Division of Biostatistics, Department of Epidemiology, Michigan State University, B601 West Fee Hall, East Lansing, MI 48823, USA.
Stat Med. 2009 Oct 30;28(24):2967-88. doi: 10.1002/sim.3675.
Statistical theory requires that analysis of study outcomes be conducted conditional on the design process. Ignoring this process may result in severely biased estimates, leading to false inferences, especially when the outcome variable is associated with design variables. We propose in this paper a class of hierarchical models to investigate the dependence between the design process and the study outcomes of primary interest. We discuss a fully parametric and a semi-parametric formulation of the hypothesized model and propose the EM algorithm to obtain maximum likelihood estimates. Our numerical results show that the semi-parametric approach outperforms the fully parametric model with respect to some key features of the model. The methodology is used to gain insight into the mechanism that generates breast cancer literacy outcomes in a study conducted among medically underserved females in Michigan.
统计理论要求根据设计过程对研究结果进行分析。如果忽略这个过程，可能会导致严重的偏差估计，从而导致错误的推断，特别是当结果变量与设计变量相关时。本文提出了一类层次模型来研究设计过程与主要关注的研究结果之间的相关性。我们讨论了假设模型的完全参数和半参数形式，并提出了 EM 算法来获得最大似然估计。我们的数值结果表明，在模型的某些关键特征方面，半参数方法优于完全参数模型。该方法用于深入了解密歇根州医疗服务不足的女性中进行的一项研究中乳腺癌知识结果产生的机制。
