Continuous-Time Multiple Imputation That Knows Why Tests Were Ordered

Hospital records are the raw material for retrospective cohort studies, which reuse past patient data to estimate how well a treatment worked. In these records, lab values and vital signs are measured at irregular times and are missing most of the time. The timing is not random. A doctor orders a lactate because the patient looks worse. Today, epidemiologists usually handle this in one of two ways. They squeeze the data into hourly or daily bins and fill the gaps with last-value-carried-forward or with multiple imputation by chained equations (MICE, which fills each column in turn by regression on the others). Or they hand the raw stream to a neural model that estimates effects directly but reports no standard errors that a biostatistician can use.

We propose a multiple-imputation engine that works in continuous time. It draws $M$ complete, smooth patient trajectories from a latent dynamical model. That model is trained jointly on the measured values and on when measurements were taken, so the ordering pattern itself informs what the unmeasured values probably were. Each completed cohort then goes through a standard, unchanged causal analysis, and the results are pooled with Rubin's rules, the textbook formulas that combine the within-imputation and between-imputation variance. We judge success by whether the pooled 95% confidence intervals for a known treatment effect actually cover it about 95% of the time. None of the continuous-time imputation or continuous-time causal papers we found measure this.

Core Claim

A continuous-time imputer yields valid multiple imputations for longitudinal causal analysis only if it has three properties. It must model the measurement-time process as a function of the latent patient state. It must include treatment and outcome as observed inputs, so the imputation model is compatible with the analysis model. It must carry uncertainty about its own parameters rather than just the noise of a single fitted model.

Existing continuous-time imputers have none of these. Drawing samples from their latent posterior should produce intervals that are too narrow and biased whenever test ordering tracks deterioration. Binned tabular MICE can have the compatibility property but loses timing information. We predict it will be close to our method when sampling is uninformative and will fail when sampling depends on the patient's state.

A useful side effect is that imputation happens in continuous time, so the analyst's choice of time grid becomes a separate decision from how the data were sampled. The same $M$ trajectories can be evaluated on a 1-hour or a 6-hour grid without re-imputing.

Relation to Prior Work

Two continuous-time imputers supply the machinery but stop short of multiple imputation.

  • Continuous-time Autoencoders (CTA) encodes a series into a continuous latent path using neural controlled differential equations (a recurrent model whose hidden state is driven by an interpolated path through the observations). It reports only point-reconstruction error, on non-clinical data. - Neural Continuous-Discrete State Space Models (NCDSSM) runs a Bayesian filter whose latent state evolves continuously between observations and is updated when they arrive. It reports forecasting error of 4.74 MSE on CMU walking and 0.34 on USHCN, but produces no completed datasets.

We reuse the NCDSSM design as our backbone because its explicit filtering and smoothing distributions make posterior draws natural. We differ from both in what is modeled (observation times, treatment, outcome) and in what is evaluated (coverage after pooling).

Modern Multiple Imputation with Functional Data is the closest multiple-imputation work. It shows that column-wise MICE produces jagged, unrealistic curves, and that binned MissForest reaches 0.28 downstream RMSE versus 0.38 for MICE on the PaTH EHR cohort. Its binning discards exact timestamps and it has no causal component. It motivates the problem and serves as a baseline.

MatchThem supplies the downstream protocol we reuse: impute, then balance within each imputed dataset, then pool with Rubin's rules. It is limited to static tables.

On the causal side, three papers estimate effects over time but none imputes or pools:

  • TE-CDE uses adversarial balancing on irregular data and explicitly leaves informative sampling only partially handled.
  • SCIP-Net uses stabilized inverse propensity weighting in continuous time.
  • IGC-Net uses iterative G-computation on discrete, pre-imputed inputs.

These papers constrain our design. We deliberately do not propose a new causal estimator. We keep the downstream analysis standard, using stabilized marginal structural models and the parametric g-formula via gfoRmula, so that any improvement can be attributed to the imputation step.

Two papers show that timing is informative. Neural ODEs for Informative Missingness reaches 0.9147 mortality AUC on PhysioNet 2012 by exploiting it. SLAN argues that imputation should be avoided altogether. We take that argument seriously but answer it from the cohort-study side, where complete data matrices are required.

Research Program

The imputer has three parts.

Latent state. A latent patient state $z(t)$ evolves under a learned continuous-time stochastic dynamic. Treatment enters as an input that changes the dynamics.

Measurement values. Each measured value $x_{k,i}$ of variable $k$ at time $t_{k,i}$ is emitted from $z(t_{k,i})$.

Measurement times. Each variable's measurement times follow a point process, meaning a model of when events occur, whose intensity $\lambda_k(t \mid z(t))$ depends on the latent state. The training objective (an evidence lower bound) therefore contains the point-process log-likelihood:

$$\log p = \sum_{k,i} \log p(x_{k,i} \mid z(t_{k,i})) + \sum_{k,i} \log \lambda_k(t_{k,i} \mid z) - \sum_k \int_0^T \lambda_k(t \mid z)\,dt,$$

Here $T$ is the end of follow-up. The last two terms reward the model for predicting when tests happen, so a burst of lactate orders pulls the smoothed estimate of unmeasured lactate toward abnormal values. The outcome and treatment history are observed channels of the same model, which makes the imputation model compatible with the analysis model.

Drawing imputations. Each of the $M$ imputations is a full trajectory drawn from the smoothing posterior. The parameters are drawn from a bootstrap ensemble: the model is refit on resampled patients, one fit per draw, which approximates the "proper" imputation Rubin's theory requires. Each trajectory is evaluated on the analyst's chosen grid, analyzed, and pooled with Rubin's rules.

Limits of identification. Because the intensity depends on the unobserved current state, this is a shared-latent selection model. Identification of the not-at-random part rests on model structure, not on the data alone. We therefore add a sensitivity parameter that scales the intensity's dependence on $z$, report how the effect estimate moves as it varies, and state openly that no imputation method can remove this assumption.

Baselines

Naive reference. The do-nothing reference is the common practice itself: hourly binning via MIMIC-Extract, last-observation-carried-forward, and a single analysis with the usual model-based standard errors. It is the realistic default, and any method that cannot beat its coverage is not worth adopting.

Strongest statistical baseline. This is binned MICE with time-since-last-measurement and measurement-count features, treatment, and outcome included, pooled via the MatchThem workflow. Alongside it we run binned MissForest (MF_B) from Modern Multiple Imputation with Functional Data, whose reported result is 0.28 RMSE on PaTH smoking-relapse prediction against 0.38 for MICE. These are the strongest competitors because they are what careful epidemiologists would actually do. Adding count features gives MICE a crude version of our timing signal.

Strongest neural causal baseline. This is SCIP-Net run directly on the irregular data without imputation, with standard errors from a patient-level bootstrap so it is comparable. The file reports no numeric RMSE for SCIP-Net ("no reported result found on Lacuna for SCIP-Net semi-synthetic MIMIC-III RMSE"). IGC-Net reports up to a 26.7% relative RMSE improvement over prior methods on semi-synthetic MIMIC-III. We include IGC-Net on binned data as a second neural baseline.

Fair comparison. All imputers receive the same patients, the same $M = 20$ draws, the same downstream estimator, and an equal hyperparameter-search budget in GPU-hours. MICE gets the same number of chained-equation iterations it needs to converge.

Ablations. Four ablations isolate the key ingredients:

  • plain CTA posterior samples;
  • our model with the intensity term removed;
  • our model without the bootstrap over parameters;
  • our model without treatment and outcome channels.

What counts as winning. Beating a baseline means 95% interval coverage within 92–97% together with lower absolute bias, in the state-dependent sampling regime, while not losing coverage in the uninformative regime.

Why not just…?

Why not just use MICE with time-gap features and MatchThem? This is the real threat. When sampling is missing-at-random given the observed history, binned MICE with timing features may match us, and we expect it to. The experiments cross the missingness mechanism with bin width to find where it breaks. Our prediction is that it breaks when ordering depends on the current unmeasured value, and when coarse bins merge the value measured before a treatment with the value measured after it. If MICE matches our coverage in the state-dependent regime too, the honest conclusion is that continuous-time imputation is unnecessary for cohort studies, and we will report that.

Why not just draw samples from CTA's latent posterior? CTA was evaluated only on point accuracy. A single fitted model conditioned only on covariates omits both the uncertainty in its own parameters and the treatment–outcome relationship. Rubin's between-imputation variance should then come out too small, and effects should be biased toward the null. The CTA ablation tests this directly.

Why not skip imputation and use SCIP-Net or TE-CDE, as SLAN argues? These methods give effect estimates, but not pooled datasets that feed standard hazard or marginal structural model analyses. TE-CDE itself leaves informative sampling unaddressed. We compare bias and bootstrap coverage head-to-head. If SCIP-Net achieves equal coverage at lower cost, imputation earns its place only through compatibility with standard statistical pipelines, and we will say so.

Proposed Experiments

Semi-synthetic MIMIC-III (primary test). We use the 25-vital semi-synthetic MIMIC-III design shared by IGC-Net and SCIP-Net, which gives a known true effect. We simulate a time-varying treatment with confounding strength $\gamma$, then mask the data under four mechanisms:

  • completely at random;
  • dependent on observed history;
  • dependent on the current true value through a Hawkes-style intensity, a point process in which each event raises the short-term rate of further events, following TE-CDE's sampling design;
  • treatment-dependent monitoring, with treated patients measured ten times as often.

Across 200 replicates we report bias, empirical coverage, and interval width. The decisive result is coverage under value-dependent masking.

PK-PD tumor benchmark. The same comparison on this simulated tumor-growth model, which uses Hawkes-process sampling, checks that the result is not specific to MIMIC-III.

Imputation calibration. On PhysioNet 2012 we hide 20% of observed values and check whether the imputers' predictive intervals cover the hidden values at their nominal rate.

Real cohort question. On MIMIC-III we estimate an effect related to vasopressor timing. There is no ground truth, so we report agreement among methods, the stability of the estimate across the 1-hour and 6-hour analysis grids, and the sensitivity curve over the intensity-dependence parameter.