New York’s bail reform law and recidivism

In recent years, many states throughout the nation have taken steps to reduce or eliminate the use of monetary bail. According to proponents, the purpose of bail “reform” is to reduce jail populations and decrease income disparities regarding how bail is applied. However, critics believe that such laws would negatively impact public safety due to more defendants committing crimes while on pretrial release. Further, there also concerns that reforms could increase failure-to-appear rates.

One state where bail reform has received considerable attention is New York. Statewide reforms first took effect on January 1, 2020, though it was later amended in April 2020 and May 2022. Currently, there isn’t a ton of research examining whether bail reform contributed to crime, but more research is starting to trickle out on this topic. There seems to be a lot of variaiton regarding the findings though. One of the more recent reports on this topic was published earlier this month by the Data Collaborative for Justice (DCJ), which claimed that bail reform decreased recidivism among pretrial releases in New York City. However, this is inconsistent with a previous study from last year, also conducted in New York City, which found the opposite: bail reform efforts significantly increased recidivism. Thus, it’s important to read these studies very carefully to determine which ones are the most relevant and helpful. In this post, I will review the latest report released by the DCJ and provide my thoughts regarding its validity.

When first enacted in January 2020, New York’s bail reform law set a limited number of “qualifying offenses” (mostly violent felony offenses) that were considered “bail-eligible,” meaning that judges had discretion to set bail. Among bail-eligible cases, judges were still required to consider additional factors into their decision, such as flight risk and the defendant’s ability to pay. Most misdemeanors and nonviolent felonies were not on the list of qualifying offenses and were therefore deemed “bail-ineligible.” In these cases, defendents were subject to mandatory release, often on their own recognizance, or with strict conditions for supervised release.

Not surprisingly, re-arrest rates for certain people (often felons and those with extensive criminal histories) increased under the original reforms. This was noticeable just three months after the law took effect, prompting amendments in April 2020 that expanded bail eligibility for certain charges that were previously “bail-ineligible.” The amendment also specified additional conditions of release that the judge could impose if desired, such as mandatory treatment or electronic monitoring. Reforms were again modified two years later, in May 2022, further expanding bail eligibility based on a defendant’s history of using a gun, whether they were charged with causing serious harm, and if they had violated and Order of Protection. The effects of these reforms were not covered in the DCJ report due to their recency.

Methods

The DCJ study sought to test the impact of bail reform on re-arrest rates among people who were released without bail in New York City. Per a long-standing agreement with the NYC Office of Court Administration (OCA), the authors were able to access non-public data from New York City courts to track arrest rates and collect information on a plethora of offender characteristics. The authors examined outcomes at two points in time: post-January 2020 (after the law first took effect) and post-May 2020 (after the law was amended). The impact of May 2022 reforms could not be assessed due to their recency.

People who were released without bail were considered the treatment group, while people who had bail set served as the comparison condition. The authors used propensity scores based on more than 60 offender characteristics to identify a matched comparison group (i.e., people who had bail set but otherwise were similar to those released without bail). Then, the authors compared recidivism rates across the groups using different types of models (including weighted logistic regression models and hazards models).

Recidivism was measured in terms of two-year re-arrest rates for the following measures: 1) any re-arrest; 2) a felony re-arrest; 3) a violent felony re-arrest; or 4) a firearm re-arrest. The authors compared re-arrest rates for people who faced bail (before the implementation of bail reform) with similar people who were released without bail (after bail reform went into effect). The authors examined the impact of bail reform overall (i.e., post-January 2020), as well as the subsequent impact of amendments passed in mid-2020 (i.e., post-July 2020).  In addition, they explored how recidivism might be moderated by individual-level factors such as criminal history, current charge, previous time spent incarcerated, and demographics.

The authors also assessed how different types of bail reform might impact recidivism differently. First, they examined the impact of mandatory release cases, where they compared cases that had bail set in the first half of 2019 (i.e., before the reforms) with comparable cases that were mandatorily released without bail in the first half of 2020 (i.e., post-bail reform).  Second, they examined how reducing the use of bail among bail-eligible cases affected recidivism, where they compared cases that had bail set in the first half of 2019 with comparable cases that were released without bail (not mandatorily, though) in the first half of 2020.

Findings

According to the authors, eliminating bail for most misdemeanor and nonviolent felony charges was associated with a reduction in two-year recidivism rates. The percentage of people experiencing any re-arrest decreased from 50% to 44% post-bail reform, while the percentage experiencing a felony re-arrest decreased from 27% to 24%. There were no impacts on violent felony re-arrests or firearm re-arrests.

For people charged with bail-eligible offenses (most of which were violent felonies), bail reform did not seem to impact recidivism in either direction.  Importantly, they did find that people facing more serious charges and people with more extensive criminal histories were more likely to recidivate (SEE NUMBERS IN TECH SUPPLEMENT).

The authors found increases in overall, felony, and violent felony recidivism for certain charges, including: 1) burglary in the second degree; (2) a case the judge deemed to involve harm to a person or property; or (3) a domestic violence case involving obstruction of breathing or blood circulation (which also classifies as assault in the third degree). These crimes were deemed bail-ineligible when the law was initially passed in January 2020. Thankfully, this was modified in mid-2020 amendments to the law, which designated these crimes as bail-eligible again.

Limitations

There are some limitatoins of this report that I feel compelled to point out. The first concerns the propensity score matching (PSM) approach. In the DCJ report, the PSM did seem to generate statistically similar groups, which allowed for a comparison between people who were subject to bail and those who were not. According to the report’s technical supplement, there were substantial baseline differences between the comparison and treatment groups. Accounting for these differences is critical when comparing two groups to ensure that comparisons are valid. In this case, the authors used PSM in an effort to balance the two groups on various baseline factors. However, the quality of the PSM method depends on a few things, and there are different types of PSM that can be used.

In the DCJ report, the authors used a PSM approach called “nearest neighbor with caliper” matching. This specific type of PSM excludes people from the data if there are no matches within a specified distance, which generally results in decent matches. This type of matching does tend to generate adequate matches, which is apparent in the DCJ report. Unfortunately though, it can greatly decrease the sample size, as unmatched treatment group members are excluded from the matched data sets if there are no matches within a specified distance (Austin, 2013; Jacovidis et al., in press; Stuart, 2010).

In the DCJ study, the authors excluded comparison group cases that were below the 26th propensity score percentile, as well as treatment group cases that were above the 74th propensity sacore percentile. This reduced the overall sample size by 25% (n=5,490), from 21,351 to 15,860 cases, and the final sample sizes for treatment and comparison groups were 12,350 and 3,510, respectively. Their approach did generate statistically comparable groups, though the substantial loss in data may have impacted the quality of the matches. In other words, the people on the upper and lower bounds were excluded from the analysis (due to not having an adequate match). Except these people are not random, so it is possible that their inclusion could have affected study results.

Similarly, the loss in data described above impacted the comparison-to-treatment group ratio (i.e., the sample size of the comparison group relative to the treatment), which can also affect validity of the results. Understanding this dynamic can be complex, though research suggests that the size of the comparison group is critical for generating adequate matches (even more so than the total sample size). In other words, when the comparison group is larger than the treatment group, the quality of matches tends to increase. This is because when the comparison pool is larger, it increases the likelihood than an adequate match can be found for each treatment group number. Unfortunately though, in the current study, the comparison-to-treatment group ratio was not good in this aspect. For the main analysis, the treatment group was much larger (n=12,350) than the comparison group (n=3,510). When a large portion of treatment group members are excluded, researchers should assess the representativeness of the analytic sample to ensure that it still matches the population, which it doesn’t seem like they did.

Conclusion

Some of the findings here are consistent with other research, particularly regarding the impact of prior criminal history and charge severity in predicting recidivism. Studies appear to be somewhat mixed regarding the aggregate effects of bail reform, but they seem consistent for certain subgroups. Specifically, it seems apparent that individuals who are charged with violent crimes, as well as those more extensive criminal histories, are more likely to see increases in recidivism following bail reform. This is not only consistent with the wider body of research on recidivism, but also mirrors bail reform impacts seen earlier in New York City, as well as other jursidictions such as Los Angeles County and Chicago.

While the DCJ study has a few interesting findings, it’s important to consider these in context. It is tempting to want to summarize complex research findings into something more digestible, but a lot of context can be lost that way. According to the authors, “eliminating bail and pretrial detention for most misdemeanors and nonviolent felonies reduced recidivism in New York City.” In my opinion, this statement is too declarative as currently. More accurately, bail reform reduced recidivism for some people with less serious charges and little to no criminal history. However, it seems to increase recidivism for people charged with violent offenses, as well as people with substantial criminal histories.

This has important implications. In particular, it seems plausible that judges could perhaps limit the use of bail among specific groups of offenders without impacting public safety. This might include nonviolent older defendants, people with clean prior records, or people charged with fraud or public order violations. However, the same is not true for all offenders, particularly offenders who are violent or who have extensive criminal histories. Thus, judges should have discretion to detain those who are violent, have lengthy criminal histories, or who have prior domestic violence convictions, without having to refer to a specified list of “bail-eligible” crimes.

Regardless of the findings discussed here, it’s important to rememebr that there large variation across jurisdictions regarding what bail reform actually is and how it is implemented. Thus, it is a challenging endeavor to study direct impacts of bail reform. At present, more research is needed to understand the overall effectiveness and unintended consequences of bail reform.