Using propensity scores for causal inference with covariate measurement error

描述（由申请人提供）：许多公共卫生研究，包括比较有效性研究，旨在回答诸如“什么对谁有效？”或“在什么条件下有效？”之类的问题。这些问题通常只能在非实验研究的背景下回答。倾向分数是非实验研究的关键统计工具，因为它有助于“苹果与苹果”的比较。然而，许多非实验研究需要结合来自多个数据集的信息，这些数据集可能有略微不同的可用测量方法。例子包括使用两种不同的量表来测量抑郁症，或者直接与父母报告的自闭症儿童行为测量。现有的倾向评分方法不能处理协变量测量有误差或在治疗组和对照组之间测量不同的情况。这在心理健康研究中是一个特别的挑战，因为研究中的许多疾病和因素不是直接观察到的，而是作为潜在构念进行建模的。该项目将开发和评估新的统计方法，用于在非实验研究中估计治疗效果，当协变量测量有误差或在治疗组和对照组中以不同的方式测量时。这项工作将结合倾向评分方法来估计治疗效果，潜在变量方法，以及处理缺失数据的多重imputation方法。目标是：1)在使用倾向评分方法估计治疗效果时，调查协变量测量误差的含义；2)在治疗组和对照组中测量误差或测量不同的情况下，开发和评估倾向评分方法；3)使用目标1和2的结果和方法来估计三项研究中的治疗效果。每组都将接受干预的一组与外部对照组进行比较，然后通过将这些估计与相同干预的随机试验报告的治疗效果进行比较来检验方法。这些方法将在三个心理健康研究的背景下进行检验，以评估(1)自闭症儿童的早期干预，(2)围产期抑郁症的预防，以及(3)使用银杏叶预防痴呆和阿尔茨海默病的有效性。

DESCRIPTION (provided by applicant): Many studies in public health, including comparative effectiveness research, aim to answer questions such as "what works for whom?" or "under what conditions does it work?" Such questions can often only be answered in the context of a non-experimental study. Propensity scores are a key statistical tool for non-experimental studies because they facilitate the comparison of "apples with apples." However, many non-experimental studies require combining information from multiple data sets, which may have slightly different measures available. Examples include depression measured using two different scales, or direct- versus parent-reported measures of behavior among children with autism. Existing propensity score methods cannot handle situations where the covariates are measured with error or are measured differently across treatment and comparison groups. This is a particular challenge in mental health research, where many of the disorders and factors under study are not directly observable and are instead modeled as latent constructs. This project will develop and assess new statistical methods for estimating treatment effects in non-experimental studies when the covariates are measured with error or are measured in different ways across the treatment and comparison groups. The work will tie together propensity score methods for estimating treatment effects, latent variable methods, and multiple imputation methods for handling missing data. The aims are: 1) Investigate the implications of measurement error in the covariates when using propensity score methods to estimate treatment effects, 2) Develop and assess propensity score methods for settings where some of the covariates are measured with error or are measured differently in the treatment and comparison groups, and 3) Use the results and methods from Aims 1 and 2 to estimate treatment effects in three studies, each comparing a group receiving the intervention to an external comparison group, and then test the methods by comparing those estimates to the reported treatment effects from randomized trials of the same interventions. The methods will be examined in the context of three studies in mental health evaluating the effectiveness of (1) early intervention for children with autism, (2) perinatal depression prevention, and (3) the use of ginkgo biloba to prevent dementia and Alzheimer's disease.