Figure 3.

Social connectivity network inferred from the WiFi co-location data identifies behavioral trends that correlate with case counts

(A) Left: distributions of daily contacts for students who tested positive (orange) versus those who remained negative (gray) over the course of each semester. Right: distributions of average exposure time per contact, in minutes, for cases (orange) versus those who remained negative (gray). p values via Wilcoxon rank sum test.

(B) Visual representation of the network metric attribute assortativity (AA). Network scenarios where the AA coefficient is equal to 1, 0, or −1 are depicted. Positive AA values indicate a higher propensity for within-group interactions, while negative values indicate a higher propensity for between-group interactions.

(C) Comparison of per-semester AA for individuals within a 10-day window of a positive test (i.e., pre-positives) versus those who never tested positive (i.e., negatives). Ninety-five percent confidence intervals (Cis) (blue) were calculated by permuting pre-positive and negative labels within the proximity network 40 times per day. The AA of the proximity network (black) was above the upper bound of the CI for 69.4% (66/95 days) of the Fall 2020 semester, implying significance at p < 0.025. Results were consistent for Spring 2021 (Figure S8).

(D and E) Relationship between smoothed case counts and smoothed AA for Fall 2020 (D) and Spring 2021 (E). Smoothing via the Savitzky-Golay filter (window length = 17, polynomial order = 4).