Every year, something like 13 million misdemeanor charges are filed in the United States. These charges, ranging from traffic violations to serious assaults, may be less flashy than felonies, but they are the main way Americans experience the criminal-justice system. We prosecute misdemeanors because, among other things, we want there to be fewer of them, and we believe prosecution deters reoffending. But a recent blockbuster paper makes a startling claim to the contrary: Prosecuting misdemeanants actually increases the likelihood that they will offend again. The paper has been heralded by supporters of progressive district attorneys who have used their position to unilaterally impose reforms on the criminal-justice system, including refusing to prosecute many misdemeanants. Boston D.A. Rachael Rollins, who provided the data for the study, has claimed it confirms the wisdom of her approach. So have other reformers such as Chicago-area state’s attorney Kim Foxx and San Francisco district attorney Chesa Boudin. Policy-makers, however, should exercise caution before reaching such expansive conclusions. The paper can just as easily be read to endorse more modest reforms — especially keeping in mind long-established principles of criminal justice on which it is silent. The paper is the work of three researchers: Rutgers’s Amanda Agan, Texas A&M’s Jennifer Doleac, and NYU’s Anna Harvey. Both Doleac and Agan have previously published research that challenges progressives’ policy preferences, so their new findings were probably not driven by a desire to achieve a politically convenient result. To conduct their study, the three obtained data on every criminal case arraigned in Suffolk County (home to Boston) between 2004 and September 2018. They analyzed the relationship between whether a misdemeanant was prosecuted and whether he was subsequently rearraigned, indicating he reoffended. The effects are startling: not being prosecuted for a misdemeanor reduces the probability of a future misdemeanor complaint by 60 percent, and of a future felony complaint by 47 percent. It also significantly reduces the probability of future violent, motor-vehicle, and disorder/theft offenses, although not of drug misdemeanors. In other words, prosecution not only does not deter subsequent crimes, it increases the chance of reoffense. This, the paper’s authors suggest, is because any deterrent effect is outweighed by the effects on misdemeanants’ labor-market prospects. Unemployment can lead to crime, and being prosecuted can increase one’s chances of becoming unemployed. It also creates a criminal record, making the offender less employable and therefore more crime-prone. How did the researchers reach these conclusions? To understand how prosecution affects reoffending risk, we can’t just compare prosecuted and unprosecuted offenders to see which commit more crimes. A person might not be prosecuted precisely because she is judged not a risk — there are confounding variables determining both that we must control for. To get around this problem, the paper uses an “instrumental variable” uncorrelated with those confounders. All misdemeanants charged in Suffolk County are arraigned by an assistant district attorney (ADA). Using each ADA’s other cases, the authors construct a measure of their “leniency,” i.e., their propensity to prosecute a given offender. Mostly, the ADAs appear to agree about prosecutions — they would typically prosecute about 70 percent of cases, and typically drop about 20 percent. But about 10 percent of the time, they vary in their inferred leniency. The assignment of these 10 percent of offenders to ADAs of varying leniency becomes the source of randomness. What this means is that most of the paper’s results apply to those “marginal” offenders. Not prosecuting all misdemeanants won’t cut everyone’s risk of offending in half, but declining to prosecute the marginal misdemeanant — the one on the line between prosecution and non-prosecution — reduces his chance of reoffending by a lot. Toward the end of the paper, the authors generalize from these marginal offenders. They find non-prosecution has a large effect on all offenders, about a 15 percent reduction in reoffense risk on average. The effect is most concentrated among those who were typically prosecuted: Being prosecuted, the authors write, made them much more likely to reoffend. Strangely, the effect on those who are typically not prosecuted is indistinguishable from zero. If ADAs had prosecuted those they typically would not, there would have been no average effect on their future propensity to offend. When I asked Doleac about this finding, she suggested that it may reflect ADA judgment about culpability: The people most likely to reoffend may also be those whom ADAs are most lenient toward — the young, the mentally ill, etc. Those least likely to reoffend — well-adjusted adults who made a mistake — are those for whom ADAs have the least sympathy, but for whom prosecution could have a big, negative impact. This suggests, in Doleac’s view, that there is a fundamental difference between culpability and risk. To me, it also indicates that ADAs are not great judges of the effects of their prosecution choices. There’s a second key detail that the authors attend to, but that has been missed in some commentary on the paper: Most of the non-prosecution effect they measure is the result of first-time offenders, who become much more likely to commit crime if prosecuted. By contrast, prosecuting repeat offenders of any sort has little discernible effect on the likelihood they will offend again in the future. This is not surprising, given that most offenses are committed by a handful of offenders — criminological research consistently finds that a small, offense-prone population drives most crime. For those outside of that population — including many first-time offenders — prosecution is unlikely to deter them from doing something they wouldn’t do anyway, but could have adverse effects that push them toward crime. So should we prosecute misdemeanants less? We can’t draw too dramatic a conclusion from one study of one county, no matter how well-designed. And while this one relies on the latest in statistical techniques, we should always be wary of findings that can only be arrived at through extensive statistical interrogation. The sheer complexity of the instrument the paper uses, combined with the very large effects it finds, should temper enthusiasm — there are simply too many researcher degrees of freedom not to. That said, we can cautiously conclude that the best evidence says the marginal misdemeanant should be prosecuted less often. But if ADAs are bad at judging the effects of their prosecution, then we shouldn’t assume they’re good at telling the marginal misdemeanant from the future serial offender. So whether the study’s results are wrong or ADAs are poor judges of how prosecution will relate to future offending, we should be wary of giving them too much leeway in deciding who is and isn’t a marginal case. We can instead offer a rule of thumb: When in doubt, err on the side of not prosecuting first-time misdemeanants. Diverting these offenders, with the threat of more serious punishment if they reoffend, could help clear dockets while minimizing crime. It would also free ADAs to focus on repeat misdemeanants. Targeting repeat offenders would mitigate the risk of abuse of first-time diversion, by making clear that a “second chance” won’t be followed by a third, a fourth, a fifth, and so on. Research on California’s “three-strikes law,” for example, indicates that increasing punishment for repeat offenders can have a powerful deterrent effect. The above approach is different from the idea that we should in general prosecute misdemeanants a lot less — a valid interpretation of the paper’s findings, but not necessarily the right one, for two reasons. First, deterrence is not the only reason to prosecute an offender. Advocates of not prosecuting misdemeanors tend to invoke “victimless” crimes such as drug possession and prostitution. But misdemeanors can also include offenses such as simple assault and auto theft — crimes that harm others. Such crimes reasonably elicit a demand for retributive justice. It offends our moral sensibilities to think that a person who commits a serious but not felonious assault could get off scot-free. Second, systematic reductions in leniency may affect all criminals’ decision-making, increasing their propensity to offend in the long-run. The paper shows that Rollins’s move toward non-prosecution of misdemeanors did not in the aggregate increase misdemeanor offenses, but the data it uses account only for the period between her election in January 2019 and March 2020, when the coronavirus crisis began. It’s entirely possible that criminals will adapt, and misdemeanor offending will increase, in the long run. Blanket policy changes can induce increases in offending. California’s 2014 increase to the threshold for felony theft, for example, predictably led to an increase in theft at the city level, indicating that offenders change their behavior in response to such shifts. Coming face to face with the justice system can be time-consuming and exhausting, and may, at the margins, increase rather than reduce a person’s propensity to offend. Even those of us highly concerned with public safety should be interested in creative solutions that minimize crime and disorder. At the same time, policy-makers should not get ahead of themselves — as some have in the rush to defund police departments and decrease the use of more serious charges. Good research is the basis of good policy, and this research makes a valuable contribution to public-safety policy. But we should be cautious in how far we go with it — careful changes around the edges are always safer than blanket transformations.