We begin by presenting descriptive results of the survey and follow with our primary analytical questions from the survey, namely, testing the hypothesis that respondents' affective state mediates their protective action.
We gathered 6,249 responses from 28 April to 5 May 2009. presents descriptive statistics of the sample.
presents the distributions of respondents' contacts within the 24 hours prior to taking the survey.
Frequencies of the categories of respondents' contacts outside the home.
presents the means of the subjective threats. Swine flu had a mean second only to injury, and the highest among the infectious sources of threat. The mean of perceived threat from swine flu fell above the Bonferroni-corrected 95% confidence interval for all other threats but unintentional injury.
Means for the perceived threat levels for different sources of risk.
presents the frequency distribution of perceived personal risk. There is a notable bimodality to this plot. This apparent bimodality is not simply attributable to sampling error since the difference between the responses
4 vs. those
5 vs. those
6 is in excess of 300. Further analysis using finite mixture models 
provides strong statistical support for the reality of the bimodal pattern (results not shown). While the majority of respondents felt that their personal risk was low, there is a second mode rating their risk as intermediate (
5). This same bimodal pattern can be seen in the frequency distribution of personal empowerment (i.e., ability to avoid infection) shown in . While most respondents indicate that they are confident they can avoid infection, a substantial second mode appears at the intermediate value.
Frequency distribution of personal risk perceptions.
Frequency distribution of personal empowerment.
shows the frequency distribution of protective behaviors. We can see that nearly 80% of respondents report washing hands more frequently, while very few avoid work or school or wear protective masks.
Frequency distribution of the protection measures undertaken by respondents.
shows the means for respondents' information sources. Not surprisingly, the most common source of information reported was the Internet. Again, mean values are plotted with their 95% Bonferroni-corrected confidence intervals. With the exception of social-networking tools (e.g., Facebook, Twitter), all other media sources are statistically indistinguishable from each other, with the social-networking tools being used significantly less.
Mean values of sources of information on swine flu cited by respondents.
The results of the model for the protection index show a number of robust trends (). In particular, we find that age increases and male gender decreases the protection index. Receiving a large amount of information from the internet, television, and health officials all increase the protection index while receiving large amounts of information from print media, friends, or social networking media has no effect. The number of household members has no discernible effect, though the number of contacts outside the home does. For the ordered factor “contacts,” the first category (<5 contacts in the past 24 hours) is the reference category. Interestingly, relative to respondents with the fewest number of contacts, all other contact categories have reduced protection indices, indicating that people with fewer contacts take more protective actions. Not surprisingly, residence in Mexico has a large positive effect, while residence in Canada or Europe decreased the index. The day that the survey was taken (29 April
1) had a negative effect on the index, indicating that respondents took less protective action as the epidemic proceeded. Respondents' reported subjective anxiety has a substantial impact on the index with high anxiety increasing protection, supporting our hypothesis that affective state mediates protective behavior.
Results of the binomial GLM for the protection index.
Increased hand-washing showed similar trends to the model for the protection index (). Male gender decreases while age and survey day increase the odds of increasing hand-washing. Receiving a large amount of information from the internet, radio, television, and health officials increase, while living in Europe or Australia/New Zealand decrease the odds. As with the overall protection index, perception of risk and subjective anxiety significantly increase the odds of increased hand-washing modestly.
Results of the binomial GLM for increased hand-washing.
Changes in Behavior
A key epidemiological question is how people's affective status and protective behaviors undertaken change as the epidemic proceeds. To develop a measure for this, we cross-tabulated individual values of the protection index and affective status by survey day. Pearson's chi-square test for independence of both tables was strongly significant (affective: χ2
<0.001; protection: χ2
54, p<0.001), indicating substantial departure of cells from the expected values. To visualize the pattern of departure from the expected values, we calculated an expected tables taken as the cross-product of the marginals of the observed table normalized by the grand sum. We combined rows of these tables to simplify the presentation, plotting the difference between observed and expected tables for a high, medium and low emotional status/protection index respectively. For example, a value of −51 on the calm affective status on day one means that there were 51 fewer responses in the calm categories than would be expected by the overall marginal distribution of responses across all days.
In and , we plot the change in respondent's protective behavior and emotional status over the first week of the survey. The lines represent the differences between observed and expected frequencies of responses for the 9-point scale simplified to three levels each. We see that by day three of the survey (May 1st), the relative number of people reporting a calm emotional state was very high, while the number of people reporting high values of the protective index declined dramatically. We interpret these results to indicate that people's response to a potential pandemic is quite sensitive to media reports.
Changes in difference between observed and expected values of the protection index over the seven days of the survey.
Changes in difference between observed and expected values of the emotional status over the seven days of the survey.
In general, individuals' survey responses to perceived risk for the eight health threats were only moderately correlated, with pairwise correlations typically well under 0.5. PCA did not reveal that a substantially reduced number of dimensions explained these correlated data – six principal components were required to explain 85% of the variance. Nonetheless, some intriguing PC loadings revealed themselves. In particular, the second PC, which explained 15.6% of the variance in the data, showed strong positive correlations with swine flu (r
0.516), bird flu (r
0.530), and terrorism (r
0.467). All three of these threats receive a great deal of media attention and their fundamental uncertainty are likely to generate an inordinate amount of fear vis-a-vis their actual threat