Why Visit Order Matters
A repeated covariate is not a single missing entry detached from time. A laboratory value recorded at an early visit and one recorded months later describe different points in a person's follow-up. When an intermediate value is missing, the measurements on either side can help predict it, but they do not make it known. A useful imputation must preserve which visits occurred, which values were observed, and when the person experienced an event or was censored.
In chained equations, repeated measurements can be represented as separate visit-specific variables. Their conditional models can then use measurements at adjacent visits, along with relevant baseline and survival information. Updating these variables in chronological order gives the imputation procedure a transparent temporal structure. It does not, by itself, guarantee a valid model or a biologically plausible trajectory.
The visit order is an organizational and modeling choice, not a claim that the data were generated by a one-way process. For an interior missing value, both the previous and next eligible measurements may be informative. The imputation model may condition on both, if that is appropriate for the analysis and the measurements exist. Conversely, a measurement at a visit after a person's event cannot serve as an observed neighbor merely because that visit appears in the study-wide schedule.
Set Up the Repeated-Measure Imputation
Before fitting equations, specify the visit schedule, the time associated with each recorded value, and the endpoint of observation for each person. Write \(Y_i=\min(T_i,C_i)\) for observed follow-up time, where \(T_i\) is event time and \(C_i\) is censoring time, and let \(\delta_i=1\) indicate an observed event and \(\delta_i=0\) censoring. The study protocol must determine how a covariate measurement recorded at or near \(Y_i\) is classified. For example, a value documented before an event may be eligible, while a nominal visit after the event is not. Do not infer a measurement's timing from its row position alone.
For each visit-specific variable \(X_j\), specify a conditional model for its missing entries. A generic model can condition on baseline predictors \(B_i\), the previous and next eligible repeated measurements, and survival predictors such as \(\delta_i\) and a suitable cumulative-hazard summary. The earlier material on survival predictors in covariate imputation explains why outcome-related information can be important; here the key additional decision is whether each proposed predictor belongs to the person's observed timeline.
This expression is schematic. The neighboring terms are included only when those visits are eligible and when the conditional model is suitable for the covariate type. At the first or last eligible visit, one neighbor may not exist. A categorical covariate needs a categorical conditional model; a continuous covariate may need transformations, nonlinear time effects, or a distribution appropriate to its scale. The model should also reflect irregular gaps between visits when a one-visit separation represents different elapsed times for different people.
As with other MICE applications, the chained equations require defensible conditional models, appropriate predictors, and stochastic draws that reflect uncertainty in model parameters and missing values. The missingness assumption must be reasonable conditional on information in the imputation models; MICE does not automatically solve nonignorable missingness. Ideally, the conditional models are compatible with a joint distribution. In practice they may not be, so the resulting imputation distribution can depend on model specification and update order. Chronological updating is a considered strategy, not proof of compatibility.
A Chronological Update Scheme
A typical sweep proceeds from the earliest visit to the latest. At each visit, fit or update its conditional model using observed values and the current completed values of predictors, then draw replacements for its missing entries. Later visit models can use the newly updated values from earlier visits. On the next sweep, the equations are revisited, so the procedure is not simply a one-pass fill from left to right.
For each person, mark scheduled measurements that are within observed follow-up according to the study's event and censoring measurement rules.
For each incomplete visit variable, specify which adjacent values, baseline variables, and survival summaries can enter its conditional model.
Draw missing values at visit 1, then visit 2, continuing through the latest eligible visit variables in the sweep.
Run enough sweeps to assess stabilization, then inspect distributions, temporal patterns, and event-time eligibility across the imputed data sets.
Chronological ordering can make predictor availability and the imputation process easier to audit. It is particularly clear when an earlier missing value is updated before it becomes a predictor for a later visit. However, if a model for an interior visit also uses the next visit, it conditions on information later in calendar time. That can be appropriate for imputing a missing value in a completed longitudinal data set, provided the next measurement was actually observed within follow-up and the analysis permits it. This is different from asserting that future measurements were available at the time of a prospective decision.
Worked Example: Two Missing Visits in a Chronological Sweep
Suppose a hypothetical biomarker is measured at days 0, 30, and 60, and all three visits occur before each participant's observed follow-up endpoint. For one participant, the day-0 value is 10 units, the day-30 value is missing, and the day-60 value is 16 units. For illustration, assume the fitted day-30 conditional model is \(X_{i2}\mid X_{i1},X_{i3}\sim N(0.5X_{i1}+0.5X_{i3},\,2^2)\). The model and variance are assumed given for this arithmetic example; they are not estimated from this one participant.
The conditional mean is \(0.5(10)+0.5(16)=5+8=13\) units. Suppose the random-number draw for this imputation is \(z=0.40\), and the predictive draw uses \(X_{i2}=13+2z\). Then \(X_{i2}=13+2(0.40)=13.8\) units. This is a stochastic imputation, not a deterministic average of the neighboring values.
Now suppose a later visit at day 90 is also scheduled and missing. If its fitted model is \(X_{i4}\mid X_{i3}\sim N(X_{i3}+0.02(90-60),\,1.5^2)\), the updated day-60 value is available in the chronological sweep. Its conditional mean is \(16+0.02(30)=16.6\) units. With a separate illustrative draw \(z=-0.50\), the imputed day-90 value is \(16.6+1.5(-0.50)=15.85\) units. The draw uses the model's uncertainty; it does not require the trajectory to increase at every visit.
In a real MICE analysis, model parameters would also be drawn or otherwise handled so that parameter uncertainty is represented. The numerical equations here isolate the role of visit order and adjacent information.
Worked Timing and Gap Decisions
Worked Example: An Event Before a Scheduled Neighbor
Consider a hypothetical participant scheduled for covariate measurements at days 0, 4, 8, and 12. An event occurs at day 7, and the observed day-4 value is 5.2 units. The day-8 and day-12 measurements are absent. Assume the protocol records covariates only while a participant is event-free and has no pre-event measurement documented at day 8.
The eligible observed visits are days 0 and 4; days 8 and 12 are not observed covariates for this participant. The missing day-8 entry must not be treated as an ordinary missing covariate value to be predicted from a nonexistent post-event measurement. It is outside the participant's observed covariate follow-up under the stated protocol. The imputation data should represent the observed at-risk intervals and endpoint consistently, rather than create a post-event covariate record.
If a day-8 value had instead been collected and timestamped before the event at day 7, that timing would be contradictory and must be resolved from the data definition, not repaired by an imputation equation. If the protocol records a measurement at the event visit before the event, it may be eligible under that explicit rule. The decision is about measurement timing, not about a value's plausibility.
Worked Example: Accounting for Unequal Visit Gaps
Suppose a hypothetical continuous marker is observed at day 10 with value 7.0 units, and is missing at day 40. A simple model fitted for this illustration is \(E(X_{ij}\mid X_{i,j-1},\Delta t)=X_{i,j-1}+0.03\Delta t\), where \(\Delta t\) is the elapsed time in days. The conditional mean at day 40 is \(7.0+0.03(40-10)=7.0+0.90=7.90\) units.
If the next eligible visit is day 100 rather than day 40, the corresponding mean under the same illustrative model is \(7.0+0.03(100-10)=7.0+2.70=9.70\) units. Applying a fixed one-visit adjustment to both gaps would ignore that the elapsed times differ by 60 days. In an analysis, the time relationship and error distribution must be estimated and checked; these hypothetical calculations show why visit index alone may not capture temporal spacing.
Diagnostics, Alternatives, and Limits
Check both the imputation algorithm and the implied repeated-measure data. Trace summaries across sweeps can reveal unstable chains or persistent drift. Compare observed and imputed distributions separately by visit, and inspect individual trajectories for abrupt jumps, impossible values, or excessive repetition. For irregular schedules, examine patterns against elapsed time, not only visit number. Review missingness and imputation counts by visit and by event/censoring status. In particular, verify that no imputed measurement is assigned to a visit excluded by the protocol's endpoint rule.
A model may fail even when the visit ordering is correct. Warning signs include sparse observed data at a late visit, separation in a categorical model, implausible extrapolation beyond observed ranges, strong sensitivity to whether the next visit is included, or markedly different imputations under alternative sweep orders. Such results call for revisiting predictors, functional forms, visit-specific samples, or the assumptions behind using neighboring values. Do not hide a structural absence of post-event measurements by coding it as ordinary item nonresponse.
Visit-specific chained equations are one option, not the only one. A long-format model can represent time through a common longitudinal conditional model, while a joint longitudinal model specifies a shared distribution for a person's repeated measurements. These alternatives can encode temporal dependence more directly, but require their own modeling assumptions and may be less convenient for mixed variable types or a broad predictor set. Simple interpolation uses neighboring measurements to construct a value but does not, by itself, provide proper stochastic imputations or incorporate survival predictors. Deterministic carry-forward is also not equivalent to chained imputation and should not be substituted merely because a neighbor is available.
No ordering scheme can compensate for weak data support, a misspecified conditional model, or missingness that depends on unobserved values in ways not addressed by the analysis. Chronological updating is most useful as a disciplined way to make visit-specific conditioning explicit and to preserve the measurement timeline. It is not a guarantee that imputed trajectories reproduce the true biological process.
Common Pitfall
Check Your Understanding
Answer each question, then compare with the solutions below.
- A continuous covariate is missing at an interior visit, with eligible observed values of 4 and 10 at the immediately preceding and following visits. Under a stated conditional mean of \(0.4X_{i,j-1}+0.6X_{i,j+1}\), what is the mean before adding a random imputation draw?
- A person's event occurs before a scheduled covariate visit, and the protocol records no measurements after an event. Should that scheduled visit be treated as an ordinary missing covariate value?
- Why might a model include elapsed time between visits rather than only the visit number?
- What does updating visit variables in chronological order guarantee, and what does it not guarantee?
Solutions. (1) \(0.4(4)+0.6(10)=1.6+6=7.6\) units; this is the conditional mean, not the imputed value itself. (2) No. Under the stated protocol it is outside the participant's observed covariate follow-up, not an ordinary missing measurement to be filled. (3) Visit gaps may differ in duration, so the same one-step change in visit index can represent different elapsed times; a time-aware model can represent that difference. (4) It specifies a transparent update sequence from earlier to later visit variables. It does not guarantee compatible conditional models, correct imputations, a causal temporal process, or valid handling of measurements beyond event follow-up.