Landmark analysis
A landmark analysis fixes a time point, uses only information available by then as predictors, and studies the outcome after it. It avoids using the future to predict the future.
Definition
In a landmark analysis, a time point (the landmark) is fixed in advance. Only patients still at risk at the landmark are included, only information available up to the landmark is used as predictors, and the outcome is followed from the landmark onward.
Why it is used
A tempting analysis relates a patient's best response, wherever it occurred, to survival from the start of treatment. That uses information that was not available at the time of prediction, and patients who survive longer have more opportunities to have a response recorded. This produces immortal-time bias: the association looks stronger than anything that could be used at the landmark.
Example
The analysis of early tumor change and survival fixes a landmark at week 16, uses only earlier tumor measurements, and evaluates the added value on held-out data. In 128 patients at the 16-week landmark, the C-index rose from 0.532 to 0.581, but the interval for the improvement included zero.
Related
The tumor-kinetic model analysis asks whether a model adds anything beyond a single early tumor change at a landmark, and the dropout analysis asks how patients who stop being measured affect the estimate.
Frequently asked questions
What is immortal-time bias?
A bias that arises when a patient must survive long enough to be classified as exposed or as a responder, so the classification itself uses survival time that is then counted as follow-up. A landmark analysis avoids it by classifying patients only on information up to a fixed time.
How is the landmark time chosen?
It should be fixed in advance by the clinical question, for example the time of the first planned tumor assessment, and not chosen after looking at the results; sensitivity to the choice should be checked.