Uptake of COVID-19 Vaccination Among Frontline Workers in California State Prisons

Key Points Question In California prisons, what proportion of prison staff who have direct contact with residents are unvaccinated, and what are their characteristics? Findings In this cohort study of 23 472 custody staff and 7617 health care staff in California state prisons, 14 317 custody staff (61%) and 2819 (36%) health care staff remained unvaccinated through June 30, 2021, despite widespread vaccine availability. Unvaccinated staff were younger and more likely to have had COVID-19; they were also more likely to work alongside other unvaccinated staff and live in communities with relatively low rates of vaccination. Meaning The study results suggest that low vaccination rates among prison staff pose continuing risks.

To limit our analytic dataset to correctional staff with the greatest likelihood of direct contact with incarcerated residents, we imposed several inclusion/exclusion criteria. Our analytic cohort was drawn from staff in 33 of CDCRs 35 prisons. Two of CDCR's 35 prisons were excluded from the analysis due to missing or incomplete staff data at those prisons at the time of analysis. We included only staff designated as custody or healthcare workers in one of the 33 included prisons (~63% of total CDCR staff). Other possible designations are contractor, education, and operations. Among the healthcare and custody staff, we included people designated as "direct-care" staff, meaning that they were classified by CDCR as having regular in-person contact with residents (~98% of custody and healthcare staff). We excluded staff who were missing data for any of our covariates (<1%), with the exception of missing race, which was coded under "Other/Unknown". We excluded staff who did not work at least five shifts between April 2021 and June 2021 (the period after vaccination was available to any staff person; 275 healthcare and 181 custody staff were excluded). Some prison staff (638 in our study sample) worked across multiple prisons, they entered the analysis assigned to their main institution. There were 3 staff who were appointed as custody and healthcare at different times. These were excluded from our analysis. There were 171 staff who had a record of being vaccinated in the community before their first recorded shift at CDCR. These were excluded from the analysis. Importantly, the fraction of co-workers vaccinated variable was constructed based on the entire custody or healthcare direct-care staff, prior to exclusions.

Outcome
Remained unvaccinated: CDCR staff were offered vaccines at their place of work beginning on December 22, 2020. Vaccine uptake was voluntary through the end of our study period. All vaccines administered on-site are recorded in CDCR data (and reflected in our dataset), and vaccines administered in the community were obtained from the California Immunization Registry, provided staff members gave their consent. The outcome reflects that the staff member remained unvaccinated (no doses of vaccine administered) through June 30, 2021.
Race group. We grouped race by race and ethnicity information provided by CDCR. Race categories and their components are as follows: Asian/PI (Asian or Filipino or Pacific Islander); Black (Black); Hispanic (Hispanic + Cuban + Mexican); White (White); Other/Unknown (includes American Indian or Alaskan Native (~0.4% of total sample) and other (~1% of total) and unknown or missing (~17% of total).
Gender: Gender is based on self-report of gender.

History of Covid-19:
The staff data included comprehensive information on PCR and antigen testing for SARS-CoV-2 infection, with the first recorded test on March 18, 2020. Testing was voluntary and/or mandatory and occurred at varying frequency over time. Because testing was infrequent through early 2020 (see counts below), we may underestimate the proportion of staff with any history of Covid-19. Despite this, we feel confident in our estimate of fraction of staff without any positive test at the end of June 2021 because of ramped up testing starting in mid-2020 (when non-fully-vaccinated staff have been tested twice-weekly and any staff who report symptoms or work in a location with an outbreak have also been tested) in combination of the timing of most community and prison outbreaks from mid-2020 and onward. Unvaccinated in zip code: Fraction of unvaccinated adults in staff home-zip code is based on the cumulative percentage of the 20-64 year-old population in a given zip-code who received 1+ doses of vaccination by June 1, 2020. Data are from CDPH (see reference in main text). We used the last known zip code of staff members to create this measure.

Month
Unvaccinated in work cohort: To build a cross-sectional measure of peer vaccine take-up, we counted, for each staff-shift, all coworkers on each prison-shift-day and all coworkers who were unvaccinated on each prison-shift-day. The individual-level measure where N is the total number of shifts worked by an individual staff person during the study period, Us is the number of unvaccinated workers on a particular prison-shift-day and Ts is the total number of workers on a particular prison-shift-day. Note, we subtract the individual for whom the measure is calculated in the counts of unvaccinated and of total workers to avoid a reflection issue.
See Appendix Table 5 for a sensitivity analysis limiting the analysis to prisons with the greatest within-institution inter-quartile range (as a measure of spread).

Shift variables:
We created a categorical variable denoting which shift a staff person worked most often (based on raw count of shifts worked over the study period): day, night, or swing shift. We controlled for the total number of shifts worked during the study period, the mean shifts per week worked (weeks with zero shifts were not counted in the mean), and their interaction.

Prison Fixed effects:
We included prison-level fixed effects to control for stable heterogeneity between prisons.

IRB and ethics approval
The study was approved by the institutional review board (IRB) at Stanford University (protocol numbers IRB-55835, IRB-55671).
The IRB approval of the study included a waiver of consent, on the basis that CDCR provided the Stanford research team with a limited data set without direct identifiers, the data had been collected for operational purposes, and the study could not practicably be carried out otherwise. Similar approval conditions were met for California Department of Public Health data.

Sensitivity Analyses (SA)
We ran three sensitivity analyses to explore questions related to the fraction of co-workers unvaccinated and the history of Covid-19 variables, as follows. These were run on the custody staff group. Predicted probabilities and 95% confidence intervals are presented in Appendix Table 5.

SA 1: Within prison dispersion in the fraction of co-workers vaccinated
As noted in the text, our measure of co-worker peer influence is crude. To estimate the relationship between co-worker vaccination uptake and individual staff decision to remain unvaccinated, we would ideally see substantial variation across individual staff workers in the rate of vaccination of their co-worker cohorts within a given prison -a signal that there are differences in cohort-preferences for vaccination. If there is not substantial variation, it could be possible that exogenous shocks over time (not accounted for by the prison fixed effects) lead to heterogeneity in cohort vaccination rates. (Note: because the patterns of vaccination uptake are strikingly similar across prisons [see Appendix Figures 1 and 2] this concern about variation in the timing of vaccination due to such shocks is substantially reduced.) To explore this further, we examined the within-prison dispersion in the fraction of (custody staff) co-worker cohorts vaccinated and repeated our main regression analysis on the sample after excluding those prisons with the least variation. Within prisons, the mean fraction of unvaccinated custody workers was between 34% and 86%, the range (highest value -lowest value) was between 4 percentage points and 24 percentage points, and the interquartile range (IQR) between 1 and 8 percentage points. We re-ran our multivariable analysis limiting the sample to those prisons with a wider distribution of values (IQR ≥ median; 17 of 33 prisons included; N = 12,221) to examine whether individuals working in prisons with a wider dispersion were more or less likely to remain unvaccinated if their co-workers we less likely to be vaccinated. The results were robust to our main analysis, custody staff were six percentage points more likely to remain unvaccinated if their co-worker cohort was at the 75 th percentile of being unvaccinated compared to working with a cohort at the 25 th percentile (compared to being 5 percentage points more likely in our main analysis). See Appendix Table 5 for results.

SA 2: Interaction of main shift worked with fraction of co-workers vaccinated
It may be the case that peer interactions may be different for night shift versus day or swing shift and hence the relationship between co-worker vaccination patterns and staff remaining unvaccinated may be different by shift. To assess the possibility that main-shiftworked is an effect modifier (e.g., the correct model specification interacts the shift-variable with the fraction of co-worker vaccinated variable) in our model, we repeated the multivariable probit regression as in the main analysis expanded to include this interaction term. The coefficient on the interaction term was not significantly different than zero and the predicted probabilities were consistent with those resulting from our main specification. See Appendix Table 5 for results.

SA 3: History of Covid-19 by time of test
Because there is some concern that vaccination uptake is related to the timing of a previously positive SARS-CoV-2 test (e.g., an individual may be advised to wait 90 days from last positive test to receive a dose of vaccine), we divided our History of Covid-19 variable into time periods as follows. The reference group (as in the main specification) is "no history of 052), and the other categories are: last positive test prior to December 22, 2020 (N = 4,619); last positive test between December 22, 2020 andMarch 15, 2021 (N = 1,423); last positive test after March 15, 2021 (N = 378). We chose the time periods based on the idea that all those in the first and second groups (e.g., before March 15, 2021) would have time to become vaccinated within our study period even after a 90-day post infection window, and those in the first group would have time to be become vaccinated during the initial vaccine push evident in Figure 1 in the main text. We find that, for residents with last positive SARS-CoV-2 test prior to December 22, they are 5 percentage points more likely to remain unvaccinated compared to those with no prior history of Covid-19. For those with last positive SARS-CoV-2 test in the first few months of the study period, they are 13 percentage points more likely to remain unvaccinated. We can compare this to an estimate of those with prior infection in the main specification being 8-percentage points more likely to remain unvaccinated than those with no prior infection. Thus, when focusing on those with a history of prior Covid-19 more than 90 days before the end of our study period, history of prior Covid-19 remains a significant predictor of remaining unvaccinated.  8  289  121  6  11  799  42  13  271  34  11  12  822  115  20  641  125  28  13  1151  75  9  316  69  17  14  736  149  7  115  64  5  15  687  69  21  133  46  17  16  382  81  8  60  23  6  17  460  88  15  115  50  12  18  530  87  13  110  42  7  19  626  63  23  110  27  17  20  570  104  7  76  28  7  21  912  52  6  183  45  11  22  762  156  11  257  91  10  23  810  103  15  298  96  18  24  540  23  15  65  15  12  25  752  66  7  115  24  6  26  938  136  14  347  96  10  27  854  114  19  318  85  19  28  938  67  8  292  70  19  29  576  155  26  70 30  666  106  21  137  54  11  31  991  254  35  256  105  15  32  868  127  26  299  101  23  33  490  63  9  178  44  10 All counts are derived from the analytic sample used in the multivariable analysis.