity of our results. Appendix Table A .1 reports an ordinal regression model that does not collapse responses into one indicator variable for "reducing activity by a lot." In Appendix Table A.2 we report the results of the baseline regression using a sub-sample that excludes individuals who reported not participating in an activity before the pandemic. Appendix Table A .3 reports a specification that differences the risk of staying at home from the risk of other activities at the individual level instead of including is as the control variable. And Appendix Table A .4 reports a version of the baseline specification that includes the belief about the risk of dying only rather than the severity index. Qualitatively, the results of all these analyses are similar to the baseline. 8 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 10, 2022. ; Perceived Role of Public Policies To assess the role of public policies in mediating the relationship between beliefs and behaviors, we asked respondents to report how much they believe they would have changed their behavior in a hypothetical scenario without a SIP order. By comparing responses in the hypothetical scenario to the reports of the actual behavioral change, we can assess how much importance individuals ascribe to the SIP policy. We re-estimate Equation 3 using responses to the hypothetical scenario. We then compare the behavior-belief elasticity as captured by β 1 between the observed and the counterfactual regime. A lower counterfactual elasticity would suggest that individuals believe that government intervention is attenuating the effect of their own risk assessment on decision-making, while a higher counterfactual elasticity would suggest that policy interventions and beliefs are complements. In the limit, a policy intervention that removes any association between subjective beliefs and behaviors (as, Table 1 reports the summary statistics for this analytic sample. While the original survey was mailed to a nationally representative draw of postal addresses, which individuals decided to respond to the survey is not random. 8 We weight the summary statistics with survey weights to account for nonrandom non-response based on observable characteristics. The average age of respondents is 47. 49% of individuals are male. 16% are Hispanic individuals while 9% are non-Hispanic Black individuals. 53% are married, 31% have a Bachelor's degree or above, and 55% reported having at least one underlying chronic health condition among hypertension, diabetes, depression or anxiety, heart disease, a respiratory disease, kidney disease, or an autoimmune disorder. 28% reported to be an essential worker. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 10, 2022. ; the right of 50%. Finally, while individuals overestimate risk levels, the within-individual rankings of risks were intuitive and consistent with contemporaneous public reports. As we show in Figure 4 and Appendix Figure A .2, respondents believed that staying at home was the lowest risk activity and that death was the least likely outcome. Several patterns emerge. First, men believe that they face a lower risk of infection than women when performing any activity. Second, Hispanic respondents believe that they face a higher risk of contracting an infection than white respondents. Non-Hispanic Black respondents also report higher perceived risk of infection (although we lack statistical power to conclude that this difference is significant). Third, people with pre-existing health conditions believe that they have a higher probability of infection. Dark grey markers and confidence intervals report the estimate of α a In most cases, these findings are consistent with the epidemiologic evidence on the relationship between demographic characteristics and COVID-19 risk. In particular, Hispanic and Black Americans seem to be correctly aware of their heightened risk in the incidence and severity of COVID-19 11 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; believe they would have made absent SIP orders. The reported reduction in activity is substantially smaller than under SIP policies, but it is far from zero. Between 21% (for shared transportation) and 32% (for restaurants) of individuals report that they would have reduced the respective activities by a lot even in the absence of a formal SIP order. Thus, individuals clearly believe that SIP orders constrain their private choices, but they also believe that they would have undertaken substantial private preventive investments in the absence of government interventions. 10 In Table 2 , we report the estimates of β 1 in Equation 3 for self-reported reductions in activity under existing SIP orders. Those believing that the infection risk associated with individual activities was greater were more likely to reduce those specific activities. For example, a 10 percentage point (19% relative to the mean of 51.5 percentage points as reported in Figure 1 ) increase in the perceived probability of infection risk from eating in a restaurant is associated with a 2.5 percentage point increase in the probability that a respondent reduced restaurant visits by a lot (this corresponds to a 3.1% increase relative to the average probability of reducing restaurant activity by a lot, which was 79% in the sample as reported in Table 2 ). The estimate implies a behavior-belief elasticity of 0.16 for restaurant visits-a one percent increase in the subjective risk from a restaurant visit is associated with a 0.16 percent increase in the probability of reducing restaurant visits by a lot. The elasticity is also positive for the other activities we examined (ranging from 0.02 for movie theater to 0.20 for shared transportation), but less precise, lacking the statistical power to reject a zero elasticity at conventional confidence levels. Taken together, our results are consistent with the existence of a non-zero prevalence elasticity between the true risk and behavioral decisions (Philipson 2000; Oster 2018) that is mediated through changes in risk beliefs. Above, we found that individuals who live in areas with a higher prevalence of risk report higher subjective risk elicitations. Here, we observe that having a higher subjective risk assessment is correlated with a stronger reduction in economic activity. The two 10. To put our estimated magnitudes of activity reduction in context, it is helpful to compare them to the estimates of activity reductions from observational data. The estimates that are conceptually closest to ours are are reported in Farboodi et al. 2021, who find that retail and recreation activity fell a median of 33 percentage points, transit station activity by 25 percentage points, and workplace activity by 28 percentage points prior to SIPs and then as SIP orders spread across the country, the first three categories fell by a further 10, 15, and 11 percentage points. These large aggregate magnitudes of activity reductions are consistent with a substantial share of individuals in our sample reporting that they reduced these activities by more than 50%. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; https://doi.org/10.1101/2022.03.08.22272111 doi: medRxiv preprint results together imply that we would expect individuals in areas with a higher prevalence of risk to undertake more preventive behaviors and reduce their economic activities more. In Appendix Table A .7 we measure the prevalence elasticity directly by correlating the geographic prevalence of COVID and the extent of economic activity reduction as reported by our survey respondents. We find that individuals in areas with a higher prevalence of COVID-19 were more likely to reduce their economic activity. The elasticity estimates (which measure the percent change in the probability of reducing an activity by a lot in response to a one percent higher prevalence of infections) range from 0.03 for movies, restaurants, and grocery shopping to 0.17 for shared transportation in the presence of government policies restricting economic activity. 11 In the second column of each column-pair in Table 2 we re-estimate the relationship between risk beliefs and privately-preferred activity reductions that respondents believe they would have chosen absent the SIP orders. In general, this counterfactual relationship is markedly stronger than the one under existing SIP orders. This is particularly true for more 'discretionary' services. For example, the point estimate for the relationship between risk beliefs and behavior is more than twice as large for the use of personal services in the absence of SIP (0.40) than when SIP is in place (0.16). Differences in the estimates of the relationship between perceived risk of infection and behavior reduction are statistically significant with a p-value for a two-sided test of < 0.1 for movies, restaurants, and services. 12 As with belief-behavior relationship, the relationship between geographic prevalence of disease and activity reduction is stronger for the hypothetical scenario of no policy intervention. We 11. Note that this result is not mechanically driven by the existence of a SIP order as only individuals exposed to a SIP are included in our analysis. 12. It is again helpful to compare our findings to the estimates of the relative contribution of SIPs versus private behaviors from observational data. Maloney and Taskin 2020 estimate about a 60 pp drop in "voluntary" restaurant reservations and that SIP policies only accounted for 8pp out of this decline. Glaeser et al. 2021 find similarly limited impacts of SIPs on restaurants. Goolsbee and Syverson 2021 also conclude that formal restrictions contributed little to the changes in behavior. The results in Brzezinski et al. 2020 and Gupta et al. 2020, who estimate the SIP-induced versus "voluntary" choice of staying at home or general mobility, respectively, are very close to ours, estimating a 50/50 effect of SIP versus private decisions. The estimates in Cronin and Evans 2020 attribute a higher share of changes to private decisions, estimating the contribution of SIP to foot traffic in most industries to be up to a quarter of total reduction, with the rest attributable to private, self-regulating behavior. 14 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; estimate higher hypothetical prevalence elasticities for all economic activities, ranging from 0.11 for restaurants and personal services to 0.21 for shared transportation. Higher belief-behavior and prevalence elasticities that we estimate for the counterfactual scenarios without "directions from your governor or other officials to stay at home or shelter-in-place" orders are consistent with individuals believing that public policy interventions are in fact constraining their private choices. Respondents believe that in the absence of SIP policies their own subjective beliefs about risk would have played a more important role in driving their behavioral choices. The relationship between the perceived risk of a poor outcome conditional on infection and activity reductions also differs depending on whether SIP is in place. In the absence of SIP policies, we find no statistically significant association between activity reductions and beliefs regarding risk of complications for each activity, with the exception of grocery shopping. For grocery shopping, people who believe they are likely to experience complications if infected are more likely than those who perceive the risk to be low to reduce this activity a lot. By contrast, in the presence of SIP policies, those who believe that they face a greater risk of a serious complication if infected are less likely to reduce several activities by a lot. In other words, those at high risk of a poor outcome conditional on infection are more likely to maintain or increase the extent to which they go to a sit down restaurant and go to a movie theater than those who perceive themselves at lower risk. This is consistent with people with high perceived health risk reducing their protective behaviors relative to those with lower perceived risk when policies create a less risky environment. While this relationship differs for grocery shopping, we note that grocery shopping is typically exempt from SIP policies. Thus, the implementation of SIP is unlikely to reduce the risk associated with this activity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. An implication of our results is that SIP policies equalize behavior restrictions, to some extent, across those with more disparate underlying beliefs about disease risk. In particular, our results imply that formal SIP orders are likely to be more constraining (relative to the level of activities that individuals would have chosen privately) for people who believe they have (and often indeed have) a lower risk of contracting a COVID-19 infection. We found that many respondents believe that they would have dramatically reduced their activities even in the absence of formal policies restricting such activities. These findings imply that epidemiologic models of SIP effects are likely to overestimate the effectiveness of policy interventions, as individuals are likely to alter their behaviors even in the absence of interventions based on their beliefs about risks. More generally our results indicate that when designing policies that aim to change individual behavior in the presence of risk, policy makers need to consider how subjective risk perceptions are formed and how they shape behavior. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Week of the COVID-19 Pandemic in the United States." Royal Society Open Science 7 (9): 200742. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; Note: The figure reports point estimates and 95% confidence intervals from regression equations that correlate the subjective risk of contracting COVID-19 when performing different economic activities, or contracting a more severe infection, with proxies of true risk exposure. Panel A uses demographic characteristics of individuals as risk exposure proxies. Demographic characteristics were self-reported by survey participants. Panel B uses prevalence of COVID-19 in the respondent's county of residence in May 2020 as a measure of true risk exposure. Regression specification as described in Equations 1 and 2 in the main text. A separate regression is estimated for each beliefs corresponding to each economic activity (bars in shades of blue from most discretionary-movies-in the lightest shade to least discretionary-grocery-in the darkest shade) and the index of disease severity. The measure of severe infection is an unweighted average of beliefs about the probability of having the following outcomes conditional on contracting the virus: having symptoms, needing medical care, needing hospitalization, not receiving treatment when needed, and death. The sample is limited to 1,127 individuals, or 92.23% of survey respondents, who selfreported that a SIP order was in effect during the time of the survey. Survey regression weights are used to account for nonresponse. Jackknife replicate weights are used to compute standard errors. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; Note: The figure illustrates the distribution of self-reported changes in economic activity. For each economic activity survey participants were asked whether they decreased the activity by a lot, decreased somewhat, have not changed, increased somewhat, decreased somewhat, or did not do this activity before the pandemic (denoted as "NA" in the figure) . In each panel, the dark blue bars reflect self-reported change in behavior under existing shelter-in-place policies. The light blue bars represent hypothetical behavioral changes that individuals believe they would have undertaken in the absence of shelter-in-place orders. The sample is limited to 1,127 individuals, or 92.23% of survey respondents, who self-reported that a shelter-in-place order was in effect during the time of the survey. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; Note: The figure illustrates the order of risk beliefs about different outcomes (having symptoms, needing medical care, needing hospitalization, not receiving treatment when needed, and death) conditional on contracting the coronavirus. We assign ranks to each risk elicited in the survey, equal risks are assigned average ranks. Ranks increase from lowest risk to highest risk. For example, 1016 people implicitly assigned the risk of death a rank up to 3 (and 461 people assigned death the lowest rank), 99 people assigned the risk of death a rank higher than 3, and for 12 people the risk of death is missing (those are not included in the chart). The sample is limited to 1,127 individuals, or 92.23% of survey respondents, who self-reported that a shelter-in-place order was in effect during the time of the survey. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; . The dependent variable is an indicator of whether the respondent decreased the economic activity by a lot. We estimate a separate model for actual self-reported change in behavior ("policy") and self-reported change behavior in the hypothetical without SIP policies ("no policy"). The independent variables include subjective beliefs of behaviorspecific risks, as well as the index of subjective severity beliefs. The measure of severe infection is an unweighted average of beliefs about the probability of having the following outcomes conditional on contracting the virus: having symptoms, needing medical care, needing hospitalization, not receiving treatment when needed, and death. Models also include the risk of staying at home and the number days since the SIP order had been enacted. The sample is limited to 1,127 individuals, or 92.23% of survey respondents, who self-reported that a SIP order was in effect during the time of the survey. Survey regression weights are used to account for nonresponse. Stanford errors using jackknife replicate weights are reported in parentheses. * for p < .1 ** for p < .05 and *** for p < .01. . It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint -Pooling across different topics studied using subjective probabilities (mortality/survival, retirement, and stock market performance), when the actual probability of an event is greater than 50%, subjective probabilities are generally understated, while when the actual probability of an event is less than 50% subjective probabilities are generally overstated (Hurd 2009)-as we also generally observe in our data. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; https://doi.org/10.1101/2022.03.08.22272111 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; https://doi.org/10.1101/2022.03.08.22272111 doi: medRxiv preprint 3. Has not changed 4. Increased somewhat (by less than 50%) 5. Increased a lot (by more than 50%) 6. I didn't do this before the pandemic CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; We use survey weights to account for non-random nonresponse based on observable demographics in all analyses. The computation of the survey weights began with the base weights, which were then adjusted for differential nonresponse to the survey request and to within-household selection, and then calibrated to external totals. The base weight is the reciprocal of the probability of selection of the address. The adjustment for nonresponse used a weighting class adjustment to redistribute base weights of eligible nonrespondents to eligible respondents within the same weighting class, after accounting for the estimated proportion of cases with undetermined eligibility that are eligible; the weighting classes were defined using address-level variables available on the sampling frame and census tract-level characteristics from the American Community Survey (ACS; 2018 5-year tables). Candidate variables included an indicator of whether a name could be matched to the address, the dwelling type (multi-unit structure or not), census region, indicators (from USPS files) of whether the address is vacant and whether the address is seasonal, and quartiles of the following census tract-level characteristics: percent below poverty, percent with less than a high school diploma, percent with a college degree or higher, percent age 65+, percent Black, and percent Hispanic. A classification tree algorithm was used to identify the classes, with survey response status as the variable being modeled. Next, the nonresponse adjusted weights were adjusted to account for the selection of one adult among the adults in the household. The adjustment factor is the number of adults in the household. For computing basic descriptive statistic point estimates, the survey weights themselves are sufficient to account for the complex sample design. But for estimating the precision of those estimates (e.g., producing standard errors and confidence intervals), it is necessary to use a method 40 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; that takes into account the precision effects of the complex sampling and estimation procedures used in this study. The method we used was to compute replicate weights using the unstratified "delete one group" jackknife with 80 random groups. We constructed the replicates by randomly sorting the sampled addresses into 80 groups and then deleting one group at a time, to result in 80 replicates. For each replicate, a set of replicate base weights is produced by first assigning a replicate weight multiplier of 0 to the addresses that were deleted in constructing the replicate and assigning a replicate weight multiplier of 80/79 to the addresses that were not deleted in constructing the replicate, then multiplying the full-sample address base weight by the replicate weight multiplier. Each set of replicate weights underwent the same set of adjustments that were applied to the full-sample weights. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; Figure A. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. ; is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. Table 2 of the main manuscript using an ordered logit instead of a linear probability model. . It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. Table 2 excluding observations for individuals who responded "I did not do this before the pandemic" from the estimation sample. . It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. . It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. . It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. Note: The table reports the mean and standard deviation of different demographic variables for (i) individuals who completed the survey and (ii) individuals who did not complete the survey. Demographic characteristics are proxied by Census tract data. The last two columns report t-statistics and p-values of a two-sided t-test on the equality of means. Census tract socio-demographic data downloaded from https://data2.nhgis.org/main. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 10, 2022. ; Figure 2 . The dependent variables are the subjective belief about the risk of contracting COVID-19 when performing different economic activities or the severity of the disease conditional on contracting the virus. The measure of severe infection is an unweighted average of beliefs about the probability of having the following outcomes conditional on contracting the virus: having symptoms, needing medical care, needing hospitalization, not receiving treatment when needed, and death. All regression control for the number of days since a SIP order had been enacted. The sample is limited to 1,127 individuals, or 92.23% of survey respondents, who self-reported that a SIP order was in effect during the time of the survey. Survey regression weights are used to account for nonresponse. Stanford errors using jackknife replicate weights are reported in parentheses. * for p < .1 ** for p < .05 and *** for p < .01. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 10, 2022. ; Table presents the results of 10 separate linear regressions. The dependent variable is an indicator of whether the respondent decreased the activity by a lot with a separate model for actual behavior and hypothetical behavior in the absence of SIP. The independent variable is the prevalence of COVID-19 cases in May 2020. The sample is limited to 1,127 individuals, or 92.23% of survey respondents, who self-reported that a SIP order was in effect during the time of the survey. Survey regression weights are used to account for nonresponse. Stanford errors using jackknife replicate weights are reported in parentheses. * for p < .1 ** for p < .05 and *** for p < .01. . It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The copyright holder for this preprint this version posted March 10, 2022. the time you spend on each of the following activities changed since the pandemic? Q3A Going to the grocery store 1. Decreased a lot (by more than 50%) 2. Decreased somewhat (by less than 50%) 3. Has not changed 4 Increased a lot (by more than 50%) the pandemic Q3B Receive personal services such as haircuts or manicures, or go to the gym 1. Decreased a lot (by more than 50%) 2. Decreased somewhat (by less than 50%) 3. Has not changed 4 Increased a lot (by more than 50%) Decreased a lot (by more than 50%) 2. Decreased somewhat (by less than 50%) 3. Has not changed 5 Decreased a lot (by more than 50%) 2. Decreased somewhat (by less than 50%) 3. Has not changed 4 Increased a lot (by more than 50%) pandemic Q3F See a movie in a theater 1. Decreased a lot (by more than 50%) 2. Decreased somewhat (by less than 50%) 3. Has not changed 4 Increased a lot (by more than 50%)