Theories of rational behavior assume that actors make decisions where the benefits of their acts exceed their costs or losses. If those expected costs and benefits change over time, the behavior will change accordingly as actors learn and internalize the parameters of success and failure. In the context of proactive policing, police stops that achieve any of several goals—constitutional compliance, stops that lead to “good” arrests or summonses, stops that lead to seizures of weapons, drugs, or other contraband, or stops that produce good will and citizen cooperation—should signal to officers the features of a stop that increase its rewards or benefits. Having formed a subjective estimate of success (i.e., prior beliefs), officers should observe their outcomes in subsequent encounters and form updated probability estimates, with specific features of the event, with a positive weight on those features. Officers should also learn the features of unproductive stops and adjust accordingly. A rational actor would pursue “good” or “productive” stops and avoid “unproductive” stops by updating their knowledge of these features through experience. We analyze data on 4.9 million Terry stops in New York City from 2004–2016 to estimate the extent of updating by officers in the New York Police Department. We compare models using a frequentist analysis of officer behavior with a Bayesian analysis where subsequent events are weighted by the signals from prior events. By comparing productive and unproductive stops, the analysis estimates the weights or values—an experience effect—that officers assign to the signals of each type of stop outcome. We find evidence of updating using both analytic methods, although the “hit rates”—our measure of stop productivity including recovery of firearms or arrests for criminal behavior—remain low. Updating is independent of total officer stop activity each month, suggesting that learning may be selective and specific to certain stop features. However, hit rates decline as officer stop activity increases. Both updating and hit rates improved as stop rates declined following a series of internal memoranda and trial orders beginning in May 2012. There is also evidence of differential updating by officers conditional on a variety of features of prior and current stops, including suspect race and stop legality. Though our analysis is limited to NYPD stops, given the ubiquity of policing regimes of intensive stop and frisk encounters across the United States, the relevance of these findings reaches beyond New York City. These regimes reveal tensions between the Terry jurisprudence of reasonable suspicion and evidence on contemporary police practices across the country.

理性行为理论假设，行为者在其行为的收益超过其成本或损失的情况下做出决策。如果这些预期成本和收益随着时间的推移而变化，随着参与者学习和内化成功和失败的参数，行为也会相应改变。在主动警务的背景下，警察拦截可以实现以下几个目标中的任何一个：遵守宪法、导致“良好”逮捕或传票的拦截、导致扣押武器、毒品或其他违禁品的拦截，或者产生善意的拦截和公民合作——应向官员表明停车的特征，以增加其奖励或好处。在形成对成功的主观估计（即先前的信念）后，官员应该在随后的遭遇中观察他们的结果并形成更新的概率估计，具有事件的具体特征，并对这些特征给予积极的重视。官员还应该了解非生产性拦截的特征并做出相应调整。理性的行为者会通过经验更新对这些特征的了解来追求“好的”或“有成效的”停留，并避免“无成效”的停留。我们分析了 2004 年至 2016 年纽约市 490 万次特里拦截的数据，以估计纽约警察局警官更新的程度。我们将使用军官行为频率分析与贝叶斯分析的模型进行比较，其中后续事件由先前事件的信号加权。通过比较有效和无效的停车，分析可以估计官员分配给每种类型停车结果信号的权重或值（一种经验效应）。我们发现了使用这两种分析方法进行更新的证据，尽管“命中率”（我们衡量拦截生产力的指标，包括追回枪支或逮捕犯罪行为）仍然很低。更新与每个月的警员停靠站活动总数无关，这表明学习可能是有选择性的并且特定于某些停靠站功能。然而，随着官员停止活动的增加，命中率下降。自 2012 年 5 月开始，一系列内部备忘录和试验命令发布后，随着拦截率下降，更新率和命中率均有所提高。还有证据表明，官员根据先前和当前拦截的各种特征进行差异化更新，包括可疑种族和拦截合法性。尽管我们的分析仅限于纽约警察局的拦截，鉴于美国各地普遍存在密集拦截搜身的警务制度，这些调查结果的相关性不仅限于纽约市。这些制度揭示了特里合理怀疑的判例与全国各地当代警察实践的证据之间的紧张关系。