On the Statistical Limitations of Landmark Analysis for Addressing Immortal Time Bias

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Landmark analysis is a common method to address immortal time bias in observational studies.
• It aims to create comparable groups by excluding early events and restarting follow-up from a fixed landmark time.

Purpose of the Study:

• To evaluate the validity and limitations of landmark analysis in estimating causal effects.
• To compare landmark analysis with other methods under potential outcomes and piecewise proportional hazards frameworks.

Main Methods:

• Derivation of closed-form results for naive, exclusion, and landmark analyses.
• Evaluation of these methods using simulations under various scenarios.
• Comparison of statistical power with randomized trials.

Main Results:

• Landmark analysis estimates are biased due to mixing 'never-treated' and 'late-treated' subjects, attenuating effect sizes and biasing towards the null.
• The method suffers from significant loss of statistical power from excluding early events and effect attenuation.
• Bias magnitude depends on true effect, treatment, and event rates, except under the null hypothesis.

Conclusions:

• Landmark analysis may not recover true causal effects and can be misleading, especially for small treatment effects.
• Caution is advised when using landmark analysis for immortal time bias.
• Time-dependent regression models offer a more robust and reliable alternative for causal effect estimation.