Connecting socio-demographic characteristics to income
The first panel of Table 2 summarizes the difference in means of several characteristics by income groups. Here, we have defined high income as the top three quintiles and low income as the bottom two quintiles. We find significant differences for most of these characteristics between high-income and low-income respondents. For example, non-white respondents were more likely to be low income than white respondents. Low-income respondents were slightly less likely to have a pre-existing health condition than those with high incomes. Of note, we find no significant income differences for beliefs in the effectiveness of social distancing. We also see that lower income people have a significantly smaller probability of increasing two of the self-protective behaviors we consider at the 5% level, while the other is significant at the 10% level. The second panel compares the difference in means between members of the third income quintile and the fifth quintile. Many of the differences that exist between income groups in the first panel persist here. However, unlike the first panel, we no longer see significant differences along individual characteristics such as race, access to open air at home, working part-time, continuing to work and lost income. When looking at behaviors we find no significant difference between middle- and high-income people increasing social distancing and increased hand washing or mask wearing behaviors. We still see a difference in the amount of people in each group changing any behavior. This is consistent with our discussion of the relationship between income and behavior in Fig. 1. We draw two conclusions from this table. First, there are significant demographic and behavioral differences between low- and high-income individuals in our sample. Second, while there are fewer differences between middle- and high-income respondents, significant differences along some demographics and one of our behaviors persist. These findings suggest that policymakers need to pay attention to lower income populations to ensure they engage in self-protective behaviors. Addressing income or job losses could be useful ways to encourage more widespread adoption of self-protective behaviors.
Figure 6 in the Appendix explores expected losses to labor and household income by labor status and income quintile. Expected losses to labor income are a much larger share of income for low-income respondents. For example, people in the first income quintile reported expected labor income losses of over 10% while respondents in the fifth income quintile expected losses of no more than 5%. We observe a similar pattern when looking at expected household income losses. First quintile expected losses range from nearly 20% to about 25% while fifth quintile losses range from 10% to just under 15%. We also find that the difference between the mean expected labor income loss for the first and fifth income quintile is statistically significant. Figure 4 assesses losses that have already occurred. We observe a similar relationship between income quintile and the magnitude of income losses in the first panel. The second panel examines changes in work status. We see that transitions to tele-working rise with income. We observe the reverse pattern for low-income people, who were most likely to have stopped working altogether. The third panel further examines job losses due to the pandemic. We find that the lowest income respondents had the least amount of job security and the highest probability of temporary unemployment. An interesting pattern is that higher income individuals were the most likely to permanently lose their jobs, despite also having the highest level of job security across all income quintiles. This may reflect selection: higher income jobs are more secure in general so a job loss reflects a large and permanent shift, e.g., a bankruptcy. Similar to expected income losses, the difference between mean household income losses for the first and fifth income quintile was statistically significant. Taken together, these figures demonstrate how the burdens of the pandemic have fallen especially hard on the lower income quintiles. These are the people who have lost relatively more income and employment, which may make it harder for them to engage in self-protective behaviors.
Next, we consider work arrangements by income. Figure 7 in the Appendix consists of two panels, which plot labor status (full-time, part-time, self-employed or not working) and how well they can work from home, respectively. According to the figure, full-time employment and the ability to work from home rise with income. Lower income people are more likely to report either that they stopped working or that they experienced no change in work status (which includes not having switched to tele-working). For example, nearly 75% of respondents in the fifth income quintile are working full-time. About the same percentage of respondents in the first quintile are not working. From the center panel of Fig. 4, we see that more than 50% of respondents in the fifth income quintile reported transitioning to tele-work post-pandemic, whereas only 10% of those in the first quintile did so. This pattern appears consistent with the fact that nearly 40% of low-income respondents report being unable to work from home.Footnote 17
The other broad categories of socio-demographic characteristics that are potentially associated with income and behavior changes are pre-existing health conditions, household features and beliefs related to the pandemic. From Table 1 we know that 45% of survey respondents reported having a pre-existing health condition such as diabetes, high blood pressure, heart disease, or asthma. Table 3 breaks down the prevalence of specific pre-existing conditions by income quintile. Given strong income-health gradients found in other literature, it is surprising that we do not find a meaningful relationship between income and most health conditions. For instance, we find higher proportions of diabetes among higher income people in our sample. This may reflect that the sample is not representative of the USA in terms of the relationship between income and health.
In contrast, housing is strongly related to income. As seen in Table 4, higher income respondents are far more likely to live in urban areas (versus suburban or rural), homes (versus apartments), and have access to open air where they reside compared to lower income respondents. These housing characteristics influence the cost of self-protective behaviors. For example, larger home types or having access to open air can make it easier to maintain social distancing. This is an area where policy actions now could alleviate the harm from future pandemics. Increasing access to open air in public housing or parks could be effective ways to address the behavior discrepancies we observe between low- and high-income respondents. As we noted earlier, we do not observe a pattern between income and beliefs in the effectiveness of social distancing.Footnote 18 Across all income quintiles, 70–78% of respondents believe social distancing is either very effective or extremely effective. This suggests policymakers have been successful in sharing some information about the pandemic. However, we will later discuss that there is a gap between these beliefs and the adoption of self-protective behaviors across income groups.
The survey data also contain respondents’ beliefs about various rates related to the disease such as infection rates, likelihood of contracting the disease, and so on. However, it is difficult to interpret responses. For example, consider beliefs about the local infection rate. The distribution of these beliefs is presented in Fig. 8. We can see that many respondents report implausibly small and large numbers. This could reflect several factors, including misinformation about the spread of illness, difficulties with probabilistic thinking, which is well documented in the literature (see, e.g., Barth et al. 2020; Lillard and Willis 2001; Delavande et al. 2006), fatalistic beliefs (e.g., Akesson et al. 2020), optimism about herd immunity, etc.Footnote 19 Another possibility is that misinformation about the pandemic initially spread by the federal government and news media may have “spoiled” the way people link less direct beliefs to outcomes. This in part motivates why we opt to include beliefs about the effectiveness of social distancing to ensure that people who believe these behaviors are effective are also the people engaging in them more frequently. In any case, these interpretational difficulties will limit conclusions we can draw using some of the beliefs variables.
Factors associated with behavior change
Our analysis until now shows that several factors are related to income. These factors could potentially help to explain differences across income groups in self-protective behaviors depicted in Fig. 1. In this section, we explore which socio-demographic characteristics are associated with behavior changes. We begin by providing summary statistics and then discuss and analyze findings from our main estimates from regressions of behavior change onto different sets of explanatory variables.
Table 5 summarizes the difference in means across various characteristics for those who increased social distancing behavior according to our metric. We find significant differences between males and females and between those who believe in the effectiveness of social distancing. These findings suggest that women are more likely than men to increase social distancing as are those who believe strongly in the effectiveness of social distancing. We see a significant difference between those who increased social distancing behaviors and work statuses. Specifically, we see that respondents who continued to work (as opposed to stopping work or transitioning to tele-work) engaged in increased social distancing at a lower rate than those who had transitioned to another work arrangement. We also see that respondents in New York and California were more likely to increase social distancing behaviors than individuals from Texas or Florida, which may be explainable by political factors. Older respondents also increased social distancing more than younger respondents. This pattern appears consistent with the elevated risk elderly people face from Covid-19. Another significant difference can be seen between respondents without access to open air at home, which we view as a barrier towards adopting this self-protective behavior. We do not find significant differences across other individual characteristics. Finally, we also see that people increasing social distancing behaviors have a significantly larger probability of increasing other self-protective behaviors. This finding is reassuring, as we would expect to see co-movement among the various measures we consider.
For our main analysis, we examine three outcomes: any behavior change, social distancing, and mask wearing or hand washing. As mentioned earlier, while we examine three different outcomes, we are testing one main hypothesis: whether socio-demographic factors predict the adoption of self-protective behaviors. We view using three measures of behavior as important. First, each of these behaviors impose a different cost for individuals. Different types of people may be more responsive to low-cost behaviors such as changing behaviors or hand washing-mask wearing as opposed to more costly behaviors such as social distancing. Second, as these behaviors are correlated with one another, common findings across each of these behaviors serves as a form of robustness for the demographic patterns we identify.
For each outcome, we estimate linear probability models as a function of income and different sets of explanatory variables. We use heteroskedastic robust standard errors. Our main findings are summarized in Table 6. We discuss these results and other findings in greater detail in the following subsections. In general, we find that income, work arrangements such as tele-working, lost income and beliefs about the effectiveness of social distancing are significantly associated with the self-protective measures we examine. Detailed results are presented in Tables 7, 8 and 9 in the Appendix. In each table, all columns include income quintiles as explanatory variables. Column (1) includes only income, column (2) adds in socio-demographic characteristics, column (3) adds in pre-existing health conditions, column (4) brings in housing characteristics, column (5) introduces work arrangements and economic loss characteristics, and column (6) adds in beliefs about social distancing and local infection rates and perceived benefits from the pandemic. Finally, in column (7) we include all of these sets of controls in a single specification. We will discuss each of these columns in the following subsections.Footnote 20
Across all three of our dependent variables we find strong, statistically significant associations with income. Higher income individuals are more likely to engage in the behaviors we examine. To fix ideas, relative to the first income quintile, a member of the fifth income quintile is 10–15 percentage points more likely to change their behaviors, 11–24 percentage points more likely to increase social distancing behaviors, and 17–25 percentage points more likely to increase hand washing or mask wearing. Put another way, when all controls are included, a member of the fifth income quintile is 13% more likely to change their behaviors, 32% more likely to increase social distancing and 30% more likely to increase hand washing or mask wearing. We find that these income effects are fairly robust to the inclusion of controls. From the baseline to the case where we include all of our controls, the size of the coefficient estimates remain fairly stable as we add additional variables, which means that these other factors do not fully explain the income gradient. The slight exception is for the increased social distancing outcome. When all of our controls are added, we only see a significant difference between the fifth and the first income quintiles, suggesting that our explanatory variables help to explain the relationship between income and what appears to be the a costly self-protective measure. In general, the income gradients presented here strongly suggest that the adoption of self-protective behaviors is a costly prospect, one that is easier for people with more income. While providing cash transfer could help, the income gradients alone do not provide very much policy guidance. Thus, we now consider whether additional factors associated with self-protective behavior adoption.
Gender, age, race and location
The next set of control variables we examine are gender, age, race and state. We do not find many significant associations between these factors and the change behaviors outcome. In the baseline case we see negative associations between males, people 56 years or older, and some regional effects. Some of these relationships lose significance when other variables are added to the analysis, though we find more robust patterns when examining increases in social distancing behavior. We find strong negative associations between males, and respondents from Florida and Texas, which maintain significance once other controls are added. To fix ideas, we find that males are 23% less likely than females to increase social distancing. This could be evidence that the pandemic is driving women into “traditional” care taker roles—staying at home to maintain the household—while males continue to work in person and are unable to adopt social distancing behaviors. Similarly, we find that relative to respondents from California, people in Texas and Florida tended to be 21% and 22% less likely to increase social distancing, respectively. These results may presage the surges in Covid-19 cases that happened in these two states that began towards the end of June 2020. This finding complements recent work that has found a significant relationship between political affiliation and beliefs about the Covid-19 pandemic and the adoption of social distancing behaviors. These are states led by governors who were less responsive to the outbreak of the pandemic, which may have had an influence on behavior. Finally, we find positive significant associations between race and people 56 years or older for the hand washing-mask wearing outcome. Specifically, we find that Black respondents are 19% more likely than white respondents to increase hand washing or mask wearing.Footnote 21 We find a similarly sized relationship for those 56 years or older. It is interesting that find these effects for increased hand washing or mask wearing. This may reflect the fact that of the three activities we examine, this one is a relatively low-cost way to self-protect for people who face risks, but are unable to engage in higher cost, less practical activities, such as social distancing.
We also examine various pre-existing health conditions, including diabetes, high blood pressure, heart disease, asthma, allergies, and other conditions. Overall, and surprisingly, these variables are not strongly correlated to behavior change. Oddly, we find a strong negative association between heart disease and increased social distancing, which may reflect that people with heart disease are generally unhealthy and thus less likely to engage in self-protective behaviors. Yet, it is surprising that health conditions more strongly associated with serious illness (e.g., diabetes, asthma, or high blood pressure) are not associated with behavior change. An exception is that we find a robust significant association between allergies and increases in hand washing and mask wearing. Allergies would presumably be less likely to be associated with unobserved factors capturing an unwillingness or inability to engage in self-protective behaviors. Another possibility that people with allergies could feel they are becoming ill even if they are not and thus be more willing to take precautions. More generally, the lack of a health gradient could be a byproduct of our non-representative sample. As mentioned previously, pre-existing conditions were not targeted for representativeness either in the data collection process. As our understanding of Covid-19 grows, future data collection efforts may want to target specific health conditions to determine if an association with self-protective behaviors exists.
Next we examine housing characteristics. We find a negative significant relationship between respondents in the countryside and changing behaviors but this association becomes indistinguishable from zero as other controls are added. We find a robust negative association for having no access to open air space at home and increased social distancing behavior. In our full control case, we find that respondents that live in homes without open air access are 20% less likely to increase social distancing behaviors. We find this to be an intuitive result. People who are more comfortable sheltering-in-place are more likely to do it. Policies aiming to slow the pandemic should take these factors into account as they suggest cramped and uncomfortable housing can potentially undermine efforts to “flatten the curve.” This result could also guide the design of future housing policy as government prepare for future pandemics. For example, communities could increase the size and availability of public parks to accommodate social distancing. Governments could also prioritize the opening of parks or other open public spaces during a pandemic, though of course the risks in terms of increased exposure would need to be weighed against the benefits in terms of higher rates of social distancing among low-income people. Another, longer run possibility would be to incorporate some open air spaces such as balconies or community gardens into the designs for public housing, which would help facilitate social distancing behavior. Finally, we find similar patterns for the two other outcome variables, though estimates are less precise in the final specification.
Work arrangements and losses
We also consider work arrangements and economic losses. In general we find fairly consistent results across all three of our outcome variables. People who transitioned into tele-working are more likely to change behaviors, increase social distancing, and increase hand washing-mask wearing. This association ranges from roughly 9–15 percentage points relative to somebody who continued to work. When all controls are included, a person that transitions to tele-work is 20 to 28% more likely to increase these self-protective behaviors. This effect is robust to the inclusion of other controls. We find a similarly sized effect for those that stopped working or never worked but significance was retained with less consistency. This result is intuitive. People who can work from home are more likely to abide by stay at home orders. Factors related to work arrangements, which vary across socio-demographic groups, can determine the sustainability and effectiveness of policies aiming to prevent the spread of illness.Footnote 22 From a policy perspective, governments could offer incentives or resources for firms to increase the availability of tele-work to their employees. Beyond preparing for future pandemics, this policy may also help firms compete for talent along non-monetary dimensions. For example, it appears that tele-medicine may remain after the Covid pandemic (Smith et al. 2020). Of course, this type of policy has its limits as some work, by its nature, requires physical contact.
We also find that realized household income losses have a significant positive association with each of these behaviors. After controlling for extreme lost income values, we find that for every $1,000 lost a respondent is 1–4 percentage points more likely to adjust each of the behaviors we examined. People who have experienced these losses have already been harmed by the pandemic. As a result, they may be more careful than others and view contracting the disease as a higher risk. Another possibility is that these people have fewer monetary resources and may not have money to cover medical expenses if they were to contract the disease. This speaks to the importance of polices that provide direct monetary relief to people during a pandemic. Our results suggest that policies such as the CARES Act, which featured direct compensation to all citizens is an effective way to promote the adoption of self-protective behaviors. Other studies of the CARES Act have shown it was a useful short-term policy to help people smooth their consumption during the pandemic (Carroll et al. 2020).
Beliefs and perceptions
The final set of variables we examined were beliefs and perceptions about the pandemic. Reassuringly, we find a fairly consistent effect for beliefs in the effectiveness of social distancing across the three behaviors we included. These findings are strongest for the changed behaviors and increase social distancing variables. We find similar results but with weaker significance for increased hand washing-mask wearing. Yet, there is a disconnect between these beliefs and individual behavior. For example, approximately 97% of respondents in the first and fifth income quintiles believe social distancing is effective. However, 80% of people in the first income quintile reported changing any behavior while 93% people in the fifth income quintile reported any behavior change. We see a similar discrepancy when looking at increases in social distancing behaviors: 45% of respondents in the first quintile and 57% in the fifth income quintile. This leads us to the somewhat depressing conclusion that is, however, entirely in line with our results: many people from lower income groups recognize the effectiveness of self-protective measures, but do not adopt them, suggesting different costs of doing so compared to higher income people. For example, low-income people tend to not have access to tele-working, which affords people the opportunity to maintain social distancing without having to sacrifice labor income.
One finding that surprised us was the negative association between beliefs about local infection rates and increases in self-protecting behaviors. As discussed previously, the distribution of respondent beliefs about local infection rates has significant mass at the low and high end, which are difficult to reconcile with reality. In Figs. 9, 10, and 11 we present Lowess smoother results for three behavioral outcomes of interest and this belief.Footnote 23 In each case, people who reported an infection rate of 20% or fewer exhibit the expected response: a rise in perceived infection rates is associated with more protective behavior. Thus, negative coefficient estimates are driven by people with implausibly high perceptions of infection rates. This could reflect respondent confusion. It could also reflect a sort of fatalism, i.e., people believe infection rates are so high that they are bound to become infected, too, and thus don’t bother to engage in protective behaviors. Fatalism is a well-documented phenomenon in several fields (see, e.g., Akesson et al. 2020; Ferrer and Klein 2015; Shapiro and Wu 2011). These findings suggest that pockets of misinformation persist within the population. It is crucial that policymakers provide accurate and complete information about the risks of the pandemic and how people can best protect themselves from infection.
We also find some positive associations between perceived benefits from the pandemic and increases in our behaviors of interest. Less pollution and more family time were two that came out as significant and tended to retain significance as other controls were added. In Appendix Table 11, we present cross-tabulations of the survey data which indicate most of the people identifying these benefits belonged to higher income quintiles.
We conduct a series of robustness checks to these specifications. The results for these analyses are available in the Supplemental Appendix of Papageorge et al. (2020). First we consider whether a respondent was engaging in these self-protective behaviors at all following the start of the pandemic. We also look at whether there were distinct behavior differences between those that had experienced a loss due to the pandemic and the pooled sample. In another test examine each state individually and pooled groups of states. In general, these analyses align with our main results. We also consider the intensive margin for increased social distancing and hand washing or mask wearing behaviors. A respondent’s income and beliefs about the effectiveness of social distancing did not have a significant association with larger increases in either of these self-protecting behaviors. Other effects are consistent with our main analysis. Finally, as mentioned previously, the data collected by Belot et al. (2020b) do not contain information on educational attainment. We use information on a respondent’s profession to construct a proxy for whether they have a college degree and incorporate it into our specifications.Footnote 24 We find that education is positively associated with increases in self-protective behaviors. The inclusion of this variable does not appreciably alter our other findings.