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

The study investigates how physical distancing policies implemented in the United States during the COVID-19 pandemic affected close interpersonal contact patterns, which fundamentally drive SARS-CoV-2 transmission. Following widespread stay-at-home and activity restrictions beginning in March 2020, the authors sought to quantify changes in the number and structure of interpersonal contacts over time and to assess implications for transmission by estimating the impact on the basic reproduction number, R0. The study introduces the Berkeley Interpersonal Contact Survey (BICS), which measures both total contacts and detailed attributes of contacts (e.g., age, relationship, location) to inform age-structured epidemiological models and identify populations at higher risk. The primary research questions are: how have contact rates and patterns changed over the course of the pandemic in the US, which demographic groups have higher contact rates, and how do these changes translate into changes in the implied R0 under physical distancing.

The paper builds on prior work linking social contact patterns to infectious disease spread, notably the POLYMOD study and subsequent contact surveys that provide age-structured mixing matrices used in transmission models. It notes limited pre-pandemic US estimates of contact rates and patterns and uses a 2015 US Facebook user probability sample as a baseline for comparison. The study references evidence from other countries documenting large pandemic-related reductions in contacts (e.g., China, UK, Luxembourg, Italy, Belgium, France, Netherlands) and the use of contact matrices and dominant eigenvalues to estimate relative changes in R0.

Design and data collection: Four survey waves were fielded: Wave 0 (pilot) March 22–April 8, 2020; Wave 1 April 10–May 4, 2020; Wave 2 June 17–23, 2020; Wave 3 September 11–26, 2020. Total respondents: 9,743 (Wave 0 n=1,437; Wave 1 n=2,627; Wave 2 n=2,431; Wave 3 n=3,248). Respondents reported the number of people they had conversational contact with on the day prior to the interview; Waves 1–3 additionally asked about physical contact. Up to three contacts per respondent received detailed characterization (contact age, sex, relationship to respondent, duration, and location). In Wave 0, respondents reported all contacts and then identified non-household contacts. From Wave 1 onward, respondents provided a household roster and then reported only non-household contacts. Sampling and recruitment: The instrument was programmed in Qualtrics; respondents were recruited via the Lucid online panel. Each wave included a US quota sample intended to be nationally representative and city-specific quota samples (New York, San Francisco Bay Area, Atlanta, Phoenix, Boston; Philadelphia added in Wave 1). All participants provided informed consent; UC Berkeley IRB approval (Protocol 2020-03-13128). Weighting: A model-based inference approach was used for the quota sample. Respondent-level calibration weights (pseudo-inclusion probabilities) were constructed to align with population margins for age categories (18–23, 24–29, 30–39, 40–49, 50–59, 60–69, 70+), sex, age-by-sex, education, race (white, Black, other), Hispanicity, household size (1–5+), and county urban/suburban/rural status (from CDC), with population values primarily from the 2018 ACS (IPUMS). Zip-to-county mapping used Sood (2016). Contact-level within-respondent weights adjusted for cases where respondents reported more than three contacts but provided details for only three; each detailed contact received weight ai = di/3, where di is total contacts for respondent i. Population-level estimates of contact characteristics combined respondent and contact weights. Statistical modeling of contacts: Negative binomial regression models were fitted separately for (1) total contacts and (2) non-household contacts. The log expected count was modeled as μi = α + Xiβ, with predictors including age category, gender, household size, survey wave, city, race/ethnicity (Non-Hispanic White, Non-Hispanic Black, Hispanic, Non-Hispanic Other), weekday indicator, and interactions (age×sex, wave×race/ethnicity, wave×city). Overdispersion was handled via a shape parameter φ. Right-censoring was addressed: Wave 0 capped at “10 or more” contacts; Waves 1–3 allowed any number but analysis top-coded at 29. Bayesian estimation used flat priors on β, weakly informative priors on α (Student-t(3,0,10)) and φ (Gamma(0.01,0.01)). Four chains with 1000 warmup and 1000 sampling iterations; R-hat ≈ 1. Conditional effects plots summarized results. Epidemiological modeling: Age-structured contact matrices were constructed for age groups 0–18, 18–25, 25–35, 35–45, 45–65, 65+. Weighted average daily contacts from respondents in age group j with contacts in age group i produced raw matrix M with entries mij. Reciprocity was imposed to obtain matrix C using population sizes Ni and Nj; for the 0–18 group (no respondents), within-group contacts were imputed by scaling the UK POLYMOD within-age contact (with school contacts removed) by the ratio of dominant eigenvalues between BICS and POLYMOD for overlapping ages. Relative changes in R0 were inferred from ratios of dominant eigenvalues of C matrices between BICS waves and baseline (2015 US Facebook probability sample; sensitivity with UK POLYMOD). Theoretical implied R0 during each wave was computed by multiplying relative changes by an assumed baseline R0 ~ Normal(mean 2.5, SD 0.5). Bootstrap (5000 samples) produced uncertainty intervals. Mask usage sensitivity: From Wave 1 onward, R0 was also estimated using only contacts where no mask usage was reported to bound the impact of masks.

- Sample and reporting: 9,743 respondents across four waves reported 49,321 total contacts; 29,880 contacts had detailed information. Analyses used calibration weights to approximate population inference. - Contact levels: Median total contacts (non-household in parentheses) per respondent by wave: Wave 0 = 2 (0); Wave 1 = 3 (1); Wave 2 = 3 (1); Wave 3 = 4 (2). Interpersonal contact increased progressively after Wave 0. - Reductions vs. pre-pandemic: Compared to a 2015 US baseline, estimated declines in daily interpersonal interactions were 82% in Wave 0, 74% in Wave 1, 68% in Wave 2, and 60% in Wave 3. Largest absolute decline was observed in ages 25–35 during Wave 0. Despite low absolute contact levels, age-assortative mixing patterns persisted. - Composition and locations of non-household contacts: In Waves 1–3, family, friends, and work colleagues accounted for most non-household contacts, with increasing interactions at work and at home by Wave 3; contacts at stores/businesses increased from Wave 0 to Wave 1. - Demographic and geographic correlates (negative binomial models for non-household contacts): Younger respondents (<45), particularly males, reported higher contact rates than older adults. Racial/ethnic patterns shifted over time: highest contact rates among Black and Hispanic respondents in Wave 1, and among White respondents by Wave 3. Household size and day-of-week showed little association. Geographic trends varied: steady increases from Wave 0 to 3 in the Bay Area and Phoenix; more uneven patterns in Atlanta, Boston, New York, and Philadelphia. - Implied transmission potential (R0): Relative declines in implied R0 compared to pre-pandemic baseline were 73% (95% CI: 72–75%) in Wave 0, 57% (53–61%) in Wave 1, 48% (43–53%) in Wave 2, and 36% (29–42%) in Wave 3. Assuming baseline R0 = 2.5, implied R0 values were: Wave 0 = 0.66 (95% CI: 0.38–0.96); Wave 1 = 1.06 (0.61–1.53); Wave 2 = 1.29 (0.74–1.86); Wave 3 = 1.59 (0.91–2.30). Sensitivity using UK POLYMOD as baseline produced qualitatively similar results. Accounting for mask usage (restricting to contacts with no reported masks) reduced the relative increase in implied R0 in Waves 1–3, suggesting masks mitigate transmission risk beyond distancing. - Cross-national context: The magnitude of contact reductions in Wave 0 aligns with declines reported in other countries (e.g., 74–88% reductions).

The findings demonstrate that US physical distancing measures in March–April 2020 achieved substantial reductions in close interpersonal contacts, which translated into a theoretical implied R0 below 1 during Wave 0, consistent with suppressed transmission potential in a fully susceptible population. As restrictions eased and economic and social activities resumed, contact rates rose across waves, particularly in work and home settings, pushing implied R0 above 1 by June and further by September. Persistent age-assortative mixing and higher contact rates among younger adults and males indicate key groups likely to sustain transmission and thus priority targets for interventions. Temporal shifts in race/ethnicity-specific contact rates highlight changing exposure risks across groups as policies and behaviors evolved. Geographic heterogeneity underscores the role of local policies and adherence. Mask usage analyses suggest non-pharmaceutical interventions beyond distancing can attenuate increases in transmission potential. Overall, tracking contact patterns provides actionable, real-time inputs for evaluating and calibrating interventions and informing age-structured epidemiological models.

This study quantifies substantial, time-varying reductions in interpersonal contacts in the US during 2020 and links these patterns to implied changes in transmission potential using age-structured contact matrices. The work provides detailed demographic, relational, and locational breakdowns of contacts, identifies groups with higher contact rates (younger adults, males), and documents geographic heterogeneity. It also shows that distancing alone substantially reduced contacts, with additional mitigation likely from mask usage. The Berkeley Interpersonal Contact Survey (BICS) offers an ongoing platform to monitor contact patterns and evaluate policy impacts. Future research should integrate these empirically derived contact matrices into dynamic, age-structured transmission models; expand to probability-based sampling; include children to capture school-related contacts; collect multilingual responses; and continue monitoring as schools/workplaces reopen and interventions change.

- Sampling: Quota-based online panel (non-probability sample) may introduce selection bias; while calibration weighting improves representativeness, it does not eliminate bias. Obtaining a national probability sample was impractical during the pandemic. - Recall and social desirability: Contacts were self-reported for the previous day, potentially subject to recall error and social desirability bias given awareness of distancing policies. - Language: Surveys were conducted only in English, excluding non-English speakers. - Age coverage: Children (<18 years) were not surveyed, requiring imputation for the youngest age group in contact matrices and limiting direct estimation of child-child contacts. - Modeling assumptions: Implied R0 estimates assume unchanged disease-specific parameters over time and use baseline contact matrices (2015 US Facebook sample or UK POLYMOD). They do not account for age-specific differences in susceptibility/infectiousness, nor for changes in transmissibility due to other factors (e.g., ventilation, variant emergence). Top-coding and censoring handling may affect tail behavior of contact distributions.