Abstract
Background: Left-truncation refers to the unrecorded interval between the onset of multiple sclerosis (MS) and the initial data entry in observational studies. This delay in study entry has the potential to bias estimates of disease-modifying therapy (DMT) effectiveness, especially when it is determined by patient and disease characteristics.
Objective: To test the hypothesis that causal effect estimates of DMTs over the full course of the disease can be reliably derived from left-truncated registry data.
Methods: We analysed MSBase data from 144 centres in 41 countries to assess how left-truncation affects causal treatment estimates. Cox marginal structural models (MSMs) estimated hazard ratios for relapses and disability outcomes under random and non-random truncation. To remediate informed left-truncation bias, we truncated data at 2 or 3 years from MS onset and applied multivariable adjustment. We also tested the robustness of MSMs using deliberately mis-specified treatment weight models.
Results: We studied 5,588 MS patients from disease onset. Without truncation, DMTs reduced relapse risk (HR=0.64; 95%CI: 0.54–0.77). Left-truncation inflated this effect, especially with shorter random truncation (1-year HR=0.36), but bias lessened with longer durations (2-year HR=0.55; 3-year HR=0.66). Not-at-random left-truncation also biased relapse estimates (HR=0.37). Disability outcomes were less affected. Fixed-time truncation (2 or 3 years) with multivariable adjustment effectively reduced left-truncation bias in estimating treatment effects.
Conclusion: MSMs can reliably estimate DMT effectiveness in left-truncated MS data. Bias from left-truncation, random or not, is greater with shorter durations and lessens over time. Disability outcomes are less affected. Fixed-time truncation and covariate adjustment reduce bias. MRI exclusion did not bias estimates.