Landmark Analysis for Conditional Survival and Dynamic Prediction
Also known as: landmark method, dynamic prediction, conditional survival estimation, Landmark Analizi (Dinamik Tahmin)
Landmark analysis, introduced by Anderson, Cain, and Gelber in 1983, estimates conditional survival probabilities for subjects who are still at risk at a pre-specified point in time — the landmark — rather than at study entry. It was developed explicitly to avoid immortal time bias that arises when subjects are grouped by an event (such as a treatment change or biomarker result) that can only occur if they remain event-free long enough to experience it.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use landmark analysis when you want to estimate survival conditional on having already survived to a given time point, or when your exposure or treatment group is defined by an event that occurs after study entry (which would otherwise create immortal time bias). It is well-suited to situations where a biomarker value, treatment switch, or response classification is only observable in survivors. A sensitivity analysis across several candidate landmark times is recommended to assess how sensitive conclusions are to the choice of t_LM. When prognostic information needs to be updated continuously throughout follow-up, the superposition model or a joint model for longitudinal and survival data should be considered as more flexible alternatives.
Strengths & limitations
- Directly prevents immortal time bias by restricting analysis to subjects who have already survived to the landmark.
- Easy to explain and implement: it is essentially a standard survival analysis run on a filtered subset with a shifted time origin.
- Allows covariate information accumulated up to the landmark to be used as predictors, enabling conditional prognosis tailored to each patient's trajectory.
- The superposition extension supports dynamic, repeatedly updated predictions without requiring a full joint model.
- The choice of landmark time is subjective; results can vary materially across different choices, making sensitivity analyses mandatory.
- Restricting analysis to survivors at the landmark can sharply reduce the effective sample size, leading to wide confidence intervals.
- Subjects who experience the event before the landmark are discarded from the conditional analysis, so landmark analysis does not describe early events.
- The standard single-landmark approach assumes that the landmark time is fixed and known in advance; post-hoc selection of the most favourable landmark inflates false-positive risk.
Frequently asked
What is immortal time bias and how does landmark analysis prevent it?
Immortal time bias occurs when subjects in one group must survive a period before they can be classified into that group — making the period between study entry and classification 'immortal' for those subjects. Landmark analysis fixes this by restricting all comparisons to subjects still alive at a single pre-specified time point, so no one in any group has an unacknowledged survival advantage built in.
How do I choose the landmark time?
The landmark time should be chosen on substantive grounds before analysing the data — for example, the time at which a treatment response assessment typically occurs, or the median time to the defining biomarker event. It should then be varied in a sensitivity analysis (e.g., t_LM ± 3 months) to show that conclusions are robust to the exact choice.
Can I include covariates in a landmark analysis?
Yes. Any covariate measured up to and including the landmark time can be used as a predictor in the model fitted to the post-landmark follow-up. However, variables that are themselves outcomes of the post-landmark period must not be used as predictors, as this would reverse the causal direction.
When should I use a joint model instead of landmark analysis?
If you need to update survival predictions continuously as new longitudinal measurements arrive — rather than at a fixed time point — a joint model for longitudinal and time-to-event data is more appropriate. The superposition landmarking model is a pragmatic middle ground: it uses multiple fixed landmark times and stacks them to approximate dynamic prediction without the complexity of a full joint model.
Sources
- Anderson, J. R., Cain, K. C. & Gelber, R. D. (1983). Analysis of Survival by Tumor Response. Journal of Clinical Oncology, 1(11), 710–719. DOI: 10.1200/JCO.1983.1.11.710 ↗
- van Houwelingen, H. C. (2007). Dynamic Prediction by Landmarking in Event History Analysis. Scandinavian Journal of Statistics, 34(1), 70–85. DOI: 10.1111/j.1467-9469.2006.00529.x ↗
How to cite this page
ScholarGate. (2026, June 1). Landmark Analysis for Conditional Survival and Dynamic Prediction. ScholarGate. https://scholargate.app/en/survival/landmark-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Joint Model for Longitudinal and Survival DataSurvival↔ compare
- Kaplan-MeierSurvival↔ compare
- Nelson-Aalen EstimatorSurvival↔ compare