Missing Data in Marginal Structural Models

Abstract

Background: The use of marginal structural models (MSMs) to adjust for time-varying confounding has increased in epidemiologic studies. However, in the setting of MSMs, recommendations for how best to handle missing data are contradictory. We present a plasmode simulation study to compare the validity and precision of MSMs estimates using complete case analysis (CC), multiple imputation (MI), and inverse probability weighting (IPW) in the presence of missing data on time-independent and time-varying confounders. Methods: Simulations were based on a cohort sub-study using data from the Osteoarthritis Initiative which estimated the marginal causal effect of intra-articular injection use on yearly changes in knee pain. We simulated 81 scenarios with parameter values varied on missing mechanisms (MCAR, MAR, and MNAR), percentages of missing (10%, 20%, and 30%), type of confounders (time-independent, time-varying, either or both), and analytical approaches (CC, IPW, and MI). The performance of CC, IPW, and MI methods was compared using relative bias, mean squared error (MSE) of the estimates of interest, and empirical power.

Addressing missing data in the estimation of time‐varying treatments in comparative effectiveness research

“…Second, this article addresses missing data in both outcomes and confounders. Previous studies addressing missing data in time‐varying settings have focused on missingness in the confounders 10‐13 . As with our study, those studies suggested that both IPW and MI provide minimal biases in estimating time‐varying treatment effects.…”

“…CER studies that use longitudinal, observational data are likely to be affected by time‐varying confounding and missing data 10‐13 . Appropriately addressing these two major issues is a necessary step for obtaining unbiased ATE estimates.…”

“…MI has received far less attention in studies with time‐varying confounding, and its relative merits in this setting are not yet well understood. Previous methodological studies have explored MI and IPW for addressing missing data in studies with time‐varying confounding 10‐13 . However, their focus has been on missing data in the confounders and monotone missing data patterns, with most studies considering a single data pattern and missing variable.…”

Statistical plasmode simulations–Potentials, challenges and recommendations

“…Further, an appropriate OGM has to be applied to generate outcomes based on the resampled covariates. 5,7,[25][26][27][28][29][30][31][32][33][34] Statistical plasmodes have often been utilized in causal (propensity-based) methods using weighted regression models. The application of these weighted regression models is often related to the presence of a complex covariance structure.…”

simstudy: Illuminating research methods through data generation

“…The SimDesign Package (Chalmers & Adkins, 2020) and the related MonteCarlo Package (Leschinski, 2019) follow a similar line of thought but focus on easy replication of the analyses and providing summaries of simulated data. simstudy has been used in a variety of fields for theoretical exploration of research methodology (Anderson, Wennberg, & McMullen, 2019;El Alili et al, 2020;Kirasich, Smith, & Sadler, 2018;Krzykalla, Benner, & Kopp-Schneider, 2020;Liu, Chrysanthopoulou, Chang, Hunnicutt, & Lapane, 2019;Nickodem, 2020;Thoya, Waititu, Magheto, & Ngunyi, 2018;Wang & Ma, 2020), power calculation for trials (Wei et al, 2019) and other simulation tasks supporting researchers (Chukwu, 2019;Forthun et al, 2020;Horry, Fitzgerald, & Mansour, 2020;Renson, Schubert, & Bjurlin, 2017).…”