Quantifying population contact patterns in the United States during the COVID-19 pandemic

SARS-CoV-2 is transmitted primarily through close, person-to-person interactions. Physical distancing policies can control the spread of SARS-CoV-2 by reducing the amount of these interactions in a population. Here, we report results from four waves of contact surveys designed to quantify the impact of these policies during the COVID-19 pandemic in the United States. We surveyed 9,743 respondents between March 22 and September 26, 2020. We find that interpersonal contact has been dramatically reduced in the US, with an 82% (95%CI: 80%–83%) reduction in the average number of daily contacts observed during the first wave compared to pre-pandemic levels. However, we find increases in contact rates over the subsequent waves. We also find that certain demographic groups, including people under 45 and males, have significantly higher contact rates than the rest of the population. Tracking these changes can provide rapid assessments of the impact of physical distancing policies and help to identify at-risk populations.


Supplementary note 1: City samples
For our city-specific samples, we recruited respondents who lived in the Designated Market Area (DMA) surrounding each city. DMAs are intended to capture media markets, and therefore often include much more than just the urban core of a city.
Our city-specific samples recruited respondents in the DMAs associated with six cities: Atlanta, the San

Supplementary note 2: Comparison data from 2015 Facebook survey
The pre-pandemic baseline contact estimates come from a survey conducted among a random sample of Facebook users. The study interviewed 4,288 respondents in the United States; respondents were randomized to report about one of two di erent types of contact: either conversational contact or sharing a meal. Here, we focus on the results for conversational contacts, which is comparable to the contact we ask about in the BICS study. The sampling frame for survey respondents was people who use Facebook; a more detailed discussion of the study design can be found in the original paper. 9 Figure 4 compares, for a given age group, contact rates among people who used Facebook in 2015 and people who we contacted through the online panel in 2020. We expect these two groups -people who use Facebook and people reached by the online panel -to be very similar; however, if there are di erences in contact rates between these two groups, then the estimates in Figure 4 could be a ected. Note that, as Figure 4 shows, our substantive conclusions are robust to using POLYMOD data from the UK as an alternate to the Facebook survey as a baseline.

Supplementary note 3: Bootstraped contact matrices
Uncertainty estimates for the descriptive results -including Figure 1 and the R 0 analysis summarized in Figure 4 -are based on the bootstrap. To obtain bootstrap resamples, we resampled respondents separately in each stratum (i.e., in each city and in the national sample). For each bootstrap resample, we first drew n c samples with replacement from among the n c respondents in city c. For resampled respondents who provided detailed reports about all contacts (i.e., those respondents who had a i = 1 for all contacts i), we used the reported detailed contacts without resampling. For resampled respondents who did not provide detailed reports about all contacts, but who reported about a subset of contacts (i.e., for each respondent who had a i > 1 for some contact i), we resampled r i contacts with replacement from among the respondent's r i detailed contacts. This second stage is intended to capture sampling variation due to the respondent choosing which contacts to report about from among her total contacts. Using this approach, we obtained 5,000 bootstrap resamples of our dataset, and these bootstrap resamples are the basis for uncertainty estimates.

Supplementary note 4: Average number of reported contacts
Here, we present additional analyses of reported contacts. First, Supplementary Next, in addition to the model fit to non-household contacts (Figure 2), we also fit a model to all contacts;

Race/Ethnicity
White, non-Hispanic

Supplementary note 5: Relationships and locations
Here, we present Tables containing the values shown in Figure 1, Panels C and D. These numerical values may be useful as numerical inputs for future modeling work.

Supplementary note 6: Sensitivity analysis for contact definition
In Wave 0 (the pilot study), contact was defined using 'conversational contact', as explained in this text: By in-person conversational contact, we mean a two-way conversation with three or more words in the physical presence of another person.
You might have conversational contact with family members, friends, co-workers, store clerks, bus drivers, and so forth.
(Please do not count people you contacted exclusively by telephone, text, or online. Only consider people you interacted with face-to-face.) After the pilot study, in Waves 1 and up, the survey instrument was modified and contact was defined using this text: Now we would like to ask you some questions about people you had in-person contact with yesterday.
By in-person contact, we mean EITHER a two-way conversation with three or more words in the physical presence of another person OR physical skin-to-skin contact (for example a handshake, hug, kiss, or contact sports) You might have in-person contact with family members, friends, co-workers, store clerks, bus drivers, and so forth.

Please do not count people you contacted exclusively by telephone, text, or online.
Only consider people you interacted with face-to-face.
In the main text, for Waves 1 and up we combine physical and conversational contacts together, since the combination of these two is most relevant for disease transmission. However, just under 10 percent of non-household detailed contacts in Waves 1 and up were reported to be strictly physical (and not conversational

Supplementary note 7: Sensitivity analysis for baseline R 0 value
Our estimate of R 0 for each survey wave was calculated assuming a distribution for R 0 for COVID-19 in the absence of physical distancing. In the main text, we assume that R 0 prior to physical distancing followed a normal distribution with mean 2.5 and standard deviation of 0.54 based on estimates from literature. 5,26 Since there is a wide range of R 0 estimates in published literature, we repeated the analysis for estimating R 0 under physical distancing with higher and lower estimates of baseline R 0 values. We used the range of estimates of state-level R 0 (in the absence of physical distancing) reported in. 29 We find that for the highest baseline R 0 estimated by Pitzer et al 29 (5.17 for Missouri), the reduction in contacts is not su cient to reduce R 0 to below 1 in Wave 0; otherwise the results remain qualitatively similar. Supplementary figure 6: R 0 estimates from the BICS contact matrices for each wave relative to two baseline R 0 values. The implied R 0 from the BICS contact matrices for each wave is calculated from age-structured contact matrices for all reported contacts (n = 3, 163 in Wave 0, n = 7, 473 in Wave 1, n = 7, 842 in Wave 2 and n = 11, 402 in Wave 3), and assuming that baseline contact patterns is equivalent to the 2015 Facebook study. The high baseline R 0 value is drawn from a normal distribution with mean 5.17 and standard deviation of 0.54. The low baseline R 0 value is drawn from a normal distribution with mean 1.92 and standard deviation of 0.54. 95% confidence intervals were derived from the bootstrap.

Supplementary note 8: Comparison of business-as-usual age-structured contact matrices
In the main text, we compare the BICS age-structured contact matrices to a baseline business-as-usual scenario from a 2015 study, 9 based on a probability sample of US Facebook users. As a sensitivity analysis, we also estimate relative changes in R 0 assuming that baseline contact patterns were equivalent to contact patterns from the UK POLYMOD study, 3 which has been widely used in many settings. Contact patterns in both studies are based on empirical estimates, similar to the BICS study. In the absence of empirical data, however, contact patterns are often derived from model-based estimates that take into account the demographic structure of the population and empricical data from other populations. We compared the contact patterns from Feehan and Cobb 9 and from Mossong et al 3 to model-based estimates of business-as-usual contact patterns in the US from Prem et al 8 . Supplementary figure 7 shows the three contact matrices, and the average number of contacts by age group. We see very similar patterns of assortativeness of contacts by age, although the absolute levels of contact di er slightly across the three studies.