Why Temporal Leakage Matters
In time-varying survival data, one person may contribute several follow-up intervals, with covariates recorded at different times. An imputation model for a missing covariate in an earlier interval can appear to fit well while using information that was not yet available then. For example, a value measured after an event might be used to fill in an exposure from before the event. If the event changed subsequent treatment or measurement, the later value can partly encode the event itself. The resulting imputation may misrepresent the earlier history.
Whether a later measurement is inappropriate depends on the purpose of the imputation. For a prospective prediction or an analysis requiring temporally valid histories, predictors must respect the time at which the target covariate is meant to be known. For some retrospective analyses, a later measurement can be a useful auxiliary predictor if its use is scientifically justified and compatible with the imputation assumptions. The central task is to make that target and the information cutoff explicit rather than treating every recorded value as interchangeable.
The cutoff is not automatically the end of a row in a start-stop data set. It depends on what the missing covariate represents and when it would be measured or used. If the target is a covariate known at the start of an interval, predictors recorded later in that interval are not available at the start. If the target is a value measured during the interval, its measurement time determines which information precedes it. State the intended ordering before specifying the imputation predictors.
Formulating an Interval-Specific Predictor Set
Let \(X_{ik}\) denote a covariate for person \(i\) in interval \(k\), and let \(t_{ik}\) be the information cutoff for imputing that value. A temporally ordered imputation model has the form
where \(Z_i\) denotes other predictors that are valid at the cutoff, such as baseline characteristics. The notation describes the information used to draw a missing value; it does not prescribe a particular distribution or fitting method. The model should also reflect the covariate type and the structure of repeated measurements. A continuous covariate, for instance, requires an imputation model suitable for its distribution and substantive role.
A practical predictor set may include baseline variables, prior observed or imputed covariate values, valid history summaries, and time indicators. It excludes measurements known to occur after the cutoff when the target is a prospective or temporally ordered history. Event status and survival-derived predictors require a separate judgment: such information can be useful in retrospective imputation for an analysis of the observed cohort, but it is not information available for a prediction made before the event. This distinction is not a rule that outcomes must always be omitted; it is a requirement to match the predictors to the estimand and use case.
State which person-interval value is missing and what time that value represents.
Identify the latest information legitimately available for imputing that target value.
For each predictor, check its measurement time, whether it is defined for that person, and whether an event could have changed it.
Use admissible predictors that also support a plausible imputation model; review the resulting histories for temporal contradictions.
Why Future Values Change the Conditional Imputation
A sequential history model represents the joint distribution through successive conditional distributions. For two repeated values, the chain rule gives
This factorization makes the forward temporal ordering explicit. Imputing an earlier value from a later one instead uses a reverse conditional distribution, such as \(P(X_{i1}\mid X_{i2})\). That reverse conditional is mathematically well-defined; it is not automatically a valid model for a prospective information history. If the later value has been affected by an event, treatment change, or selective follow-up, conditioning on it can make the earlier imputation reflect those later processes.
There is an important qualification. Retrospective imputation may intentionally use later observed repeated measures to predict an earlier missing value. That can be reasonable when the target is a retrospective completed data set, later values are genuinely observed, and the joint model and missingness assumptions support the use of those values. Temporal leakage is a failure when the model is meant to represent information available at an earlier time, or when later values are treated as if they were unaffected and valid measurements of the earlier covariate. The date order alone does not settle the question; the target and assumptions do.
Worked Example: A Post-Event Measurement Is Not an Earlier History Value
Consider a hypothetical patient with a covariate \(X\) measured at baseline, at month 2, and at month 6. The month-2 value is missing. An event occurs at month 4, after which treatment is changed. The recorded values are \(X_{i0}=4\) at baseline and \(X_{i6}=12\) at month 6. Suppose the target is the value at month 2, before the event.
For a temporally ordered imputation of the month-2 value, baseline \(X_{i0}\), baseline characteristics, and other information available by month 2 may be predictors. The month-6 value is not available by the month-2 cutoff, and it is measured after the event and treatment change. It therefore does not belong in the prospective history for this target.
The month-6 value might still be considered in a retrospective imputation model, but only with an explicit justification: the question would be whether it helps impute a missing earlier value in a completed historical data set, not what could have been known at month 2. Treating \(12\) as if it were simply another pre-event measure ignores the possibility that post-event treatment changed \(X\).
Worked Example: Quantifying the Effect of an Invalid Predictor
Suppose, purely for illustration, that a fitted imputation equation for an earlier covariate value is
For a patient with an observed prior value \(X_{i1}=4\), the prediction using the valid history is
Now suppose someone adds the later month-6 value to the equation, using the hypothetical fitted equation \(\widehat{X}_{i2}=2+0.6X_{i1}+0.5X_{i6}\). With \(X_{i6}=12\), the result is
The additional \(6\) units in the prediction come entirely from the future measurement. These equations are illustrative, not estimates from study data; their coefficients are assumed for the calculation. The arithmetic shows how strongly a later value can influence an earlier imputation, but does not establish that either prediction is statistically unbiased. In this scenario, the post-event value may reflect the treatment change, so the second equation does not represent a month-2 information history.
Worked Predictor Audit Across Intervals
The following hypothetical records show why the same predictor can be admissible for one target and inadmissible for another. Suppose the analysis uses intervals ending at months 3 and 6. A covariate is intended to describe exposure during each interval, and the imputation target is the exposure known at the start of each interval.
| Interval target | Information cutoff | Candidate predictor | Decision for temporally ordered imputation |
|---|---|---|---|
| Exposure for months 0–3 | Start of month 0 | Baseline age | Include if relevant and observed |
| Exposure for months 0–3 | Start of month 0 | Exposure recorded at month 3 | Exclude: recorded after the cutoff |
| Exposure for months 3–6 | Start of month 3 | Exposure from months 0–3 | Potentially include: prior history |
| Exposure for months 3–6 | Start of month 3 | Exposure recorded at month 6 | Exclude: recorded after the cutoff |
Worked Example: Building Predictors for Two Imputation Targets
In the hypothetical table, a patient has age 60 at baseline, exposure 5 in months 0–3, and a missing exposure for months 3–6. The value recorded at month 6 is 9. The target is exposure at the start of months 3–6, so the cutoff is month 3.
Age 60 is baseline information and the earlier exposure value 5 is history available by month 3. Both can be considered as predictors. The value 9 is recorded at month 6, later than the cutoff, so it is excluded from this temporally ordered imputation. A suitable model might use the prior exposure and age, for example \(X_{i,2}\sim P(X_{i,2}\mid X_{i,1},\text{age}_i)\), with its distribution chosen for the covariate and its data.
If the patient experienced an event before month 6, the month-6 value may additionally be affected by that event or by changes in care. This strengthens the need to distinguish it from the month-3 history. It does not, by itself, prove that month-6 data can never help in a retrospective imputation; it means that such a use requires a different target and a defensible model.
Diagnostics, Alternatives, and Limitations
Temporal leakage is often a design problem rather than a numerical warning from a fitted model. A model can converge and produce plausible-looking imputations while using invalid predictors. Audit the data structure and predictor list directly:
- Check timestamps and definitions. Confirm when each covariate was measured and whether it represents a value before, during, or after the interval.
- Flag post-event values. Determine whether a candidate predictor was collected after an event or a treatment change that could alter it.
- Review histories person by person. Look for imputations that imply a future state was already present earlier without supporting measurements or a defensible model.
- Compare model specifications. If appropriate to the analysis, compare imputations from a temporally restricted model with a justified retrospective model. Large changes indicate sensitivity to future information, not proof that one model is correct.
- Validate at the intended prediction time. For a prospective use, do not evaluate a procedure with predictors that would be unavailable at that time.
The main alternatives serve different purposes. An interval-specific sequential model respects a prospective information history but may be less statistically efficient if later measurements contain useful information. A retrospective repeated-measures model may use later values as auxiliary predictors, but requires assumptions that justify their relationship to the missing earlier value and must account for event-related changes. Omitting all time-varying predictors avoids one form of leakage but may discard important history and weaken the imputation model. Last observation carried forward is not a general solution: it substitutes a fixed prior value rather than representing uncertainty or plausible within-person change.
No predictor rule can repair an ill-defined target, a covariate that is undefined after an event, or an unsuitable imputation model. A value that does not exist for a person-time should not be treated as an ordinary missing measurement merely because a rectangular data table expects an entry. Also, excluding future values may be inappropriate when the stated retrospective target and imputation assumptions specifically call for their use. The defensible choice follows from the scientific question, the measurement process, and the analysis model.
Common Mistake
Check Your Understanding
Answer each question using the stated information cutoff. Brief solutions follow the questions.
- A covariate for months 0–2 is missing, and its information cutoff is the start of month 0. A measurement is available at month 5. Is it admissible for a prospective, temporally ordered imputation? Why?
- A missing month-4 value is imputed for a retrospective completed data set. Is a month-8 measurement automatically forbidden? State the key qualification.
- For the missing exposure at the start of months 3–6, the observed exposure for months 0–3 is 7 and baseline age is 50. The cutoff is month 3. Name two candidate predictors and one value that should be excluded if it was measured at month 6.
- Suppose the event occurs at month 4 and treatment changes afterward. Why is a month-6 covariate value a particularly concerning predictor of a missing month-2 value?
- A colleague proposes excluding every time-varying predictor to prevent leakage. What important limitation does that choice introduce?
Solutions. (1) No. Month 5 is after the start-of-month-0 cutoff, so it is future information for this prospective target. (2) Not automatically. It may be useful if the retrospective target, measurement process, and imputation assumptions justify conditioning on that later value; it must not be presented as information available at month 4. (3) The month 0–3 exposure and baseline age are candidates because both precede the month-3 cutoff. Exclude the month-6 measurement for a temporally ordered imputation. (4) The month-6 value may reflect the event-related treatment change, so it could encode a process that occurred after the month-2 target and distort its imputation. (5) It may discard informative prior history, weakening the imputation model and failing to represent within-person change. The better approach is to select temporally valid predictors, not to remove all time-varying information.