OBJECTIVES:

To demonstrate the use of Longitudinal Targeted Maximum Likelihood Estimation (LTMLE) methods to derive causal average treatment effect estimates for survival and health-related quality of life (measured by the EQ5D) outcomes from longitudinal real-world data with risk of bias from immortal-time and time-varying confounding.

METHODS:

LTMLE is a double-robust method that accounts for time-varying confounding by modelling both treatment and outcome mechanisms, and produces an unbiased estimate of the causal treatment effects if either one of the mechanisms are correctly specified. Machine Learning algorithms can be used to estimate components of the treatment and outcome mechanisms, to increase the likelihood of correct model specification while retaining valid statistical inference. We apply LTMLE to data from the European Myelodysplastic Syndromes (EUMDS) registry and estimate causal average treatment effects of erythropoiesis-stimulating agents (ESA) for intermediate-1 to low-risk Myelodysplastic Syndromes patients. To account for the challenge of repeated measurements, long follow up and relatively small sample, we restrict the treatment and outcome models to only use a maximum of or two lags of covariates.

RESULTS:

Accounting for time-varying confounding changes the predicted intervention-specific mean (counterfactual) outcomes and causal average treatment effects compared to a naïve analysis. We find no statistically significant effect of using ESA on patients' EQ5D scores or on cumulative survival. These results hold with and without using machine learning for the treatment and outcome models.

CONCLUSIONS:

This study demonstrates the appropriate use of longitudinal causal methods in studying the treatment effect of therapies under sustained exposure, accounting for immortal-time and time-varying confounding risk of bias, which are usually neglected in these analyses. The challenges highlighted in the paper provide a lesson for future analyses that attempt to apply LTMLE in complex real-world data settings, especially in the case of a small sample size with a long follow-up period.