A Bayesian Approach to Selection Bias Applied to Racial Profiling

本项目提出研究新的统计方法，以控制选择偏差时，面对显著的数据限制。一般来说，使用非随机数据样本可能导致模型参数的估计有很大偏差。反过来，这可能导致基于系统偏见估计的不正确结论。该项目控制了回归分析中的选择偏差，并应用于种族貌相背景。具体来说，该项目提出了一个统计模型，该模型将估算一个选择模型，而不需要对未被选中的个人进行个人层面的数据。在控制选择偏差时，项目的前半部分量化了使用总体水平数据而不是个人水平数据的潜在精度损失。项目的后半部分将新模型应用于种族貌相的实际选择问题，模拟在高速公路上行驶的汽车的选择（而不仅仅是停车的汽车子集）。这个模型将量化种族定性的成本和收益，包括搜查无辜司机的比率和从这些拦截中查获的毒品。虽然这些实证结果并不能明确地回答种族定性作为一项政策是否符合宪法的高度要求，但它将为用公正的实证结果回答这个问题提供一个开端。虽然该项目侧重于种族貌相背景下的选择偏差，但所使用的方法广泛适用于许多实证问题，并且在无法获得个人层面数据时为控制选择偏差提供了重要的一步。

This project proposes to investigate new statistical methods to control for selection bias when faced with significant data limitations. In general, using non-random samples of data can lead to quite biased estimates of model parameters. This, in turn, can lead to incorrect conclusions based on systematically biased estimates. This project controls for selection bias in regression analysis, with an application to the racial profiling context. Specifically, the project proposes a statistical model that will estimate a selection model without the need for individual-level data for the non-selected individuals. The first half of the project quantifies the potential loss in precision from using aggregate-level data rather than individual-level data when controlling for selection bias. The second half of the project applies the new model to the actual selection problem of racial profiling, modeling the selection for search of cars driving down the highway (rather than just the subset of stopped cars). This model will quantify the costs and benefits of racial profiling, including the rate of searching innocent motorists and the drugs seized from such stops. While these empirical results will not definitively answer whether racial profiling as a policy meets the high demands of the constitution, it will provide a beginning to answer this question with unbiased empirical results. Although the project focuses on selection bias in the racial profiling context, the method used has wide applicably to many empirical questions, and provides a significant step in controlling for selection bias when individual-level data is unavailable.