Does Bail Reform Increase Crime in New York State: Evidence from Interrupted Time-Series Analyses and Synthetic Control Methods

Abstract

In 2019, New York State passed bail reform legislation that limited the use of money bail and expanded pretrial release. The bail reform law took effect on January 1, 2020. We evaluated the effect of this law on crime rates in New York State. Interrupted time series analyses (ITSA) were used to examine whether the bail reform was significantly associated with a crime increase. When a significant association was detected, we examined whether this relationship was causal. The causal relationship was tested using synthetic control methods (SCM). We found that the rates of murder, larceny, and motor vehicle theft increased after the bail reform. We then employed SCM to create a comparison group to control for potential confounders like the pandemic. By comparing New York State with its synthetic control, we found the increases in the murder rate, larceny rate, and motor vehicle theft rate were not statistically significant. Our findings suggest that the effect of bail reform on crime rate increases is negligible.

Keywords:

• Bail reform
• crime rates
• interrupted time series analyses
• synthetic control methods
• quasi-experiment

Notes

1 Misdemeanors that are bail eligible under the 2019 bail law: PL 130 (misdemeanor sex offenses) and criminal contempt in domestic violence cases.

2 Nonviolent felonies that are bail eligible under the 2019 bail law: witness intimidation, witness tampering, nonviolent felony sex offenses, incest, money laundering for terrorism, sexual performance of a child. Among lower court arraignments from 2019 to 2021 (N = 559,817), about 3% are bail-eligible nonviolent felonies and misdemeanors under the 2019 bail law (see Table B1 in Appendix B).

3 The number 447 is obtained by differencing the numbers of re-offenses before and after the order: 447 = 24,504*17%–20,435*18.2%. The numbers of pretrial releasees and reoffending rates before and after the order were reported by the Office of the Chief Judge in Cook County Circuit Court (2019). 141 additional crimes and 34 additional violent crimes were calculated similarly: 141 = 9,133*17.8%–8,891*16.7%; 34 = 9,133*3.2%–8,891*2.9%. The numbers of pretrial releasees and reoffending rates before and after the order were reported by the Stemen and Olson study (2020).

4 The New York State Office of Court Administration (OCA) and DCJS only released pretrial rearrest data from 2020 to 2021(New York State Unified Court System, 2022). Statewide pretrial rearrest data in 2019 are not available for comparison.

5 The pretrial release rate in New York City increased from 2019 to 2020. But the releasee population in New York City did not increase because the number of arraignments decreased greatly from 2019 to 2020. There were 160,920 total arraignments in 2019 and 83,868 total arraignments in 2020 (Lu et al., 2022).

6 Releasee population and reoffending counts were obtained by summing up monthly reoffending data shown in the “People in Community with Pending NYC Case” dashboard (New York City Criminal Justice Agency, 2022), filtered by selecting violent felony as most series charge and slicing by most serious charge of new offense.

7 New violent felony counts were obtained by summing up monthly reoffending data shown in the “People in Community with Pending NYC Case” dashboard (New York City Criminal Justice Agency, 2022), filtered by selecting violent felony as most series charge of the new offense.

8 As of June 2022, the 2021 index crime data is preliminary and subject to change (New York State Division of Criminal Justice Services, 2022b). During the revision process of this research, NYS DCJS updated the data. We requested and obtained the updated monthly data from January 2017 to May 2022. The interrupted time series analyses conducted using data from January 2017 to May 2022 are reported in Table C1 and Table C1.1 in Appendix C.

9 The “widest practicable donor pool” should not include states that experienced similar reforms “during the post-intervention period” (McDowall et al., 2019, p.163). Although Abadie et al. (2010, p.498) stated they discarded states that adopted similar programs during the “sample period,” they considered only the post-intervention period in two of their examples (Abadie et al., 2010, 2015). In our SCM, the sample period is from January 2017 to December 2020, and the post-intervention period is from January 2020 to December 2020. Table C2 in Appendix C shows none of the states in the donor pool experienced similar bail reforms during the post-intervention period. Four jurisdictions enacted large-scale bail reform during the sample period: Alaska, Connecticut, Illinois and Harris County, Texas. The potential contamination is minimum for the following reasons. First, the bail reform in Alaska, Connecticut, and Illinois took effect shortly after the onset of our sample period. In other words, during most of our sample period, their bail reform had been ongoing. And the first five to fifteen months did not affect our main results according to the “discard-a-segment” tests in Table 3, Table 5, and Table 7. Second, although Harris County is the most populous county in Texas, the number of offenses in Harris County accounted for about 19% of all offenses in Texas (Texas Department of Public Safety, 2021). It is hardly likely that reform in Harris County would represent state-level changes. Third, when included in the sample, Illinois and Connecticut obtained zero weights in the murder rate model. Alaska and Texas obtained zero weights in all models. In other words, most of the time they were not “selected states.” When they were selected, “discard-a-state” tests in Table 2, Table 4, and Table 6 show robust results when they were dropped. Therefore, most of the time our main results are identical or similar whether excluding them or not. We reported SCM using the uncontaminated donor pool excluding AK, CT, IL, and TX in Appendix C. The results, including the construction of synthetic controls, the in-sample placebo tests, and the robustness tests, replicated the results in the main text.

10 According to the Florida Department of Law Enforcement (2022), their UCR records are issued annually. The data records in FBI Return A master files show zeros in most months of a year and nonzeros in the 12^th month or sometimes in the 6^th month of that year. Although both zero crime and no reporting are shown as zeros in the dataset (Kaplan, 2022), at the state level it is not possible that zero crime occurs in 11 out of 12 months a year. Therefore, a reasonable conclusion from the examination of the data is that Florida reports annual or semi-annual data.

11 For the murder rate, among models that yielded significant autoregressive/moving average coefficients within boundaries of stationarity/invertibility and successfully removed residual autocorrelations, BIC selected a first-order autoregressive model with additive seasonal (quarterly) effects, that is, AR(1,4), as the best one. The ARIMA (1,0,0)(1,0,0)₁₂ model was also an adequate model. The effects of bail reform from an ARIMA (1,0,0)(1,0,0)₁₂ model and an AR(1,4) were very similar. Therefore, Table 1 reports results from ARIMA (1,0,0)(1,0,0)₁₂ models for all seven time series.