The Handling of Missing Data Induced by Time-Varying Covariates in Comparative Effectiveness Research Involving HIV Patients
Support for This Research
This is a Patient Centered Outcomes Research Institute (PCORI)-sponsored project to improve methods for analyzing comparative effectiveness studies
Antiretroviral Therapy in HIV Patients and Cardiovascular Disease
Advances in antiretroviral therapy (ART) have dramatically reduced mortality from HIV, enabling reclassification of HIV as a chronic condition. Numerous studies suggest that some drugs increase the risk of cardiovascular disease, although findings are inconsistent. Studies differ largely due to methodological choices, including study design, definition of exposure, and approaches to handling missing data. It is crucial to incorporate information on drug exposure and other confounders over time; patients vary their regimens over time and for reasons that may be related to their condition. Thus, in order to not implicate the wrong drug, information on changes in regimen and other factors must be considered.
Studies with Time-Varying Confounders Can Induce a Missing Data Problem
Including information over time complicates the analysis, however. One such complication is the introduction of missing data. Common methods for handling missing data yield misleading descriptions of relationships. Appropriate methods for handling missing data are computationally burdensome; software does not exist for many situations and the analyst must rely on his/her own programming skills to implement specialized techniques. Multiple imputation (MI) is a reasonably accessible and theoretically sound method for handling missing data. Available in mainstream software, its special application is required due to the unique issues posed by time-varying covariates and outcomes that are only partially observed for those individuals who do not experience a cardiovascular event during the observation period. We propose an extensive simulation study to evaluate commonly applied methods to this setting, to investigate the performance of standard MI in this context, and to adapt and evaluate MI methods utilized in a longitudinal setting where the outcome is fully observed to this particular setting. Based on our findings we will develop concrete guidelines on how to use MI in the context of partially observed outcomes and time-varying covariates. We will develop user-friendly open-source software in order to optimize the use of recommended methods and to eliminate lack of software as a barrier to employing missing data methods. Finally, we will illustrate methods considered on data from the US veteran population of HIV-infected individuals using the Veterans Health Administration's rich longitudinal Clinical Case Registry (CCR), the analysis of which motivated this proposal. This work has the potential to greatly impact patients living with HIV. Currently, there is no consensus on which ART agents increase cardiovascular risk.
Our proposal will address the implications of methodological choices for handling missing data when conducting comparative effectiveness research in the longitudinal setting. Importantly, the development of guidelines will unify analyses enabling combination of evidence across studies in the form of meta-analyses, and accessibility to software will eliminate barriers to incorporating missing data methods into analyses.
Aims for Our Proposal
Aim 1: Develop methods to simulate correlated time-varying covariates and right-censored outcomes
Aim 2: Characterize variability in common missing data methods applied in this context
Aim 3: Develop new methods for handling missing data for these studies
Aim 4: Provide open-source software for implementing recommended methods
Update on Our Findings
To evaluate methods for handling missing data in studies of comparative effectiveness research (CER), we relied on simulations. This was challenging, as commonly CER studies involve multiple correlated time-varying covariates and a right-censored outcome. We described practical guidelines for generating right-censored outcomes as a function of time-varying covariates for CE studies (Montez-Rath et al., 2017), and provided a publically available novel simulation tool for simulating entire CE studies that reflect real complexities of CER studies (Montez-Rath et al., Submitted to SIM in 2017). This work is essential for evaluating a variety methods – not only missing data methods -- relevant for CER. Our simulations have demonstrated that an improved passive multiple imputation (MI) approach is ideal for handling interaction terms that are categorical (Mitani et al., 2016). In addition, when rate of change is the outcome of interest in longitudinal studies, active MI of the outcome under a two-stage slope model can provide large benefits over a mixed effects model without MI particularly when the data are not missing at random (Desai et al., 2016). We found considerable heterogeneity across findings among commonly applied techniques (Desai et al, Submitted to Observational Studies in 2016). In particular, our simulations demonstrated that standard MI of time-varying covariates provides advantages over typical approaches under certain conditions even when the data are not missing at random. Concrete guidelines for performing MI in CER studies with time-varying covariates and right-censored outcomes as well as an illustration for how to interpret findings from disparate sensitivity analyses will be submitted in February 2017.