Treatment Benefit Before Dropout and Reference Behavior Afterward

One Trajectory, Two Different Sources of Information

When a participant assigned to an active treatment drops out, the outcomes recorded before dropout remain outcomes under the treatment the participant received. A reference-based imputation rule does not rewrite that observed history. Instead, it specifies how to predict the participant’s unobserved outcomes after dropout. Keeping these two parts separate is essential: the first is observed information; the second depends on an assumption.

This distinction matters when describing treatment benefit over follow-up. An active-arm advantage at visits before dropout is part of the observed data, subject to sampling variability. A reference-like trajectory after dropout is not a later observation showing that the advantage disappeared. It is an assumption used to construct a distribution for missing outcomes and, ultimately, to conduct a sensitivity analysis.

Definition: Let \(Y_t\) denote the outcome at visit \(t\), with \(A\) denoting the active arm and \(R\) the reference arm. For an active-arm participant whose last observed visit is \(s\), the observed history is \(Y_{0:s}=(Y_0,\ldots,Y_s)\), and the post-dropout outcomes are \(Y_{s+1:K}=(Y_{s+1},\ldots,Y_K)\). A reference-based assumption retains the observed history and specifies a predictive distribution for the post-dropout outcomes using a reference-arm model or rule.

The superscript or arm label on an observed outcome describes the participant’s randomized treatment assignment; it does not say that every later outcome is observed. For instance, an active-arm participant can contribute observed active-arm outcomes through visit \(s\), followed by imputed outcomes whose distribution is defined using reference-arm behavior. The imputed segment is not a claim that the participant switched treatment in fact.

What “Reference Behavior Afterward” Means

Under jump to reference (J2R), the post-dropout predictive distribution is based on the reference-arm distribution, conditional on the observed history. Thus, the rule does not normally mean “replace every missing value by the reference-arm mean.” A participant’s observed history can be higher or lower than the reference-arm history, and the reference model’s conditional prediction uses that information. The exact conditional prediction also depends on the fitted means and covariance structure.

For example, suppose the reference-arm model for a pair of visits has mean vector \((\mu_{R,1},\mu_{R,2})\) and covariance matrix with entries \(\sigma_{R,11},\sigma_{R,12},\sigma_{R,22}\). Under the bivariate normal model, the conditional distribution at visit 2 given a visit-1 value \(y_1\) is:

Y_2 \mid Y_1=y_1 \sim N\left( \mu_{R,2}+\frac{\sigma_{R,12}}{\sigma_{R,11}}(y_1-\mu_{R,1}), \ \sigma_{R,22}-\frac{\sigma_{R,12}^2}{\sigma_{R,11}} \right).

Here the reference model supplies the conditional relationship, while the participant’s observed value supplies the conditioning history. The observed active-arm value is retained; it is not replaced by a reference-arm value. Copy-reference (CR) uses a different reference-based distributional restriction, as established in the earlier material on Copy-Reference distribution restrictions and covariances. The present distinction is about which part of the trajectory is observed and which part is governed by the selected post-dropout rule.

Key interpretation: “Active treatment before dropout, reference behavior afterward” describes a hybrid trajectory construction for a dropout pattern. It does not assert that a participant’s treatment assignment or observed outcomes changed at dropout, and it does not establish that the reference-like post-dropout behavior is a fact about the participant.

Deriving the Conditional Prediction

The conditional-normal formula follows by separating the reference-arm outcome vector into observed-history coordinates \(Y_O\) and missing coordinates \(Y_M\). Suppose their joint distribution under the reference model is multivariate normal, with means \(\mu_{R,O}\) and \(\mu_{R,M}\), and covariance blocks \(\Sigma_{R,OO}\), \(\Sigma_{R,MO}\), and \(\Sigma_{R,MM}\). If \(\Sigma_{R,OO}\) is invertible, the conditional distribution is:

Y_M\mid Y_O=o \sim N\left( \mu_{R,M}+\Sigma_{R,MO}\Sigma_{R,OO}^{-1}(o-\mu_{R,O}), \ \Sigma_{R,MM}-\Sigma_{R,MO}\Sigma_{R,OO}^{-1}\Sigma_{R,OM} \right).

The conditional mean has two parts: the reference-arm mean for the missing visits, plus an adjustment for how the observed history differs from the reference-arm mean history. The adjustment is a linear prediction based on the reference-arm covariances. The conditional covariance describes remaining uncertainty after conditioning on that history; it is not zero merely because a conditional mean has been calculated.

For J2R, this formula applies when the chosen model uses the reference-arm joint distribution to predict post-dropout outcomes conditional on the observed history. Its use requires a suitable normal model (or an approximation adequate for the analysis), an estimable reference-arm mean and covariance structure, and an invertible observed-history covariance matrix. The reference-based restriction itself is an identifying assumption for the missing outcomes, not a fact verified by observing the dropout participants’ unobserved values.

Worked Example: Predicting a Post-Dropout Outcome Conditionally

Suppose a hypothetical trial measures a symptom score in points at visits 1 and 2. For the reference arm, assume the bivariate normal model has means \(\mu_{R,1}=6\) and \(\mu_{R,2}=6.5\), variances \(\sigma_{R,11}=4\) and \(\sigma_{R,22}=9\), and covariance \(\sigma_{R,12}=2\). An active-arm participant has an observed score of 7 points at visit 1 and drops out before visit 2. Under the J2R conditional prediction, calculate the visit-2 mean and variance.

Assumption check. For this calculation, assume the reference-arm bivariate normal model is an adequate predictive model, its covariance estimates are applicable to this participant’s history, and the observed visit-1 score is within a range where the reference conditional relationship is credible. The variances are positive and \(\sigma_{R,11}=4\) is nonzero, so the conditional formula is defined.

The conditional mean is:

6.5+\frac{2}{4}(7-6)=6.5+0.5=7.0\text{ points}.

The conditional variance is:

9-\frac{2^2}{4}=9-1=8\text{ points}^2.

Thus, the reference-based predictive distribution is \(Y_2\mid Y_1=7\sim N(7,8)\), with conditional standard deviation \(\sqrt{8}=2.828\) points. The prediction is not simply the unconditional reference mean of 6.5 points: it is adjusted upward because the observed history is one point above the reference mean at visit 1 and the model assigns a positive covariance to the two visits. This calculation predicts a missing outcome; it does not turn it into an observed value.

Reading Pattern-Specific Trajectories

A pattern-specific trajectory should be read in chronological order. First identify the last observed visit. Values up to and including that visit are the participant’s observed outcomes under the randomized active treatment. Next identify the missing visits. Their predictive distribution is specified by the chosen reference-based rule, conditional on the observed history as required by the model. The location of dropout matters because it determines how much observed active-treatment history is retained.

The following hypothetical summary illustrates the distinction. The post-dropout entries are conditional predictive means for the indicated active-arm dropout pattern, not observed outcomes and not necessarily the unconditional reference-arm means.

Visit	Reference-arm mean (points)	Active pattern: last observed at visit 1	Status for that pattern
Baseline	10	10	Observed
1	12	15	Observed
2	14	15.5	J2R conditional prediction
3	15	16.0	J2R conditional prediction

The visit-1 active-minus-reference contrast in this example is \(15-12=3\) points, based on observed means. At visit 2, the displayed difference \(15.5-14=1.5\) points compares a conditional imputation mean for one dropout pattern with the reference-arm mean. It should not be described as a directly observed treatment contrast. Also, the post-dropout prediction need not equal the unconditional reference mean: the observed active history influences its conditional value.

Worked Example: Retaining the Observed Segment in a Completed Mean

Consider a hypothetical active-arm dropout pattern with two participants and three visits. Both have observed scores of 10 points at baseline. At visit 1, their observed scores are 14 and 16 points. Both then drop out. A fitted reference-based model gives conditional visit-2 predictive means of 13 and 15 points, respectively. Compute the pattern’s average trajectory using these conditional means, and identify which entries are observed.

Assumption check. Treat the stated visit-2 predictions as outputs of a specified, adequate reference-based model conditional on each participant’s observed history. The visit-0 and visit-1 scores are given as observed; only visit 2 is missing. The arithmetic describes the mean under these predictions, not the mean of observed visit-2 outcomes.

At baseline, the pattern mean is \((10+10)/2=10\) points. At visit 1, it is \((14+16)/2=15\) points. These are observed-data averages. At visit 2, the average conditional prediction is:

\frac{13+15}{2}=14\text{ points}.

The resulting illustrative trajectory is \((10,15,14)\) points. Its first two entries are based on observed outcomes; its final entry is based on conditional predictions. If one averages all three visit-specific means, the result is \((10+15+14)/3=13\) points, but that summary combines observed and model-based information. It must be reported as an imputation-based summary, with the reference assumption stated, rather than as though all three visits had been observed.

When the Distinction Matters for Treatment Benefit

An active-arm group mean at a visit is generally an average over the participants whose outcomes are available at that visit. If dropout is related to outcome history, that available-case mean need not represent the mean that would have been seen in the full randomized group. Reference-based imputation addresses missing outcomes through explicit assumptions; it does not remove the need to distinguish observed summaries from completed-data estimates.

A completed-data contrast can combine observed active-arm outcomes before dropout with model-based post-dropout values. Consequently, the estimated treatment contrast at later visits depends on the selected reference-based rule and its fitted model. It is not merely a continuation of the observed pre-dropout contrast, nor does a reference-like prediction prove that treatment benefit biologically vanishes after a participant stops attending.

Worked Example: Separating an Observed Contrast From a Predicted One

In a hypothetical trial, the observed active-arm mean at visit 1 is 15 points and the observed reference-arm mean is 12 points. For active-arm participants in a specified dropout pattern, the conditional visit-2 prediction is 15.5 points; the observed reference-arm visit-2 mean is 14 points. Calculate both contrasts and state their evidential status.

Assumption check. The visit-1 means are assumed to be computed from observed outcomes in their respective arms. The visit-2 active value is explicitly a conditional reference-based prediction for a dropout pattern, while 14 points is the observed reference-arm mean. These quantities can be subtracted arithmetically, but they do not have the same status as observed data.

The visit-1 observed contrast is \(15-12=3\) points, active minus reference. The visit-2 contrast using the specified prediction is \(15.5-14=1.5\) points. The first is an observed-data contrast for the available participants summarized; the second is a model-dependent contrast that includes imputed active-arm outcomes. Neither arithmetic result alone establishes a causal effect for all randomized participants. Interpretation of a treatment effect requires the estimand and analysis assumptions to be specified.

Diagnostics, Alternatives, and Limitations

The post-dropout segment is especially sensitive to model fit and support. Review the reference-arm visit means and covariance estimates, the number of reference participants contributing to the relevant visits, and whether the observed active-arm histories being conditioned on are represented in the reference data. Implausibly extreme conditional predictions, unstable covariance estimates, or histories far outside the reference arm’s range are warnings that the prediction relies heavily on extrapolation. A good fit to observed reference data does not verify the missing-outcome assumption for active-arm dropouts.

The main comparison is with alternatives that use a different assumption for missing outcomes. A missing-at-random analysis uses a different basis for predicting missing values; it does not impose the same post-dropout reference behavior. CR is another reference-based rule, but it does not impose the same distributional restriction as J2R. These choices can yield different completed-data estimates even when they use the same observed data. The earlier material on missing data provides the standard missing-data terminology; here, the central point is to report which assumption governs the unobserved segment.

Reference-based imputation has important limits. It cannot recover the unobserved outcomes from the data alone, and its predictions depend on the reference model, its covariance structure, and the plausibility of the chosen reference behavior for the dropout pattern. The imputation rule is a sensitivity assumption, not a diagnostic that can confirm what would have happened. In addition, a conditional mean is not a full uncertainty calculation: inference must account for uncertainty in missing outcomes and model parameters using an appropriate analysis procedure.

Pitfall: treating the whole trajectory as reference-arm data. This wrongly discards observed active-arm outcomes before dropout. Retain those values; apply the reference-based rule only to the missing segment.
Pitfall: equating a conditional prediction with the reference mean. This ignores the observed history. Under a conditional reference model, the prediction includes the covariance-based adjustment for that history.
Pitfall: describing an imputed trajectory as observed treatment behavior. The correction is to label predictions as model-based and state the post-dropout assumption that produced them.
Pitfall: interpreting the rule as a testable fact. Missing outcomes cannot confirm the imposed post-dropout behavior. Treat the rule as an assumption to be justified and varied in sensitivity analyses.

Key takeaway: In this setting, the active-treatment history through the last observed visit remains intact. Reference-based assumptions govern only the distribution of subsequent missing outcomes; any later trajectory or contrast that uses them is conditional on that assumption.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

An active-arm participant has observed scores through visit 2 and misses visits 3 and 4. Under a reference-based rule, which outcomes are observed and which are governed by the rule?
In a bivariate reference model, \(\mu_{R,1}=8\), \(\mu_{R,2}=9\), \(\sigma_{R,11}=4\), and \(\sigma_{R,12}=1\). Find the conditional mean at visit 2 when the observed visit-1 score is 10.
Why is a post-dropout conditional prediction not generally equal to the unconditional reference-arm mean?
A visit-2 active-minus-reference contrast is 2 points, but the active-arm visit-2 value includes imputation. What must be stated when interpreting the contrast?

Solutions. (1) Visits 0 through 2 remain observed active-arm outcomes. The chosen reference-based assumption governs predictions for visits 3 and 4, conditional on the available history as specified by the model. (2) The conditional mean is \(9+(1/4)(10-8)=9+0.5=9.5\) points. (3) The conditional mean adjusts the reference mean using the observed history and the reference model’s covariance relationship; it equals the unconditional mean only in special cases, such as a zero covariance with the conditioning history. (4) State that the active value is imputed, identify the reference-based assumption and model, and describe the contrast as model-dependent rather than wholly observed.