Jump-to-Reference Mean Restrictions

What “Jump to Reference” Means

In a longitudinal clinical trial, a participant assigned to an active treatment may stop that treatment before all outcome visits are observed. A reference-based restriction describes how the unobserved outcomes after that point are to be represented. Jump-to-reference (J2R) makes a particular mean-level assumption: after discontinuation, the active-arm participant’s expected outcome follows the corresponding reference-arm predictive mean, rather than continuing an active-treatment mean trajectory.

The word “jump” describes the assumed change in the mean behavior after discontinuation. It does not mean that an individual’s outcome is observed to jump, nor does it assert that treatment discontinuation causes a particular biological change. The post-dropout outcomes are unobserved; the restriction specifies a model for their means.

Definition: Let $A=1$ denote assignment to active treatment and $A=0$ assignment to the reference treatment. Let $Y_t$ be the outcome at visit $t$, and let $D=d$ indicate that visit $d$ is the last observed visit for a participant. Write $H_d$ for that participant’s observed outcome history through visit $d$, and $X$ for baseline covariates. Under J2R, for each post-dropout visit $t>d$, the assumed conditional mean for the active-arm participant is the reference-arm predictive mean for the same history and covariates: $$ E(Y_t\mid A=1,D=d,H_d,X;\mathrm{J2R}) = E(Y_t\mid A=0,H_d,X). $$ The right-hand side is evaluated using the specified reference-arm model. This is a restriction on conditional means, not a claim that the participant was assigned to the reference arm.

The conditioning matters. A reference-arm predictive mean given an observed history need not equal the reference arm’s unconditioned mean at that visit. For example, participants with unusually high outcomes at visit $d$ may have different expected later outcomes from the reference-arm average. The reference predictive mean compares like histories under the reference model; it is not automatically the raw average among all reference participants.

This formulation presumes that the reference model defines a meaningful predictive mean for the histories and covariates encountered among active-arm participants who drop out. If a simpler model uses only covariate strata, the same idea applies within each stratum. In a setting where histories are treated as identical or irrelevant, the rule can reduce to setting the post-dropout mean equal to the reference-arm mean at that visit.

Assumptions and Scope of the Restriction

J2R is an identifying assumption about outcomes that are missing after dropout. It is not a consequence of randomization alone. Randomization supports comparisons of assigned groups at baseline, but it does not reveal the unobserved post-dropout outcomes of participants assigned to active treatment.

To state and use the restriction coherently, the analyst must specify the outcome scale, visit times, reference arm, dropout definition, covariates and history included in the reference prediction, and the reference model. The following conditions are also important:

Comparable outcomes: the outcome must have the same interpretation and measurement scale across treatment arms and visits.
A defined reference prediction: the reference-arm data and model must support prediction at the relevant visits for the active participants’ covariates and observed histories. Severe lack of overlap makes the required prediction extrapolative.
A meaningful history: the conditioning history must be defined consistently. If it includes outcomes through dropout, those outcomes must have been observed for the pattern being considered.
A substantive post-dropout assumption: after discontinuation, the expected trajectory is taken to behave like the reference-arm prediction, even though the participant was randomized to active treatment and may have received it previously.

These conditions do not make J2R empirically true. Observed outcomes can help assess whether the chosen reference model is compatible with observed data and whether it predicts plausibly in regions with support. They cannot directly check the assumed means of outcomes that remain unobserved after dropout.

Worked Example: Replacing an Active Mean at a Later Visit

Suppose a hypothetical two-arm trial records a score at visits 1, 2, and 3. Higher scores are better. For a simple illustration, assume that the relevant observed histories are identical for the participants being compared, so that the reference predictive means at visits 1–3 are 10, 12, and 14 points. The active-arm mean trajectory while on treatment is 10, 15, and 18 points. A participant assigned to active treatment has visit 2 as the last observed visit.

Under J2R, the participant’s assumed mean at the only post-dropout visit, visit 3, is the visit-3 reference predictive mean:

E(Y_3\mid A=1,D=2;\mathrm{J2R})=14\text{ points}.

For comparison, a straight-line continuation of the two observed active-arm means would add the visit-1-to-visit-2 increase of $15-10=5$ points, giving $15+5=20$ points at visit 3. That is not the J2R restriction. J2R uses 14 points here because the stipulated post-dropout mean follows the reference trajectory.

At visit 2, which is observed for this participant, J2R does not replace the observed score with a reference value. The restriction concerns post-dropout visits only. The example is a mean-level illustration; it does not specify the variance or joint distribution of visit-3 outcomes.

How the Mean Contrast Changes

The defining consequence of J2R can be expressed directly. Define the conditional mean contrast at a post-dropout visit as the active-arm J2R mean minus the reference predictive mean, evaluated at the same history and covariates. The restriction sets that contrast to zero at every such history:

\Delta_t(h_d,x) = E(Y_t\mid A=1,D=d,H_d=h_d,X=x;\mathrm{J2R}) - E(Y_t\mid A=0,H_d=h_d,X=x) =0,\qquad t>d.

Averaging these conditional means over any specified distribution of histories and covariates also gives a zero difference between the two quantities averaged over that same distribution. This follows by taking expectations of the conditional equality; it does not imply that every marginal comparison in the trial is zero. In particular, averages over different covariate or history distributions can differ.

This distinction is useful in reporting results. “The post-dropout conditional mean contrast is zero under J2R” is precise. “The active and reference arms have the same outcome mean after dropout” may be misleading if it sounds like a claim about the observed arm averages, which can have different participant compositions and include different treatment histories.

Worked Example: Using a Reference Predictive Mean Given History

Suppose, hypothetically, that a fitted reference-arm model gives the following visit-3 predictive mean for a participant with visit-2 score $y$:

E(Y_3\mid A=0,Y_2=y)=4+0.5y.

An active-arm participant drops out after visit 2 with an observed score of 16 points. Under J2R, the assumed mean at visit 3 is the reference prediction evaluated at that history:

4+0.5(16)=4+8=12\text{ points}.

Thus $E(Y_3\mid A=1,D=2,Y_2=16;\mathrm{J2R})=12$ points. If the reference arm’s unconditioned visit-3 mean were 11 points, substituting 11 would not follow the stated conditional rule. The correct target in this example is 12 because it conditions on the participant’s observed visit-2 score.

The calculation assumes the reference prediction is appropriate for a history of 16 points and that the model’s linear expression applies there. If such histories are rare or absent in the reference arm, the numerical prediction may rely on extrapolation; the arithmetic does not remove that limitation.

Covariates, Patterns, and Averaging

J2R is applied to a dropout pattern and its post-dropout visits, not just to an overall trial mean. A participant who drops out after visit 1 has a different observed history from one who drops out after visit 3. Their reference predictive means may therefore differ, even for the same later visit.

Covariates create a related issue. If reference means differ across baseline groups, the post-dropout mean for each participant should use the prediction for that participant’s covariates when the restriction is defined conditionally on those covariates. The resulting average for a dropout pattern depends on the covariate distribution within that pattern. It need not equal an overall reference-arm average.

Worked Example: Standardizing Across Covariate Groups

Consider a hypothetical trial with two baseline strata, low-risk and high-risk. At visit 3, the reference predictive means for the relevant histories are 10 points in the low-risk stratum and 14 points in the high-risk stratum. Suppose the trial’s target population is 75% low-risk and 25% high-risk.

The visit-3 mean obtained by standardizing the J2R predictions to that target population is:

0.75(10)+0.25(14)=7.5+3.5=11\text{ points}.

Now suppose the active-arm participants who dropped out are 25% low-risk and 75% high-risk. Their average J2R prediction, averaged over that dropout group’s own covariate mix, is:

0.25(10)+0.75(14)=2.5+10.5=13\text{ points}.

Both calculations use the same stratum-specific J2R restrictions, but they answer different averaging questions: 11 points is standardized to the trial population, whereas 13 points averages over the dropout group’s covariate distribution. Treating either number as the universal reference mean would hide the population being averaged over.

Justification and Interpretation

The zero-contrast implication follows from the defining equality, with no additional distributional claim. For fixed $d$, $t>d$, history $h_d$, and covariates $x$, J2R sets the active-arm conditional mean equal to the reference conditional mean. Subtracting the latter from both sides gives a conditional difference of zero. If the target average uses weights $w(h_d,x)$ that sum or integrate to one and are applied to both conditional means, then:

\sum_{h_d,x} w(h_d,x) \left[ E(Y_t\mid A=1,D=d,h_d,x;\mathrm{J2R}) - E(Y_t\mid A=0,h_d,x) \right] = \sum_{h_d,x} w(h_d,x)(0)=0.

For continuous histories or covariates, the sum is replaced by an integral. The key qualification is that the same weights are used on both sides. If the two means are averaged over different distributions of histories or covariates, the result need not be zero. This is why a conditional restriction and an unadjusted arm-level comparison are not interchangeable.

J2R is specifically a mean restriction. It does not, by itself, specify the variance of post-dropout outcomes, their correlations across visits, or their full predictive distribution. Many different distributions can share the same mean. A complete imputation model needs additional choices beyond the J2R mean restriction; those choices should not be silently attributed to J2R.

Diagnostics, Alternatives, and Limitations

Because the restricted outcomes are missing, there is no direct diagnostic that proves the post-dropout J2R mean is correct. Useful checks instead examine the ingredients and implications that can be assessed:

Check whether the reference model fits observed reference-arm outcomes adequately across visits, histories, and important covariate groups.
Check overlap: compare the active dropout participants’ observed histories and covariates with those represented in the reference arm. Predictions far outside that support deserve particular caution.
Check whether the chosen conditioning set is clinically and statistically defensible. Omitting an important covariate or history feature can make an apparently suitable reference prediction poorly matched to the participant.
Check that the stated visit-specific predictions behave plausibly on the outcome scale and are consistent with the intended dropout timing.

J2R is one sensitivity assumption among alternatives. Under an active-arm continuation assumption, post-dropout means follow an extrapolation of the active-arm behavior rather than switch to reference predictions. A sensitivity shift can instead alter a chosen predicted mean by a specified amount. These alternatives encode different assumptions about outcomes after dropout; observed data alone generally cannot choose among them.

The principal limitation is substantive: J2R assumes reference-like mean behavior after active-treatment discontinuation, although prior active treatment may have persistent effects and the reasons for discontinuation may be related to prognosis. The restriction also depends on how the reference prediction is conditioned and averaged. It is best reported with the reference arm, visit range, dropout definition, conditioning variables, and target population made explicit.

Common pitfall: “J2R means the missing active-arm score equals the reference-arm average at that visit.” This can be wrong because the rule is defined using the specified reference predictive mean for the same history and covariates. Use the raw arm average only when the model and conditioning set make it the appropriate reference prediction.

A second common error is to infer that J2R makes the entire post-dropout outcome distributions identical. It equates the means specified by the restriction; equality of variances, correlations, or full distributions requires additional assumptions.

Check Your Understanding

Answer each question using the stated conditioning and target population.

An active-arm participant’s last observed visit is visit 2. Does J2R replace the observed visit-2 outcome, or restrict the assumed mean at later visits?
A reference model gives $E(Y_4\mid A=0,H_2=h,X=x)=17$ points. What post-dropout mean does J2R assign at visit 4 to an active participant with that history and covariates?
Why might the J2R mean for an active dropout group differ from the unconditioned reference-arm average?
Does a J2R mean restriction imply equal post-dropout variances and correlations across arms?

Solutions. (1) It restricts the assumed means at visits after visit 2; it does not replace observed outcomes. (2) It assigns a visit-4 conditional mean of 17 points. (3) The dropout group may have a different mix of observed histories or covariates, and the rule uses reference predictions conditional on those features rather than necessarily using the overall reference average. (4) No. J2R specifies means only; variance and correlation assumptions must be stated separately.