Adjusting for unmeasured confounding in nonrandomized longitudinal studies: a methodological review

Abstract

OBJECTIVE: Motivated by recent calls to use electronic health records for research, we reviewed the application and development of methods for addressing the bias from unmeasured confounding in longitudinal data. DESIGN: Methodological review of existing literature SETTING: We searched MEDLINE and EMBASE for articles addressing the threat to causal inference from unmeasured confounding in nonrandomised longitudinal health data through quasi-experimental analysis. RESULTS: Among the 121 studies included for review, 84 used instrumental variable analysis (IVA), of which 36 used lagged or historical instruments. Difference-in-differences (DiD) and fixed effects (FE) models were found in 29 studies. Five of these combined IVA with DiD or FE to try to mitigate for time-dependent confounding. Other less frequently used methods included prior event rate ratio adjustment, regression discontinuity nested within pre-post studies, propensity score calibration, perturbation analysis and negative control outcomes. CONCLUSIONS: Well-established econometric methods such as DiD and IVA are commonly used to address unmeasured confounding in non-randomised, longitudinal studies, but researchers often fail to take full advantage of available longitudinal information. A range of promising new methods have been developed, but further studies are needed to understand their relative performance in different contexts before they can be recommended for widespread use.