Robust estimates of the prevalence of drinking-and-driving using secondary data

项目摘要 研究人员和政策制定者需要可靠的估计， 酒后驾车公共卫生和交通安全资源的优化配置， 执法，以及对刑事犯罪的适当惩罚主要取决于是否：a）相对较少 驾驶员将其他道路使用者暴露于极端风险，或B）相对多的驾驶员将其他道路使用者暴露于 中度风险。现有的调查和统计方法都强加了有问题的行为假设， 产生一个无信息的广泛的值。为了克服这些挑战，我们提出了一个偏差校正 Levitt和Porter（2001年）和Dunn和Tefft（2020年）开发的LPDT方法的版本（以下简称BC- LPDT），其放宽了一种驱动程序类型与另一种驱动程序类型交互的概率 只取决于道路上每个驾驶员类型的份额。该项目通过以下方式建立重要的研究基础设施： 1)增加最大似然函数以考虑驾驶员的时空排序; 2） 估计驾驶员相互作用概率;以及3）生成流行率和碰撞的无偏估计 利用现有的行政数据来分析与酒后驾车相关的风险。我们将把这个方法应用到两个 数据来源：死亡分析报告系统（法尔斯）（目标1）和坠机报告抽样系统 （目标2）。法尔斯是美国致命车祸的普查，将使我们能够恢复 酒精参与（BAC> 0）和酒精受损（BAC ≥ 0.08）的司机造成致命的碰撞。CRSS是一个全国性的 所有机动车碰撞的代表性样本，使我们能够恢复造成任何 个碰撞测试所附的流行率估计数将通过解决 在过去四十年中，流行率是否稳步下降（与 NRS估计）或自20世纪90年代末以来停滞不前（与以前的LPDT结果和崩溃统计数据一致）。 前者意味着现有的政策组合继续显示出重要的回报;后者意味着现有的政策组合继续显示出重要的回报。 政策已经耗尽了它的回报，创新是必要的，以减少公共卫生成本的饮酒， 开车证明这一方法的可行性和稳健性的好处，包括 方法和数据集（目标3）之间的估计是显而易见的。首先，bc-LPDT方法可以应用于崩溃 每年发布的数据，大大提高了流行率和撞车风险的频率和及时性 估计比可能的调查工具。第二，产生新估计的边际成本， 从这些数据来看，患病率和相对风险几乎为零。第三，bc-LPDT方法非常灵活： 该方法可以应用于较小的地理区域（州一级）和子群体（年龄，性别，种族， 种族）或司机，车辆和碰撞的特征，以调查酒精参与如何相互作用 与属性，如嗜睡，药物参与，和分心。因此，估计和方法， 该项目将是公共卫生监测的宝贵产出，也是未来政策评估的投入。

Project Summary Researchers and policymakers require reliable estimates of both the prevalence and the relative risk of drinking-and-driving. The optimal allocation of public health and traffic safety resources, the efficient level of enforcement, and the appropriate penalty to criminal offenses depend critically on whether: a) relatively few drivers expose other road-users to extreme risk, or b) relatively many drivers expose other road-users to moderate risk. Existing surveys and statistical methods each impose questionable behavioral assumptions and produce an uninformatively wide range of values. To overcome these challenges, we propose a bias-corrected version of the LPDT method developed by Levitt and Porter (2001) and Dunn and Tefft (2020) (hereafter, bc- LPDT) that relaxes the assumption that the probability of one driver-type interacting with another driver-type depends only on the share of each driver-type on the road. This project builds vital research infrastructure by: 1) augmenting the maximum likelihood function to account for the spatiotemporal sorting of drivers; 2) estimating driver interaction probabilities; and 3) generating unbiased estimates of the prevalence and crash risk associated with drinking-and-driving using existing administrative data. We will apply this method to two data sources: the Fatality Analysis Reporting System (FARS) (Aim 1) and Crash Report Sampling System (CRSS) (Aim 2). FARS is a census of fatal crashes in the US and will allow us to recover the relative risk of alcohol-involved (BAC>0) and alcohol-impaired (BAC≥0.08) drivers causing a fatal crash. CRSS is a nationally representative sample of all motor vehicle crashes, allowing us to recover the relative risk of causing any crash. The accompanying estimates of prevalence will make an immediate contribution by resolving the outstanding question of whether prevalence has fallen steadily over the past four decades (consistent with NRS estimates) or stagnated since the late 1990s (consistent with previous LPDT results and crash statistics). The former would imply the existing policy portfolio continues to show important returns; the latter that existing policy has exhausted its returns and innovation is necessary to reduce the public health cost of drinking and driving. The benefits of demonstrating the feasibility and robustness of this approach, including reconciling estimates across method and dataset (Aim 3) are manifest. First, the bc-LPDT method can be applied to crash data that are published annually, greatly increasing the frequency and timeliness of prevalence and crash risk estimates than is possible with survey instruments. Second, the marginal cost of generating new estimates of prevalence and relative risk from these data is nearly zero. Third, the bc-LPDT method is exceptionally flexible: the approach can be applied to smaller geographies (state-level) and subpopulations (age, gender, race, ethnicity) or characteristics of drivers, vehicles, and crashes to investigate how alcohol involvement interacts with attributes, such as drowsiness, drug involvement, and distraction. Thus, the estimates and methods from this project will be valuable outputs for public health surveillance and inputs for future policy evaluation.