Towards more dynamic modeling of high-dimensional register-based data

简介：我们的目标是开发统计降维技术，以控制药物安全性观察性研究中潜在混杂因素的高维空间。这些数据通常是通过药物处方或报销和住院的常规行政报告系统产生的。这类研究面临的挑战是，常规数据收集没有针对研究目的进行优化，因此往往缺乏直接解决混杂因素的变量，因此，通常使用大量可用的指标来代替潜在的混杂因素。提取混杂信息的传统方法在很大程度上是基于变量选择技术，而我们认为通过特征选择可以更好地解决这个问题。然而，为此目的需要发展专门的统计方法。**长期研究目标：除了高维度外，通过常规报告系统获得的数据具有连续时间的特点。为了适应这一点，我们将研究降维方法的动态推广，以控制因时变个人水平特征引起的混淆。这就要求将降维作为时间的连续平滑函数进行建模。**最新进展：由于与暴露（处方药）相比，所关注的不良后果通常很少，因此通过暴露/处方建模而不是通过结果建模来控制混淆是可取的。两种众所周知的方法是倾向评分调整/匹配或治疗逆概率（IPT）加权；我们开发了前者的连续时间变体和后者的贝叶斯版本。如果暴露模型只包含预测暴露的变量，则效应估计的方差可能会被夸大，这促使研究将暴露模式建模与降维相结合。**未来五年的短期目标：我们将研究针对特定结果和暴露变量量身定制的降维，以捕获由高维协变量空间造成的最大混淆量，并开发其估计方法。此外，我们将开发推断程序来估计暴露的影响基于使用估计的维数减少结合倾向得分调整和IPT加权。由于记录的协变量是大型事件，而不是连续值测量，因此估计降维需要基于似然的方法，包括贝叶斯方法。**影响：通过基于人群的观察性研究监测药物治疗的长期安全性是最可行的。通过我们的研究，我们的目标是促进更有效地利用行政卫生保健数据库作为监测药物安全的资源。这类观测数据的一个主要来源是加拿大通过省级卫生保健计划实施的普遍公共卫生保健系统。作为省级卫生保健数据库及其利用的一个例子，我们提到魁北克的RAMQ和MED-ECHO数据库以及加拿大观察性药物效应研究网络（CNODES）的合作框架。我们研究的目的是产生直接适用于此类研究的统计方法。

Introduction: We aim to develop statistical dimension reduction techniques for controlling for confounding in observational studies of drug safety with a high-dimensional space of potential confounders. Such data are commonly generated through routine administrative reporting systems on drug prescription or reimbursement and hospitalizations. The challenge in such studies is that the routine data collection has not been optimized for research purposes, thus often lacking variables to directly address confounding, and as a result, a large number of available indicators potentially informative of the underlying confounders are typically used instead. The conventional approaches to extract confounder information have largely been based on variable selection techniques, whereas we argue that the problem is better addressed through feature selection. However, development of specialized statistical methodology is needed for this purpose.**Long-term research objectives: In addition to high dimensionality, data coming through routine reporting systems are characterized by continuous time. To accommodate this, we will investigate dynamic generalizations of the dimension reduction approaches to control for confounding due to time-varying individual-level characteristics. This calls for modeling of the dimension reductions as continuous smooth functions of time.**Recent progress: Since the adverse outcomes of interest are often rare compared to the exposures, which are prescription medications, it is preferable to approach controlling for confounding through modeling of the exposure/prescription rather than through modeling of the outcome. The two well known methods for this are propensity score adjustment/matching or inverse probability of treatment (IPT) weighting; we have developed a continuous-time variation of the former and a Bayesian version of the latter. If the exposure model includes variables only predictive of the exposure, variance of the effect estimate may be inflated, motivating research into combining modeling of the exposure patterns with dimension reduction.**Short-term objectives for the next five years: We will investigate dimension reductions tailored to specific outcome and exposure variables to capture maximal amount of confounding accounted for by the high-dimensional covariate space, and develop methodology for their estimation. Furthermore, we will develop inference procedures for estimating exposure effects based on using the estimated dimension reductions in conjunction with propensity score adjustment and IPT weighting. Because the recorded covariates are by and large events, rather than continuous-valued measurements, estimation of the dimension reductions requires likelihood-based, including Baysian, approaches.**Impact: Monitoring the long-term safety of drug therapies is most feasible through observational population-based studies. Through our research, we aim to contribute to more efficient utilization of administrative health care databases as a resource for monitoring drug safety. A major source for such observational data is the universal public health care system in place in Canada, implemented through provincial health care plans. As an example the provincial health care databases and their utilization, we mention the RAMQ and MED-ECHO databases of Quebec and the collaborative framework of the Canadian Network for Observational Drug Effect Studies (CNODES). The objective of our research is to produce statistical methodology directly applicable in such studies.