Bail Pretrial Reform

What the evidence says

Across the strongest-designed studies here, pretrial detention and cash-only bail move a defendant through the system toward conviction and incarceration. A controlled analysis by Donnelly (2018) finds that, after holding case severity and criminal history constant, cash-only bail and pretrial detention increase the likelihood of conviction, of pleading guilty, of being incarcerated, and of receiving a longer sentence. DOJ figures cited by Jones (2013) put a number on the conviction gap: 78% of defendants held on bail are eventually convicted versus 60% of released defendants, an 18-percentage-point difference, and detained defendants are four times more likely to be sentenced to incarceration and receive sentences 86% longer than released defendants. Donnelly's Delaware data independently arrive at the same ~4x figure for prison-sentence likelihood, and show almost 20% of pretrial detainees get two-year-plus sentences versus 1% of those released before trial. Page's (2022) review states the same direction without new magnitudes: detained defendants plead guilty more often and earlier, and are more likely to be sentenced to prison and to longer terms, than comparable released defendants.

The pretrial stage also carries most of the measured Black-White disparity in these downstream outcomes. Donnelly's decomposition finds pretrial conditions account for 43.5% of the explainable Black-White disparity in convictions, 37.2% of the disparity in guilty pleas, and under 25% of the disparity in average incarceration sentence length; a companion claim puts the combined contribution at 30% to 47% of the explained disparity across dispositions. That disparity begins upstream, at the bail-setting stage itself: White defendants are released before trial at 76.7% versus 69.5% for Black defendants (a 7.2-point gap), and Donnelly attributes about 62% of that gap (4.2 of 6.8 points) to the disparate impact of individual judges' decisions rather than to differences in observable case characteristics — controlling for criminal history and crime type only shrinks the disparate-impact estimate by 17%. Arnold (2017) finds Black defendants are 11.2 percentage points more likely to be assigned monetary bail than White defendants and, when bail is set, face amounts $14,376 higher. That racial bias is not uniform: it is substantially lower among high-volume judges (full-time judges hearing an average of 6,239 cases a year, and the most experienced part-time judges hearing thousands of cases over a career) and much larger among the least experienced part-time judges (a few hundred cases per career), and it appears among both white and Black judges. The same study finds no measured bias against Hispanic defendants. Judge-level leniency itself has a directly estimated causal effect: a one-percentage-point increase in a judge's predicted leniency raises a defendant's release probability by 0.96 points, and targeting the most discriminatory New York City judges with a race-neutral release quota is estimated to cut the average level of disparate impact by 36%.

On the public-safety rationale for financial conditions, the evidence does not support a deterrent effect. Ouss (2023), studying Philadelphia's elimination of cash bail for a set of eligible offenses, finds the policy increased release on own recognizance by 11 percentage points (22% relative to the pre-reform mean) with no statistically detectable change in the pretrial detention rate, in failure-to-appear, or in new charges within six months — point estimates near zero, tight enough to reject even small increases at the 5% level. Page's review states the same conclusion in general terms: there is no evidence money bail keeps the public safer, and only mixed evidence it affects appearance rates. Kleinberg's (2017) counterfactual simulation of algorithmic risk-based release goes further, estimating that crime could fall by up to 24.8% with no change in jailing rates, that reductions appear across all crime categories including violent ones, and that a policy targeting violent crime specifically could cut murder, rape, and robbery by 57.3% without increasing detention; applied nationally, the same approach is estimated to cut crime by 18.8% or jail populations by 24.5% holding the other constant. That paper also documents what current judicial decisions leave on the table: judges release 48.5% of the defendants the algorithm scores as the riskiest 1%, and those released defendants fail to appear at 56.3%, are rearrested at 62.7%, and are arrested for murder, rape, or robbery at 4.8%.

Pretrial detention carries direct economic and family costs independent of case outcome. Wakefield (2020) finds detention adds a 6-percentage-point drop in the share of people with labor earnings beyond the loss already associated with conviction, and raises the hazard of losing a household partnership by a factor of 1.29 for detainees who are not convicted and 2.15 for those who receive an additional prison sentence. Arnold estimates detained defendants lose about $40,000 in earnings and government benefits. Jail populations and time spent inside have both grown over the same period this literature covers: the national jail population rose from 182,288 in 1980 to 740,700 in 2016, unconvicted people now make up 74% of the roughly 631,000-person jail population, and average length of stay rose from 14 days in 1983 to 26 days in 2019. Bail dollar amounts have risen alongside this: the share of defendants facing bail of $10,000 or more grew from roughly 20% in 1990 to nearly 60% in 2009, and the median bail amount now represents eight months of income for the typical detained defendant.

Where the evidence splits

Page's review states there is "only mixed evidence" that money bail affects appearance rates, without specifying which studies pull in which direction. Ouss's Philadelphia analysis, by contrast, reports a precise null: point estimates near zero and a rejection, at the 5% level, of any failure-to-appear increase larger than 0.009 percentage points from removing cash bail for eligible offenses. These are not strictly incompatible — Page is characterizing a broader literature and Ouss is one study within it — but the sources do not explain what produces the "mixed" pattern Page describes (jurisdiction, offense type, or era), so the split is unexplained by the sources on this page.

Thresholds and dose-response

Racial bias in bail decisions varies with judge experience: bias is substantially lower among full-time judges (an average of 6,239 cases per year) and experienced part-time judges (thousands of cases over a career), and much larger among the least experienced part-time judges, who hear only a few hundred bail cases in their entire career.

Jail stays of under three days are associated with decreased compliance with subsequent court dates, even at that short duration.

Open empirical questions

• What happens to recidivism, appearance, and conviction outcomes beyond the six-month windows measured here — do the Philadelphia No-Cash-Bail null results (Ouss) hold over multi-year follow-up?
• Kleinberg's crime and jail-population reductions are counterfactual simulations against historical judicial decisions; no source here reports outcomes from an actually deployed algorithmic risk-ranking system.
• No source here establishes a specific bail-dollar-amount or income-ratio threshold at which release probability or detention length changes.
• Whether the racial disparities in bail-setting and downstream outcomes documented here narrow, persist, or shift under reforms like New Jersey's or Philadelphia's over a multi-year horizon is not measured in these sources.