Building a Landmark Dataset from Longitudinal Records

From Visit Records to Landmark Records

Longitudinal records are often stored with one row per visit, while event records usually have one row per patient. Landmark analysis needs a different unit of analysis: a patient considered at a particular landmark time. Reshaping the records requires more than joining tables. Each resulting row must identify whether the patient is eligible at that time, which measurement history was available then, and what happened afterward.

The key boundary is the landmark. Information recorded after it cannot be used as a predictor for that landmark, and follow-up for its outcome begins at that landmark. Applying both rules consistently avoids using future information to predict an earlier risk.

The strict inequality adopts the convention that an event at the landmark makes a patient ineligible at that landmark. If event and landmark timestamps can coincide, the analysis protocol must state their ordering; the data-processing rule must implement that convention consistently. A patient censored on or before the landmark is also ineligible because event-free status through the landmark cannot be established from the observed record.

A row should keep three kinds of information distinct: eligibility at the landmark, predictor history available by the landmark, and outcome follow-up after the landmark. The predictor history can include the visit times and measurements available up to \(s_\ell\), or a precomputed representation of that history. The choice of representation is a separate modeling decision. Whatever is stored, it must be based only on information available by the landmark.

Constructing a Patient-Landmark Row

Begin with a prespecified list of landmark times and a consistent time origin. For each patient and landmark, determine eligibility from event and censoring records. For eligible pairs, select only visits at or before the landmark, then attach the subsequent event and censoring follow-up using elapsed time from the landmark. Retaining explicit eligibility and follow-up fields makes the construction auditable.

Define Post-Landmark Follow-Up Carefully

For an eligible pair, let the post-landmark observed follow-up time be \(R_{i\ell}=\tilde T_i-s_\ell\), and let the event indicator be \(\Delta_{i\ell}=\Delta_i\). Since eligibility requires \(\tilde T_i>s_\ell\), \(R_{i\ell}\) is positive. These fields describe residual follow-up from the landmark, rather than time from enrollment.

These definitions assume a single time scale for events and censoring, a consistently defined event of interest, and correctly linked patient records. The construction itself does not make censoring non-informative. For example, interpreting survival estimates from the rows generally requires appropriate assumptions about censoring; the data reshaping cannot make those assumptions true. If a different event type or a competing event is relevant, its outcome coding must be specified rather than silently treating it as ordinary censoring.

For a fixed horizon \(h>0\), the event indicator by that horizon is known if an event occurs within \(h\), or if the patient is observed event-free for at least \(h\). Censoring before \(h\), without an earlier event, leaves the binary horizon outcome unknown. This distinction matters because labeling every such patient “no event” would incorrectly treat incomplete follow-up as evidence of remaining event-free.

Why the Row Definition Matches the Landmark Question

At a landmark, the target population consists of patients who have remained event-free through that time. If \(T\) denotes event time and \(S(t)=P(T>t)\), the conditional probability of an event by \(s+h\), given survival beyond \(s\), is

The first equality in the calculation uses the event “event-free at \(s\) and event by \(s+h\),” which is \(s<T\le s+h\). Thus the row must represent a patient known to be event-free at \(s\), and its outcome follow-up must begin at \(s\). Including a patient who had the event before the landmark, or treating time since enrollment as if it were residual time, would not represent this conditional risk.

This argument clarifies the data structure, not the assumptions needed to estimate the risk. In particular, observed follow-up may end through censoring, and the handling of censoring must be appropriate for the analysis. The landmark row records what was observed; it does not resolve incomplete follow-up by itself.

Multiple Landmarks and Alternative Data Structures

A landmark dataset differs from a one-row-per-patient dataset because a patient may be considered at more than one landmark. It also differs from the original visit-level file: visit rows record measurements over time, whereas a landmark row represents the patient’s status and available history at a specified prediction time. A joint longitudinal-survival analysis uses the longitudinal and event records in a different modeling structure; converting to landmark rows does not preserve every feature of that model.

If the same patient contributes eligible rows at several landmarks, those rows are not independent observations. The data may be arranged as one row per eligible patient-landmark pair, but later analyses must account for the actual sampling and dependence structure as appropriate. Conversely, collapsing multiple landmarks to a single patient row changes the set of prediction times and should not be done merely to make the table simpler.

Diagnostics, Common Errors, and Limits

After construction, check the dataset against its source records. Useful checks include confirming that no row has an event or censoring time at or before its landmark, that predictor visits do not occur after the landmark, and that residual follow-up equals the observed end time minus the landmark. Also verify that the event indicator agrees with the source event record and that each eligible patient-landmark pair appears exactly once.

A particularly common mistake is to label a patient with no recorded event by the horizon as a non-event, even when censoring occurred before the horizon. That reasoning confuses “no event observed” with “observed event-free through the horizon.” The correction is to mark the horizon status unknown when follow-up ends too early, unless a suitable analysis explicitly handles the censoring.

• Future-information leakage: A visit after the landmark is included because it appears in the same patient file. Exclude it; availability in the source data does not make it available at prediction time.
• Wrong time origin: Follow-up is stored from enrollment, then interpreted as time after the landmark. Subtract the landmark time for each eligible row.
• Unverified eligibility: A patient censored before the landmark is treated as event-free at the landmark. Exclude that candidate pair because status through the landmark is not known.
• Duplicate rows: A many-to-many join between visit and event files replicates patient-landmark records. Check pair uniqueness after joining and retain the visit history separately or summarize it according to the analysis plan.

The resulting records are only as reliable as their source times, event definitions, and censoring information. They do not, on their own, decide how to represent a patient’s biomarker history, how to handle missing measurements, or how to fit a prediction model. Those are distinct choices, and a well-constructed dataset makes them explicit rather than hiding them in a join or a recoding step.