Abstract

Improving outcomes for individuals under correctional supervision is a core focus for criminal justice researchers and professionals. Here we analyze a cohort of over 500 females on parole in Utah between 2013 and 2017, using time-to-event analysis. We find that risk factors for females on parole are different when the outcome variable denotes a technical violation versus a new crime, with 45 percent of females being returned to prison at the end of the 1-year follow-up time. Utah’s high recidivism rate coupled with the cost of female incarceration merits further research on this important public policy issue.


Definitions

In many US states, an individual sentenced to prison may be granted parole to serve their remaining sentence in the community. In the state of Utah, a parolee can be sent back to prison due to one or more technical violations or a new crime, with the latter considered a greater public safety concern. A return, or revocation due to a technical violation does not involve a new crime but means that the individual did not comply with his or her supervision terms. These violations range from for example, not showing up to an appointment to absconding from parole. This study explores survival time for females on parole in Utah, using time-to-event analysis. We examine two distinct outcomes: A revocation from parole to prison due to one or more technical violations and a revocation due to one or more new crimes.


Data & Method

Data

Records of 508 unique1 females serving time on parole between 2013 and 2017 were obtained from the Utah Department of Corrections’ (UDOC) database O-Track, each followed for 1-year after their respective parole start date. As informed by the national literature,2 demographics including age, race-ethnicity, and region (urban vs rural) of the parole, serve as important covariates. Additionally, criminal justice variables, including offense type (a binary classification that = 1 if it was a violent offense), offense severity (Felony 1,2, and 3 where 1 is the most severe), and the parolee’s risk to re-offend, defined by the Level of Service: Risk-Need-Responsivity (LS:RNR) instrument3 were included as important covariates.

Figure 1 shows the actual revocation rates by type of return. Overall, 45 percent of the females were returned to prison on either a technical violation or a new crime by the end of the 1-year follow-up time, with 38 percent being returned on technical violations and 7 percent being revoked on new crimes. The mean time spent on parole (within the 1 year follow-up time) was 265 days. Table 1 shows summary statistics and is available in the Appendix.


**Figure 1:** One-Year Return to Prison Rate by Type

Figure 1: One-Year Return to Prison Rate by Type


Method

Survival Analysis - Overview

While most prominently used in medical research, survival analysis, or time-to-event analysis have shown to have useful properties in understanding performance for individuals placed under correctional supervision. One of the reasons is its censoring properties, where individual’s sentence length often differs due to reasons unknown at supervision start. The Cox Proportional Hazard Model (Cox model) is the most commonly used model in the survival analysis literature.

A compelling feature of the Cox model and unlike the Kaplan-Meier Estimator, is that it allows the analyst to control for variables believed to have an effect on the outcome, thereby reducing possible differences across groups. The Cox model estimates the probability of survival at time t, given that the individual has already survived until that time, expressed as the Hazard Ratio. A Hazard Ratio (HR) of:

  • 1 indicates no difference between groups;
  • below 1 denotes an improved survival probability, and
  • above 1 indicates a reduced survival probability.


Empirical Results

Technical Violations

Figure 2 show the Cox proportional hazard model’s hazard ratios with a 95 percent confidence interval for the outcome associated with one or more technical violations.4 As seen, high risk females show a worsen survival probability with a hazard ratio of 1.8 (at alpha= 5%). This implies that females that are high to intensive risk to re-offend are nearly twice as likely of being revoked to prison than low to moderate risk females. In contrast, the variable denoting no prior parole starts shows an improved survival probability and has a hazard ratio of 0.54. The other factor associated with improved survival probability include increases in age (HR=0.97).


**Figure 2:** Hazard Ratios with error-bands \vspace{20pt}

Figure 2: Hazard Ratios with error-bands


New Crimes

Figure 3 show the Cox proportional hazard model’s hazard ratios with a 95 percent confidence interval for the outcome associated with one or more new crimes. As seen, the variable denoting offense severity5 (HR=3.6) and the region that the female was paroled to (HR=0.44) are significant at the 5% testing level and are associated with an improved survival probability. This implies that females convicted of a Felony 1 & 2 offense are less likely to be revoked on a new crime in comparison to those convicted of a Felony 3 offense. The rather large error bands around this Hazard Ratio does however, indicate a high degree of uncertainty. Other covariates, including, the variable denoting risk to reoffend lack statistical support at the 5% testing level.

**Figure 3:** Hazard Ratios with error-bands \vspace{20pt}

Figure 3: Hazard Ratios with error-bands


Survival Curves

Technical Violations

Figures 4-6 illustrate survival curves with a 95 percent confidence interval, overall, and by selected group characteristics when the outcome concerns a revocation due to technical violations.6 As expected, survival curves are the steepest at supervision start and become more gradual with time.

The difference in survival probability across groups is clearly seen amongst the selected covariates. For example, the slope for low/moderate risk parolees is significantly less steep than for high to intensive risk parolees starting at around 150 days after supervision start and continues throughout the remaining follow-up time. Similarly, the difference in survival is also seen amongst females with no prior parole starts and those who have had multiple prior parole starts, in which shows a significantly improved survival probability starting at a similar number of days after supervision start while continuing throughout the remaining follow-up time.


**Figure 4:** Cox Proportional Hazard Survival Curve

Figure 4: Cox Proportional Hazard Survival Curve


**Figure 5:** Cox Proportional Hazard Survival Curve by Risk Level

Figure 5: Cox Proportional Hazard Survival Curve by Risk Level


**Figure 6:** Cox Proportional Hazard Survival Curve by Number of Prior Parole Starts

Figure 6: Cox Proportional Hazard Survival Curve by Number of Prior Parole Starts


New Crimes

Figures 7-8 illustrate survival curves, overall and by offense severity with a 95 percent confidence interval when the outcome concerns a revocation due to a new crime.7 In contrast to Figures 4-6, the slope of the survival curve appears flatter at parole start and then increases with time (at around the 200 day mark). As seen in Figure 8, due to high degrees of uncertainty around the estimate, the confidence bands are close to overlapping throughout the 1-year follow-up time.


**Figure 7:** Cox Proportional Hazard Survival Curve

Figure 7: Cox Proportional Hazard Survival Curve


**Figure 8:** Cox Proportional Hazard Survival Curve by Offense Severity

Figure 8: Cox Proportional Hazard Survival Curve by Offense Severity


Moving Forward

Study Limitations

Utah’s current data infrastructure limits the ability to fully account for short times that a parolee may have spent in jail time across the selected time period. Additionally, the small share of females being revoked to prison on a new crime may affect traditional statistical significance tests and hence findings from this analysis should be seen as preliminary. Other factors, including, an individual’s personal motivation remains statistical challenges to detangle when estimating effects in quasi-experimental designs.


Future Research

Findings from this study support similar conclusions drawn from previous research on Utah’s overall parolee population. In moving forward, research may look at additional definitions and statistical methods to evaluate recidivism for female parolees. Furthermore, while the relationship between technical violations and the prevention of new crimes is complex, exploring the nature and severity of technical violations and new crimes that occurred can further provide meaningful information that can aid in the development of action-oriented policy responses. Indeed, the findings from this study highlight the continuing need to conduct thoughtful analyses to identify factors linked to barrier and successes for females on parole.


Appendix

Table 1: Summary Statistics

mean sd min max
recid_nc_1year 0.07 0.25 0.00 1.00
recid_tech_1year 0.38 0.49 0.00 1.00
days_on_parole 264.54 126.13 1.00 365.00
age 34.43 8.02 19.95 79.43
minority 0.26 0.44 0.00 1.00
urban 0.73 0.44 0.00 1.00
high_risk 0.74 0.44 0.00 1.00
severity 1.34 0.53 1.00 3.00
first_parole 0.21 0.41 0.00 1.00


Further Readings


Cox, David R (1972). Regression Models and Life-Tables. Journal of the Royal Statistical Society, Series B. 34 (2): 187–220.

Fox J. & Weisberg S. (2018). Cox Proportional-Hazards Regression for Survival Data in R. Available at https://socialsciences.mcmaster.ca/jfox/Books/Companion/appendices/Appendix-Cox-Regression.pdf

Goel MK, Khanna P, & Kishore J. (2010). Understanding survival analysis: Kaplan-Meier estimate. Int J Ayurveda Res 2010;1:274–8. Available at https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3059453/

King, R. S., & Elderbroom, B. (2014). Improving recidivism as a performance measure. Washington, DC: Urban Institute. Available at https://www.urban.org/research/publication/improving-recidivism-performance-measure

Stone, R., Morash, M., Goodson, M., Smith, S., & Cobbina, J. (2018). Women on Parole, Identity Processes, and Primary Desistance. Feminist Criminology, 13(4), 382–403.

The Council of State Government (2019). Confined and Costly: How Supervision Violations Are Filling Prisons and Burdening Budgets. Available at https://csgjusticecenter.org/confinedandcostly/

Examining Parole Revocation Patterns (2018). Utah Commission on Criminal and Juvenile Justice. Available here https://justice.Utah.gov/JRI/Documents/Examining_Parole_Revocation_Patterns.pdf

Yesberg, J. A., Scanlan, J. M., Hanby, L. J., Serin, R. C., & Polaschek, D. L. (2015). Predicting women’s recidivism: Validating a dynamic community-based ‘gender-neutral’tool. Probation Journal, 62(1), 33-48.

Yukhnenko, D., Blackwood, N., & Fazel, S. (2019). Risk factors for recidivism in individuals receiving community sentences: a systematic review and meta-analysis. CNS spectrums, 1-12.

  1. A unique parole start is defined as a combination of a unique identification number (O-Track number) and parole start date and only includes individuals with complete information on the covariates used here.↩

  2. See Yukhnenko, Blackwood & Fazel (2019) for a review of predictors of recidivism.↩

  3. The LS:RNR is a validated risk assessment instrument that categorizes offenders as low, moderate, high and intensive risk to re-offend. UDOC implemented this version of the Level of Service in 2015. The risk-to-reoffend variable was reclassified into a binary variable (low/moderate versus high/intensive risk to re-offend) due to concerns about the comparability of different risk assessment versions through time.↩

  4. The performance of the Cox proportional hazard model is evaluated via the concordance statistic, which calculates how well the observed response agrees with a given explanatory variable. The concordance statistic lies between 0 and 1. A value of 0 implies no concordance, 0.5 random predictions, while a value of 1 denotes perfect concordance. Both models showed a concordance statistic of > 0.5.↩

  5. The variable denoting offense severity was made into a binary variable due to low number of observations.↩

  6. Survival curves were generated holding each covariate fixed at their mean values, with the variables denoting offense severity evaluated at Felony 3.↩

  7. Survival curves were generated holding each covariate fixed at their mean values.↩