Modeling Shifts in Community Corrections Populations following COVID-19: Evidence from a Midwest Metropolitan Area

ABSTRACT

The onset of the COVID-19 pandemic resulted in short-term strategy shifts by criminal justice agencies, including population reductions and early release in institutional settings and restricting in-person check-ins in community corrections, among many others. Years into the pandemic, the impact of how local criminal justice agencies navigated these changes in practice is still coming into focus. This paper models shifts in community corrections populations in a Midwest US metropolitan community corrections agency following the onset of the COVID-19 pandemic. Using monthly population data from early 2017 to mid-2022, we employ Bayesian interrupted time-series models to describe changes in probation, parole, and supervision absconder populations over this period. The results suggest that both probation and parole populations, as well as absconders, were in decline leading up to March 2020, but then exhibited varying trends following the pandemic. Probation populations decreased markedly, while parole absconding increased.

KEYWORDS:

• COVID-19
• parole
• probation
• absconding

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Replication materials for the analyses presented in this manuscript are available from: https://www.openicpsr.org/openicpsr/project/179841/version/V1/view.

Notes

1. We considered additional options for the residual correlation structure, including a second-order autoregressive structure, and the possibility of no systematic correlation between monthly observations. Model comparison via the LOO cross validation score (see Multi-model comparison strategy section) suggested the first-order structure was preferable. Further, we also considered possibilities for error terms between the active supervision and absconder models to correlate. This may occur because individuals leaving one population count may appear in the other (e.g., as individuals under active supervision abscond, they produce a drop in the active supervision population and a corresponding increase in the absconder population). We explored modeling the active supervision and absconder populations for parole and probation as simultaneous equations with a cross-equation residual correlation parameter. The posterior estimates for these correlations were highly uncertain and straddled zero, and produced no substantive differences in conclusions over the models presented here. Finally, autocorrelation functions (ACF) and partial autocorrelation functions (PACF) suggested that seasonality controls were not necessary.

2. Alternate likelihoods were considered as well, including Gaussian using the raw population count as the outcome, log-normal, Poisson, and negative binomial. The model using a Gaussian likelihood on the log-transformed outcome produced the best fit to the data per comparison of LOO scores.

3. Weakly informative priors, designed for conservative parameter regularization, are applied to all parameters except the autoregression terms ($\phi$). Markov chains for initial models using less stringent priors (e.g., Normal(0, 0.5)) had difficulty exploring the posterior parameter space, producing low effective sample sizes, poor chain convergence diagnostics, and oddly shaped bimodal posterior distributions. A relatively tight prior of Normal(0, 0.1) on the autoregression parameter was sufficient to resolve this issue for the parolee models, while a (still tight, but) more relaxed prior of Normal(0, 0.25) was sufficient for the probationer models.

4. We also considered the possibility of more complex functional forms for time, including cubic parameterizations. On inspection, these models tended to produce dramatic overfitting of the pre-COVID trend, to the extent that they generated ludicrous predictions for the population counts had the pre-COVID trend continued into the post-COVID period (e.g., millions of monthly probation absconders by June 2022). As such, we opted to scope our comparisons to the more reasonable model set presented here.

5. A benefit to the LOO is that it provides diagnostics in the form of Pareto k estimates, where values greater than 0.7 suggest problematic data points which complicate comparisons of predictive accuracy across models. In these instances, candidate models can be refit leaving these specific observations out to exactly compute the predictive accuracy (rather than approximating them; Vehtari et al., 2017). On this basis, the models evaluated in Table 2 were refit 1 to 3 times each to obtain their corrected LOO scores and stacking weights.

6. As the data were received, there were approximately 1,100 unique conviction offense charges among the 3.3 million supervisee-month observations. To the best of our ability, these offenses were categorized into the broad typologies of persons offenses, property offenses, controlled substances, motor vehicle offenses, weapons offenses, and other. We were able to categorize 99.95% of the observations into one of these categories.

Additional information

Funding

This project was supported by Grant No. [2018-86-CX-K033] awarded by the Bureau of Justice Statistics, Office of Justice Programs, and U.S. Department of Justice. Points of view in this document are those of the author and do not necessarily represent the official position or policies of the US Department of Justice.