Why These Errors Matter
Survival and time-varying data encode more than a list of numerical values. A recorded follow-up time may end in an event or in censoring; one person may contribute many time intervals; and a covariate measured later may not describe the person’s earlier state. An imputation model that ignores these features can create records that look complete but misrepresent the observed information or the intended analysis.
This tutorial focuses on four errors that can undermine chained-equation imputation: treating a censored follow-up time as an event time, omitting cumulative-hazard information when imputing a missing covariate, treating repeated rows for a participant as independent cases, and using future covariates without a defensible temporal interpretation. Each error should be checked against the target estimand and the structure of the substantive analysis.
This notation assumes right censoring and that the event and censoring times are recorded on a common time scale. Standard survival analyses additionally require appropriate handling of the study’s sampling and risk-set structure. For the usual independent-censoring interpretation, censoring must be independent of the event time conditional on the variables included in the model. If this condition is implausible, the censoring process may need to be modeled or otherwise addressed; merely changing the imputation predictors does not guarantee valid inference.
Error 1: Treating a Censored Time as an Event
A frequent coding error is to use \(Y_i\) as if it were \(T_i\) for every participant. That is correct only for participants with \(\Delta_i=1\). For a censored participant, \(Y_i\) is the end of observed follow-up, not the time at which the event occurred. Assigning an event at that time invents information and changes the outcome record.
When the goal is to impute a missing covariate, event information can help predict that covariate. But the outcome information supplied to the imputation model must preserve the distinction between events and censoring. When the goal is to impute an event time itself, a censored observation supplies a lower bound on that time, not an observed event time; the draw must respect \(T_i>Y_i\).
Worked Example: A Follow-Up Time Is Not Necessarily an Event Time
Suppose a hypothetical participant has \(Y_i=8\) months and \(\Delta_i=0\). Another participant has \(Y_j=5\) months and \(\Delta_j=1\). What event information is observed for each?
For participant \(i\), \(Y_i=8\) means follow-up ended at 8 months, and \(\Delta_i=0\) means the event was not observed by then. Thus the record supports \(T_i>8\) months, not \(T_i=8\) months. For participant \(j\), \(\Delta_j=1\), so the event was observed at \(T_j=Y_j=5\) months.
If an analysis or imputation file instead assigns an event indicator of 1 to participant \(i\), it creates a false event at month 8. If it assigns the event time \(T_i=8\) without retaining the censoring indicator, it likewise treats an unknown time as known. The corrected survival record retains the pair \((Y_i,\Delta_i)=(8,0)\).
Error 2: Omitting Cumulative-Hazard Information
For imputing a covariate associated with survival, a model that includes event status alone may discard important information about how long each participant was followed. In a Cox proportional-hazards model, that information is represented in part by cumulative hazard up to the observed time. A commonly used survival summary is the estimated cumulative baseline hazard at \(Y_i\), denoted \(\widehat{\Lambda}_0(Y_i)\), together with \(\Delta_i\). The exact imputation model should still reflect the substantive model, covariates, and design.
The justification follows from the event-time likelihood. Under a proportional-hazards model with hazard \(\lambda_i(t)=\lambda_0(t)\exp(\beta x_i)\), the event-time likelihood contribution, apart from terms not involving \(x_i\), is proportional to
Taking its logarithm gives
The event indicator appears in the event term, while cumulative hazard at the observed follow-up time appears in the survival term. Consequently, \(\Delta_i\) alone does not generally carry the same information about \(x_i\) as the pair \(\{\Delta_i,\Lambda_0(Y_i)\}\). In practice, \(\Lambda_0\) is usually estimated, for example using a Nelson–Aalen cumulative-hazard estimate. This is a model-based motivation for including cumulative-hazard information in a covariate imputation model, not a guarantee that these two summaries alone make every chained-equation model compatible with every survival analysis.
Worked Example: Why Follow-Up Duration Adds Information
Consider two hypothetical participants whose event indicators are both 0. Suppose the estimated cumulative baseline hazards at their observed follow-up times are 0.10 and 0.60, respectively. Compare what is retained by using only event status with what is retained by also including cumulative hazard.
Using event status alone gives \(\Delta_1=\Delta_2=0\), so the two records look identical on that predictor. Adding cumulative hazard distinguishes them: the first has \(\widehat{\Lambda}_0(Y_1)=0.10\), while the second has \(\widehat{\Lambda}_0(Y_2)=0.60\). Under the Cox likelihood, the survival-term factors for a given covariate value \(x\) are \(\exp\{-0.10\exp(\beta x)\}\) and \(\exp\{-0.60\exp(\beta x)\}\).
For illustration, if \(\beta x=0\), these factors are \(\exp(-0.10)=0.9048\) and \(\exp(-0.60)=0.5488\). Both participants remained event-free through their observed follow-up, but the second record contains information from a longer period of cumulative hazard. The numerical difference does not establish a particular covariate value; it shows why reducing both records to \(\Delta=0\) loses information relevant to the survival likelihood.
Error 3: Treating Repeated Rows as Independent Participants
Time-varying data are often stored in start–stop or person-period form, with one row per participant per interval. Such rows can be necessary: each row can represent the covariate values and risk status during a particular interval. The error is not having multiple rows. The error is treating those rows as independent participants when fitting or imputing a participant-level variable, or when calculating uncertainty, without accounting for their shared participant.
Rows from the same participant share characteristics and often share the same missing value. Duplicating a participant-level value across intervals can also make participants with more intervals contribute more weight than participants with fewer intervals. The consequence depends on the model and data layout: estimates may be reweighted, uncertainty may be understated, or both. Do not assume that a repeated-row file is equivalent to a one-row-per-person file.
Worked Example: Unequal Row Counts Change the Weighting
Suppose a hypothetical data set has two participants, with a participant-level binary covariate \(X\). Participant A has \(X=0\) and four interval rows; participant B has \(X=1\) and one interval row. The person-level mean is
If the five rows are incorrectly treated as five independent participants, the row-weighted mean is
The difference is not a small-sample correction: the row-level calculation answers a different weighted question, giving participant A four times the representation of participant B. A valid person-period model may appropriately use all five rows to represent time at risk, but it must preserve participant identity and use a structure suitable for within-participant dependence. For a participant-level imputation, the target and unit of analysis are the participant, not the number of interval rows.
Before fitting a model to long data, verify which variables vary within a participant, which are repeated copies of one participant-level value, and whether interval rows represent distinct risk periods. Check participant identifiers, the number and duration of intervals per participant, and whether event status is coded only in the appropriate interval. Remedies depend on the task: impute a participant-level variable once per participant and carry it consistently across that participant’s records, or use an imputation model that explicitly accommodates the multilevel or longitudinal structure. A standard independent-row model is not made valid merely by including a participant identifier as an ordinary numeric predictor.
Error 4: Using Future Covariates Without a Temporal Justification
For a time-varying covariate \(X_i(t)\), the value used to describe exposure or status at time \(t\) must correspond to the intended time point. Using a measurement made after \(t\) to fill in the value at \(t\) can leak future information into the past. In a risk-set or prediction setting, this can make a participant’s earlier history depend on information that was not yet available, and can distort associations between covariates and event timing.
This is a diagnostic warning, not a rule that future measurements can never help impute a missing earlier measurement. In a retrospective analysis, a later measurement might be a defensible predictor of an earlier missing value if a model for their relationship is justified and the target estimand permits it. The key distinction is between using later data as a predictor for a clearly defined retrospective imputation and assigning that later value as though it had been observed earlier. The latter changes the covariate history.
Worked Example: A Later Measurement Cannot Automatically Stand In for an Earlier One
A hypothetical participant has an exposure measured at months 1 and 3, but the month-1 value is missing. The observed month-3 value is 10 units. Suppose the month-1 value was 4 units in a hypothetical complete record, but that value is unavailable to the analyst. What can be concluded from the month-3 measurement alone?
Nothing in the observed record establishes that the month-1 value was 10. The exposure may have changed between months 1 and 3. Replacing the missing month-1 value with 10 asserts \(X_i(1)=X_i(3)\), a constancy assumption that must be justified rather than silently imposed.
For example, if the exposure were known to be constant over that interval by design, using 10 could be appropriate. If it can change, the month-1 value should instead be imputed using a model appropriate to the measurement schedule and covariate history, while preserving its time index. In a month-1 risk-set analysis, month-3 information should not simply be treated as an exposure already present at month 1.
A Practical Error-Check
A useful review follows the data from its meaning to its model implementation. The checks below are intended to identify structural errors before interpreting imputation results.
For each participant, verify that \(Y_i\) is the observed event or censoring time and that \(\Delta_i\) distinguishes the two. Confirm that only event records are treated as observed event times.
For missing-covariate models tied to a survival analysis, check whether event status and suitable cumulative-hazard information are represented. Confirm that estimated hazard summaries are evaluated at the correct observed time and account for delayed entry when relevant.
Count unique participants as well as rows. Identify repeated participant-level values and determine whether the imputation and analysis models account for within-participant dependence and unequal interval counts.
For each time-varying predictor, record when it was measured and which interval it represents. Confirm that the completed covariate history does not silently assign later values to earlier times.
Check that the imputed data retain valid event/censoring coding, participant identity, time ordering, and the substantive model’s required variables. Investigate implausible transitions, duplicated records, or intervals outside observed follow-up.
These checks diagnose data and model-structure problems; they do not prove that the imputation model is correctly specified. Estimates can still be sensitive to the assumed survival model, the censoring mechanism, sparse events, or poorly measured covariates. In particular, cumulative-hazard summaries are estimated quantities, and a convenient predictor set may not reproduce every feature of a complex time-varying substantive model.
Choosing a Remedy and Recognizing Limits
| Approach or error | What it retains | Main caution |
|---|---|---|
| Use \(Y_i\) as an event time for everyone | Follow-up length only | Invents events for censored participants; retain \(\Delta_i\). |
| Use event status alone as survival information | Whether an event was observed | Can omit follow-up information represented by cumulative hazard. |
| Fit an independent-row model to repeated records | All interval rows | Can change participant weighting and ignore within-person dependence. |
| Use a later covariate as an earlier value | A convenient observed value | Assumes temporal constancy or otherwise changes the covariate history. |
| Use a survival-aware or longitudinal imputation model | Can preserve censoring and repeated-measure structure | Requires suitable model assumptions, diagnostics, and alignment with the substantive analysis. |
Complete-case analysis avoids making these particular imputations, but it can lose information and may be biased when the cases with complete covariates are not representative for the target analysis. Simple single imputation of a censored time or copying a later measurement is not a general substitute: it can encode false certainty or incorrect timing. A joint model or a carefully specified chained-equation approach may better represent survival and longitudinal structure, but neither is automatically valid. The appropriate choice depends on the estimand, data-generating assumptions, measurement schedule, and substantive model.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A participant has \(Y=12\) months and \(\Delta=0\). What is known about the event time?
- Two censored participants both have \(\Delta=0\), but their estimated cumulative hazards at follow-up are 0.2 and 0.8. Why might an imputation model retain both quantities?
- A participant-level binary variable is copied onto six interval rows for one person and one row for another. What can go wrong if all rows are treated as independent cases?
- When could a later covariate measurement be used to help impute an earlier missing value, and what would be an unjustified use?
- Name two checks that can reveal time-varying data structure errors before interpreting the imputed analyses.
Solutions. (1) The event was not observed by month 12, so the record indicates \(T>12\) months; it does not identify the event time as 12. (2) Event status records whether an event was observed, while cumulative hazard at follow-up carries information about the duration and event-free survival contribution; the two records are not equivalent in the survival likelihood. (3) The first participant may receive excess weight, and ordinary independent-row uncertainty calculations may ignore dependence among that person’s rows. (4) A later measurement may be a predictor in a justified retrospective model if the target permits it; simply assigning it as the earlier value assumes it represents the earlier state and can leak future information into an earlier risk period. (5) Examples include counting unique participants and rows, checking event status is coded in the correct interval, reviewing measurement times against interval definitions, and verifying that repeated participant-level values remain consistent.