Statistical Methods for Correlated Data in Health Research

我的研究计划的总体目标是开发能够对统计实践产生实际影响的统计方法。今后五年的具体目标如下：1)制定复杂的观察员间研究和可靠性研究的方法，并取得两种结果。在我的这部分研究中，我将开发同时评估观察员之间的一致性和观察员内的可靠性的方法。将渐近结果的有限样本性能与Bootstrap进行比较。这些结果将被用来分析几项正在进行的大型研究，我是这些研究的联合调查员。2)开发社区干预试验的方法和事件发生时间的结果。这些试验的一个决定性特征是，它们通常涉及少量的大集群。然而，可以安全地应用于此类试验的方法并未得到很好的开发。我将建议对渐近方法进行小样本调整。我还将开发计算密集型方法，如Bootstrap。3)双结果整群随机试验中组内相关系数的可信区间程序的建立。研究人员经常使用以前文献中的点数估计，而没有充分认识到与这些估计相关的固有的精确度不足。一个适用于具有少量大群集的试验的可信区间程序将具有很高的实用价值。我将根据我之前为大量小星系团开发的结果开发新的程序。还将开发针对由整群随机化试验产生的二元结果的新相关性参数的可信区间程序。4)发展具有两个结果的整群随机试验的因果中介分析方法。虽然中介分析是阐明干预如何起作用的自然工具，但在整群随机试验的背景下，几乎没有做过任何研究工作。我将利用反事实框架开发相关的统计方法。

The general goal of my research program is to develop statistical methods that can have a real impact on statistical practice. Specific objectives for the next five years are as follows: 1) Development of methods for complex inter-observer and reliability studies with binary outcomes. In this part of my research, I will develop methods for concurrent assessment of both inter-observer agreement and intra-observer reliability. The finite sample performance of asymptotic results will be compared to the bootstrap. These results will be used to analyze several ongoing large studies for which I am a co-investigator. 2) Development of methods for community intervention trials with time-to-event outcomes. A defining feature of these trials is that they often involve a small number of large clusters. However, methods that can be safely applied such trials are not well developed. I will propose small sample adjustments to asymptotic methods. I will also develop computationally intensive approaches such as the bootstrap. 3) Development of confidence interval procedures for the intraclass correlation coefficients in cluster randomization trials with binary outcomes. Investigators frequently use point estimates from previous literature without fully appreciating the inherent lack of precision associated with these estimates. A confidence interval procedure that is applicable for trials with small number of large clusters would have high practical value. I will develop new procedures based on my previous results developed for large number of small clusters. Confidence interval procedure for new dependence parameters for binary outcomes arising from cluster randomization trials will also be developed. 4) Development of methods for causal mediation analysis of cluster randomization trials with binary outcomes. Although mediation analysis is a natural tool for illuminating how an intervention works, virtually no research work has been done in the context of cluster randomization trials. I will develop relevant statistical methods using the counterfactual framework.