Measured voluntary avoidance behaviour during the 2009 A/H1N1 epidemic

Managing infectious disease is among the foremost challenges for public health policy. Interpersonal contacts play a critical role in infectious disease transmission, and recent advances in epidemiological theory suggest a central role for adaptive human behaviour with respect to changing contact patterns. However, theoretical studies cannot answer the following question: are individual responses to disease of sufficient magnitude to shape epidemiological dynamics and infectious disease risk? We provide empirical evidence that Americans voluntarily reduced their time spent in public places during the 2009 A/H1N1 swine flu, and that these behavioural shifts were of a magnitude capable of reducing the total number of cases. We simulate 10 years of epidemics (2003–2012) based on mixing patterns derived from individual time-use data to show that the mixing patterns in 2009 yield the lowest number of total infections relative to if the epidemic had occurred in any of the other nine years. The World Health Organization and other public health bodies have emphasized an important role for ‘distancing’ or non-pharmaceutical interventions. Our empirical results suggest that neglect for voluntary avoidance behaviour in epidemic models may overestimate the public health benefits of public social distancing policies.

the Storm Data set includes a short description of the event, start date of the weather, end date of the weather in addition to state, and county and city identifiers of where the weather was taking place. We match the location of the event to the respondent's home location with precision to the city or county level depending on level of geographic censorship in the ATUS.
We collect Google search data to represent a subjective measure of risk. The measure is subjective because it is based on news coverage and general interest in the population rather than

Regression Model
In addition to regression models 1-3 described in the main text, we include two additional models in the supplementary material. Model 4 includes a dummy for age 65 and older individuals and an interaction of that dummy variable with the number of laboratory-confirmed cases to test whether avoidance behaviour was stronger among more sensitive individuals. Model 5 includes a similar set of terms for parents spending time with children. The ATUS does not survey individuals less than 15 years; however, respondents specify which family members were present during each activity and their demographic information. We use parents' time at home with children as a proxy for the behaviour of children less than 15 years old. Model 6 includes a third set of interactions between parents with children and Saturday to further investigate avoidance behavior on the weekend.

Regression Results
We present the full set of parameter estimates for all variables used in the regression model in Supplementary Therefore, we find no conclusive evidence of additional avoidance behavior by parents with children.
Our model accounts for approximately 20% of the variation in time spent at home ( 0.20 in models 3-6).The magnitude and statistical significance of estimates on employment, weather, and day of week demonstrate that our covariates capture certain critical factors that influence time-use patterns.

Epidemic Simulations Adjustment
Laboratory confirmed cases represent only a fraction of total cases suspected in the population [2]. Reed et al. [2] estimated that 43,677 laboratory confirmed cases through July 23, 2009 represented between 1.8 and 5.7 million cases in the population. Using the most conservative estimate, we assume that 2.4% of simulated cases are confirmed by laboratory testing.
Furthermore, confirmed cases are measured at the national level so we scale the simulated number of infected individuals (based on Phoenix MSA) to the national level by multiplying daily prevalence by 3.47 (9,734 national cases/2,800 Phoenix MSA cases during the peak of the epidemic in the third week of October [3]). The product of the proportion of laboratory-confirmed cases and the proportion of national cases to those reported in Phoenix MSA yields 8.33%.

Sensitivity of Household Contact Scalar
We investigate the sensitivity of our results to the assumption of  Supplementary Table S4.
As increases, time spent engaging in household contacts is more infectious in households with at least one infected person. If =3, household contacts are three times more infectious per minute than public contacts. The attack rate rises in the simulations with and without avoidance to over 80% of the population, but the percent change between the two simulations with and without avoidance falls by 7.9 percentage points. If =5, the attack rates with and without avoidance reach nearly 100% of the population. Under such a severe epidemic, most households become infectious, which exacerbates the effect of increased infectivity of household contacts.

Probabilistic Contact Matrix (PCM).
We The total time of public exposure on an average day between groups and at location ℓ

Sampling Uncertainty
Like any sampling method, the ATUS samples include uncertainty, which we propagate through our simulations. In the main text, we illustrate sample uncertainty with 95% confidence bars around cumulative attack rate in Figure 2. In Supplementary Figure S4 Supplementary Figure S4 shows that the sampling errors are asymmetric near the peak prevalence. The pattern of uncertainty is consistent across both simulations because of the mechanics of the model and the source of the uncertainty. We hold the contact matrix constant during each simulation. When a particular Monte Carlo sample from the ATUS yields a contact matrix with high contact minutes (i.e., respondents that happen to collocate more frequently), the simulated epidemic will spread through the population faster and the peak prevalence will occur earlier. The Matlab code and bootstrap simulation data are available upon request.