What is landmark analysis?
A way of comparing groups defined by something that happens after baseline, without giving one group a period in which it could not have the outcome.
Landmark analysis is a way of answering questions about an exposure that is only known some time after follow-up begins — whether a patient responded to treatment, whether they received a transplant, whether a biomarker fell — without introducing immortal time bias.
The method is simple. Choose a time after baseline — the landmark. Exclude anyone whose outcome occurred before it. Classify everyone still at risk by what had happened to them by the landmark, and start the clock there. Because exposure is fixed using only information available at the landmark, and follow-up begins at the landmark, nobody is credited with time during which their outcome could not have occurred.
The cost is real, and it is the reason the landmark has to be chosen on clinical grounds rather than by trying several. Everyone who had the outcome before the landmark is discarded, so a late landmark answers a question about a progressively healthier group. An early landmark keeps more people but misclassifies those whose exposure status changed shortly afterwards. The estimate is conditional on surviving to the landmark, which is a different quantity from the one a trial would report, and it should be described that way.
The alternative is to treat the exposure as what it is: something that changes over time. A model with a time-dependent covariate uses each patient's exposure as it actually evolved, and does not throw away the early events. That is usually the better answer when the timing itself is informative; landmark analysis earns its place when the question genuinely is “among those still alive and event-free at this point, what happens next?”