Schuler Megan S, Leoutsakos Jeannie-Marie S, Stuart Elizabeth A
Department of Mental Health, Johns Hopkins University Bloomberg School of Public Health, Baltimore, MD 21205, phone: 803-730-3476, fax: 814-863-0000.
Department of Psychiatry and Behavioral Sciences, Johns Hopkins University School of Medicine, Baltimore, MD 21224.
Health Serv Outcomes Res Methodol. 2014 Dec;14(4):232-254. doi: 10.1007/s10742-014-0122-0.
Confounding is widely recognized in settings where all variables are fully observed, yet recognition of and statistical methods to address confounding in the context of latent class regression are slowly emerging. In this study we focus on confounding when regressing a distal outcome on latent class; extending standard confounding methods is not straightforward when the treatment of interest is a latent variable. We describe a recent 1-step method, as well as two 3-step methods (modal and pseudoclass assignment) that incorporate propensity score weighting. Using simulated data, we compare the performance of these three adjusted methods to an unadjusted 1-step and unadjusted 3-step method. We also present an applied example regarding adolescent substance use treatment that examines the effect of treatment service class on subsequent substance use problems. Our simulations indicated that the adjusted 1-step method and both adjusted 3-step methods significantly reduced bias arising from confounding relative to the unadjusted 1-step and 3-step approaches. However, the adjusted 1-step method performed better than the adjusted 3-step methods with regard to bias and 95% CI coverage, particularly when class separation was poor. Our applied example also highlighted the importance of addressing confounding - both unadjusted methods indicated significant differences across treatment classes with respect to the outcome, yet these class differences were not significant when using any of the three adjusted methods. Potential confounding should be carefully considered when conducting latent class regression with a distal outcome; failure to do so may results in significantly biased effect estimates or incorrect inferences.
混杂因素在所有变量都能被完全观测到的情况下已得到广泛认可，但在潜在类别回归背景下对混杂因素的认识以及解决混杂问题的统计方法正在缓慢兴起。在本研究中，我们关注在潜在类别上对远端结局进行回归时的混杂因素；当感兴趣的处理是一个潜在变量时，扩展标准的混杂方法并非易事。我们描述了一种最近的一步法，以及两种纳入倾向得分加权的三步法（模态和伪类别分配）。使用模拟数据，我们将这三种调整方法的性能与未调整的一步法和未调整的三步法进行比较。我们还给出了一个关于青少年物质使用治疗的应用实例，该实例考察了治疗服务类别对后续物质使用问题的影响。我们的模拟表明，相对于未调整的一步法和三步法，调整后的一步法以及两种调整后的三步法都显著降低了由混杂因素引起的偏差。然而，在偏差和95%置信区间覆盖方面，调整后的一步法比调整后的三步法表现更好，尤其是在类别分离较差时。我们的应用实例也凸显了处理混杂因素的重要性——两种未调整的方法都表明各治疗类别在结局方面存在显著差异，但使用三种调整方法中的任何一种时，这些类别差异并不显著。在对远端结局进行潜在类别回归时，应仔细考虑潜在的混杂因素；否则可能会导致效应估计出现显著偏差或推断错误。
