Association of Jail Decarceration and Anticontagion Policies With COVID-19 Case Growth Rates in US Counties

Key Points Question Were jail decarceration and government implementation of anticontagion policies associated with the spread of SARS-CoV-2 in US counties? Findings In this cohort study of 1605 counties in panel regression models, an estimated 80% reduction in US jail populations would have been associated with a 2% reduction in daily COVID-19 case growth rates, with considerably greater COVID-19 reductions in counties with above-median population density and above-median proportion of Black residents. In analyses of anticontagion policies, nursing home visitation bans were associated with a 7.3% reduction in COVID-19 growth rates, followed by school closures (4.3%), mask mandates (2.5%), and prison visitation bans (1.2%). Meaning The findings of this study suggest that anticontagion policies, including jail decarceration to minimize carceral outbreaks and their spillover to surrounding communities, appear to be necessary for epidemic control, public health, and mitigation of racial health disparities.


eAppendix 1. Robustness Checks
Some policies in our dataset do not feature dates of change after August 31, introducing concern for possible measurement error. We therefore conduct an additional analysis that restricts to an end-date of August 31 in order to investigate whether, when controlling for anti-contagion policies without this possible measurement error, daily jail population remains correlated with COVID-19 daily growth rates both quadratically and linearly. The results of this analysis, which can be found in eTable 3, are consistent with our broader analysis.
As an additional analysis of anti-contagion policies and as a robustness check to evaluate possible selection bias in our analysis sample, we drop the jail variables in order to evaluate anti-contagion policies for all counties without restricting to those for which we have jail population data. This increases our number of observations from 319,084 (when the sample is restricted to only those counties for which we have jail data) to 725,407 (when unrestricted by jail data availability). The results of this analysis (see eTable 4) are broadly consistent with the results reported below for our analysis restricted to counties for which we have jail data; the estimates in Column 2 are roughly a factor of 0.8 of those in Column 1, with the exception of two policies in which the difference is about the same magnitude but in the other direction. This mitigates concerns regarding selection bias. An additional way to assuage possible concerns for outliers in our sample is to drop one stateweek at a time, as presented in eFigure 2. This figure, in combination with Figure 1, provide a robustness check for possible outliers that might account for our results. eTable 5 repeats the regressions from Table 2 while using a one-week lag between intervention and COVID-19 growth rate rather than a two-week lag. The patterns observed with this one-week lag are larger by a factor of 1.4 for jail population; the remaining coefficients remain numerically very similar to the patterns reported in Table 2. eFigure 3 presents binned scatter plots for each of four lags between jail population and COVID-19 growth rate.
As seen in eTable 6, the association between daily jail population and daily COVID-19 growth rate two weeks later persists even when controlling for a two-week lag of COVID-19 growth--that is, when controlling for the association of earlier COVID-19 growth rates on subsequent COVID-19 growth rates. Jail cycling is expected to function as multiplier of incoming COVID-19 cases from the community that then disseminate back to community contexts in a positive feedback loop that compounds community COVID-19 growth rates. 18 Consistent with this expectation, controlling for this lagged COVID-19 growth diminishes the magnitude of the observed coefficients for the effects of jail population. The quadratic relationship nonetheless remains statistically significant (p < 0.001) with this additional control.
In eFigure 2, to assuage concerns of spatial or temporal outliers driving our main findings, we drop one state-week (i.e., one week of one state) at a time and present the coefficient estimates corresponding to the linear model in Table 2, column 2. All values were above conventional significance levels.
Figures S4A-B also display the relationship separately by time period and intensity of SARS-CoV-2 outbreak in a county. The patterns we observe are also much larger when caseloads are higher.
Accordingly, the patterns we observe are much larger from April to August 2020, prior temporary decline in US COVID-19 cases in the autumn months before their rapid increases beginning in November.
Notably the patterns are large prior to the adoption of any anti-contagion policy in a given county (i.e., when jail cycling policies cannot be associated with other anti-contagion policy measures). eTable 7 presents a comparison of demographic characteristics between the counties included in our analytical sample (in-sample) and the counties that were not included owing to the lack of jail population data (out-sample). We performed a two-sample t-test between the two sets of counties on selected demographic variables. The analysis sample is slightly biased towards minority, poorer, and younger populations.

eAppendix 2. Limitations
We are limited by the non-universal coverage of data on county-level policy interventions. Our sources allow us to observe 1,144 counties--approximately 72% of all counties in our analysis sample--with at least one policy variable in addition to state-wide CUSP variables. More specifically, we use countyspecific, non-CUSP-reported data on stay-at-home policies for 554 counties; restaurant closures for 332 counties; non-essential business closures for 591 counties; and school closures for 481 counties. Our data do not include intra-state county-level variation for other variables such as mask mandates, gym closures, or prohibitions on nursing home visitation. eTable 1. Estimated relationships between COVID-19 daily growth rate and log daily jail population and anti-contagion policies in multivariate regression analysis with county fixed effects.  The table compares, as a robustness check, estimates from regressions using a cubic polynomial term on jail population variable with that of the quadratic and linear regressions which we tabulated in Table T2.

Independent variable
We can see that the COVID-19 case growth rate, upon adding a cubic polynomial term also finds a significant positive association with jail population. eTable 2. Estimated relationships between growth rate in COVID-19 cases and log imputed daily jail cycling and all intervention policies in multivariate regression analysis with county fixed effects.  This table shows that when restricting the data to August 31 (after which some of the policy variables are not updated in our data sources as inactive if they have been lifted during this period), even when controlling--without measurement error in data after August 31--for anti-contagion policies, daily jail population correlates with COVID-19 daily growth rates both quadratically in column 1 and linearly in column 2. The association between jail population and COVID-19 daily growth rates appears to grow quadratically.

Independent variable
eTable 4. Estimated relationships between COVID-19 daily growth rate and all the anti-contagion policies in multivariate regression analysis with county fixed effects. Nursing eTable 4 shows that the association between policies and COVID-19 growth remain similar when, in

Independent variable
Column 1, we remove daily jail population as a control but keep the sample of counties for which we have data on jail population, and in Column 2, when include all US counties without restricting to eTable 5. Estimated relationships between growth rate in COVID-19 cases and log jail population and all intervention policies in multivariate regression analysis with county fixed effects and one-week lag between intervention and COVID-19 growth rate.