A multiple regression imputation method with application to sensitivity analysis under intermittent missingness

Missing data is a common problem in general applied studies, and specially in clinical trials. For implementing sensitivity analysis, several multiple imputation methods exist, like sequential imputation, which restricts to monotone missingness, and Bayesian, where the imputation and analysis models differ, entailing overestimation of variance. Also, full conditional specification provides a conditional interpretation of sensitivity parameters, requiring further calibration to get the desired marginal interpretation. We propose in this paper a multiple imputation procedure, based on a multivariate linear regression model, which keeps compatibility in sensitivity analysis under intermittent missingness, providing a marginal interpretation of the elicited parameters. Simulation studies show that the method behaves well with longitudinal data and remains robust under demanding constraints. We conclude the possibility of situations not covered by the existing methods and well suited for our proposal, which allows more efficient handling of a given multivariate linear regression structure. Its use is illustrated in a real case study, where a sensitivity analysis is accomplished.

缺失数据是各类应用研究中的普遍问题，在临床试验（clinical trials）中尤为突出。为开展敏感性分析（sensitivity analysis），现有多种多重插补（multiple imputation）方法：例如限制于单调缺失模式（monotone missingness）的序贯插补（sequential imputation）法，以及插补模型与分析模型存在差异、会导致方差高估的贝叶斯插补法。此外，全条件设定（full conditional specification）法可对敏感性参数（sensitivity parameters）给出条件层面的解释，但需进一步校准才能获得预期的边际解释（marginal interpretation）。本文提出一种基于多元线性回归模型（multivariate linear regression model）的多重插补流程，该方法在间歇性缺失（intermittent missingness）场景下的敏感性分析中保持兼容性，并可对所导出的参数给出边际层面的解释。仿真研究表明，该方法在纵向数据（longitudinal data）场景下表现良好，且在严苛约束条件下仍具备稳健性。我们得出结论：存在现有方法未能覆盖的场景，而本文提出的方法恰好适配这类场景，能够更高效地处理给定的多元线性回归结构。最后通过一则实际案例研究演示了该方法的应用，并完成了敏感性分析。