Stanford’s institutional review board granted approval (IRB-55835).
Data and Variables
CDCR provided data on all incarcerated people aged ≥18 years residing in its prisons during the period March 1 through October 10, 2020. The data were provided at the person-day-level, which allowed daily tracking of changes in any time-varying information until residents’ release or death.
The data included variables indicating residents’ demographic (sex, age, and race or ethnicity) and health characteristics; location; participation in prison labor, education, and other activities; and COVID-19 testing history. The data contained each resident’s security level (1 [lowest] to 4 [highest]) which determines housing locations and eligibility for work and other activities (16). Locational information specified the room in which each resident spent the night. Rooms were defined as discrete spaces, at least partially enclosed by solid walls, and were classified as cells or dormitories of varying sizes.
The health information included indicators for diagnosed medical conditions and COVID-19 risk score. This score, developed by CDCR, is an integer-based estimate of each resident’s probability of severe health outcomes following COVID-19 infection. Scores correspond to the presence of demographic and clinical characteristics identified in the literature as risk factors for severe COVID-19-related illness, and CDCR considers scores ≥3 to indicate high-risk (Table S1).
Testing information included dates and results. CDCR has been testing residents using real-time PCR and antigen tests since April 2020. Testing expanded during the study period, eventually employing both reactive mass testing and periodic surveillance testing. Prisons experiencing large outbreaks tested residents at particularly high rates (Figure S1). By Fall 2020, all prisons were testing 5–25% of residents every 2 weeks.
We created variables to describe risk factors for COVID-19 exposure and transmission, based on residents’ housing situation and their participation in out-of-room activities. With respect to housing, we calculated the daily number of residents housed in each room (square footage of rooms was unavailable). This count variable had a bimodal distribution, with many residents living alone or with one other resident, or in substantially larger rooms (≥10 residents), and relatively few in between. For some analyses, we dichotomized this variable, distinguishing residents in “cells” (1–2 occupants) and “dormitories” (≥3 occupants); this aligned with CDCR’s conception of the major division in room sizes.
With respect to activities, the resident-day-level data included information on residents’ out-of-room participation in labor (e.g., janitorial), education (e.g., high school classes), and other activities (e.g., religious services) (17, 18). From April 2020, educational activities were confined to residents’ rooms. Consequently, we focused on labor and other out-of-room non-educational activities excluding recreation and meals which were not comprehensively tracked. We specified variables indicating whether each resident participated in each activity type and variables indicating whether each resident or a roommate did so in the previous 2 weeks.
Finally, we classified prisons into 5 categories based on the predominant resident security levels, housing configurations, and CDCR advice: reception centers, medical prisons, low security and general population prisons, high security prisons, and mixed security and medium security prisons (details provided in Table S2).
We calculated changes over the study period in the size and composition of the incarcerated population, in housing, and in participation in out-of-room activities (samples for each analysis shown in Figure S2).
We used person-level survival analysis to estimate the association between room occupancy and labor activities, respectively, and rates of COVID-19 infection. We focused on sustained within-prison transmission, therefore limiting the analysis to prisons with outbreaks involving substantial resident-to-resident spread. We defined prisons with outbreaks as those having ≥50 cumulative cases during the study period and ≥10 incident cases detected on at least one day in that period. We defined an outbreak’s start date as 14 days prior to the first day with ≥10 incident cases.
We specified several additional prison-level and resident-level eligibility criteria for the survival analysis (Figure S3). Briefly, to allow ≥90 days follow-up, we excluded prisons (n=7) with outbreaks that began after July 12, 2020; to minimize confounding, we excluded one prison with testing rates that differed substantially between cells and dormitories; and we excluded prisons (n=3) whose outbreaks were seeded by mass introduction of cases (e.g., San Quentin), because their epidemic growth may have been atypical. Within eligible prisons, we included residents present on the day the outbreak began who were tested for COVID-19 at least once during the 90-day period.
The observation period for eligible residents’ observation ran from the start of their prison’s outbreak until the sample collection date of their first positive test or their last negative test. Release or transfer to another prison were also censoring events.
We fit a multivariable Cox proportional hazard regression model to estimate the associations of interest. The outcome variable indicated the sample collection date for the first positive COVID-19 test result among residents who had a positive test. The main exposure variables were room occupancy at the outbreak start and room-level labor participation during the 14-day period prior to the outbreak start. We also included prison fixed effects. We assessed the appropriateness of the proportional hazards assumption by inspecting plots of Schoenfeld residuals.
We conducted sensitivity analyses. Because there may be systematic differences in how residents with higher COVID-19 risk scores or higher security levels mix with other residents, we added baseline values of these covariates. We varied the required follow-up period for prison inclusion. We allowed the observation period for residents who did not test positive, exit the prison, or die to extend to the end of the study period, regardless of their last negative test date. We estimated the model clustering standard errors at the room and prison levels.
Role of the Funding Sources
This research was supported by the Horowitz Family Foundation, the National Institute on Drug Abuse (R37-DA15612), the National Science Foundation (DGE-1656518), and the Open Society Foundations. The funders had no role in the study’s design, conduct, or reporting, or in the publication decision.