Statistical Methods For Human Studies

在开发灵活的半参数贝叶斯方法以评估环境和遗传预测因子与健康反应分布之间的关系方面取得了实质性进展。 我们开发了一种方法，可以有效地识别预测疾病表型的单核苷酸多态性。 该方法依赖于一种新的非参数贝叶斯方法，该方法根据SNP对表型的影响自适应地将SNP分配到未知数量的聚类中。 这种自适应的概率聚类在降低维度的同时保持了灵活性。 我们还开发了根据患者结局分布对医院进行聚类的方法。 以前的聚类方法集中在分布的某个方面，例如典型患者的平均结果。 相反，我们的方法将具有相同分布特征的医院分组。 这一点很重要，因为两家医院的平均值可能相同，但病情最重和最健康的患者的结果却截然不同。 这篇文章被选为最佳论文，将于明年在美国统计协会杂志的理论和方法部分发表，并将在年度联合统计会议的特别会议上重点介绍。 我们还开发了灵活的建模框架，允许健康结果的分布随预测因子的非参数变化。 这使人们能够评估对群体中最敏感个体的不同影响，而不是假设环境暴露的治疗效果或不良影响对所有个体都是相同的。 我们使用相关的想法来评估怀孕期间体重增加轨迹对孕产妇和儿童结局（如肥胖）的影响。

Substantial progress has been made in development of flexible semiparametric Bayesian methods for assessing relationships between environmental and genetic predictors and the distribution of health responses. We developed an approach that allows efficient identification of single nucleotide polymorphisms predictive of a disease phenotype. The proposed method relied on a novel nonparametric Bayes approach, which adaptively allocated SNPs into an unknown number of clusters based on their impact on the phenotype. This adaptive, probabilistic clustering maintained flexibility while reducing dimensionality. We also developed methods for clustering of hospitals in terms of their distribution of patient outcomes. Previous approaches for clustering focus on a certain aspect of the distribution, such as the mean outcome for a typical patient. Our approach instead groups hospitals having identical aspects of all features of the distribution. This is important, since two hospitals can have the same mean, while having very different patient outcomes among the sickest and healthiest patients. This article was selected at the best paper to be published in the theory and methods section of the Journal of the American Statistical Association next year, and will be highlighted in a special session at the annual Joint Statistical Meeting. We have also developed flexible modeling frameworks that allow distributions of health outcomes to vary nonparametrically with predictors. This allows one to assess differential effects on the most sensitive individuals in a population instead of assuming a treatment effect or adverse effect of an environmental exposure is identical for all individuals. We have used related ideas to assess the impact of weight gain trajectories during pregnancy on maternal and childhood outcomes, such as obesity.