Department of Biostatistics, University of Michigan School of Public Health, 1420 Washington Heights, Ann Arbor, MI 48109, United States.
Accid Anal Prev. 2010 Mar;42(2):530-9. doi: 10.1016/j.aap.2009.09.019. Epub 2009 Oct 23.
The Crash Injury Research Engineering Network (CIREN) database contains detailed medical and crash information on a large number of severely injured occupants in motor vehicle crashes. CIREN's major limitation for stand-alone analyses to explore injury risk factors is that control subjects without a given injury type must have another severe injury to be included in the database. This leads to bias toward the null in the estimation of risk associations. One method to cope with this limitation is to obtain information about occupants without a given injury type from the National Automotive Sampling System's Crashworthiness Data System (NASS-CDS), which is a probability sample of towaway crashes, containing similar crash information, but less medical detail. Combining CIREN and NASS-CDS in this manner takes advantage of the increased sample size when outcomes are available in both datasets; otherwise NASS-CDS can serve as a sample of controls to be combined with CIREN cases, possibly under a sensitivity analysis that includes and excludes NASS-CDS subjects whose status as a control is uncertain. Because CIREN is not a probability sample of crashes that meet its inclusion criteria, we develop a method to estimate the probability of selection for the CIREN cases using data from NASS-CDS. These estimated probabilities are then used to compute "pseudo-weights" for the CIREN cases. These pseudo-weights not only allow for reduced bias in the estimation of risk associations, they allow direct prevalence estimates to be made using medical outcome data available only in CIREN. We illustrate the use of these methods with both simulation studies and application to estimation of prevalence and predictors of AIS 3+ injury risk to head, thorax, and lower extremity regions, as well as prevalence and predictors of acetabular pelvic fractures. Results of these analyses demonstrate combining NASS and CIREN data can yield improvements in mean square error and nominal confidence interval coverage over analyses that use either the NASS-CDS or the CIREN sample alone.
Crash Injury Research Engineering Network（CIREN）数据库包含大量在机动车事故中严重受伤的乘员的详细医疗和碰撞信息。CIREN 进行独立分析以探索受伤风险因素的主要限制是，没有特定受伤类型的对照受试者必须有另一种严重受伤才能被纳入数据库。这导致在估计风险关联时偏向于零。应对这种限制的一种方法是从 National Automotive Sampling System 的 Crashworthiness Data System（NASS-CDS）中获取没有特定受伤类型的乘员信息，NASS-CDS 是拖曳事故的概率样本，包含相似的碰撞信息，但医疗细节较少。以这种方式将 CIREN 和 NASS-CDS 结合使用，可以在两个数据集都有结果时利用增加的样本量；否则，NASS-CDS 可以作为与 CIREN 病例相结合的对照样本，可能在灵敏度分析中包括和排除 NASS-CDS 受试者，这些受试者作为对照的状态不确定。由于 CIREN 不是符合其纳入标准的碰撞的概率样本，因此我们开发了一种使用 NASS-CDS 数据估算 CIREN 病例选择概率的方法。然后，这些估计的概率用于计算 CIREN 病例的“伪权重”。这些伪权重不仅可以减少风险关联估计的偏差，还可以使用仅在 CIREN 中可用的医疗结果数据直接进行患病率估计。我们使用模拟研究和对 AIS 3+头部、胸部和下肢区域受伤风险的患病率和预测因素以及髋臼骨盆骨折的患病率和预测因素的估计来说明这些方法的使用。这些分析的结果表明，与单独使用 NASS-CDS 或 CIREN 样本的分析相比，结合 NASS 和 CIREN 数据可以提高均方误差和名义置信区间覆盖度。
