Worked Single-Visit J2R Imputation

What a Single-Visit J2R Imputation Calculates

Suppose a participant was assigned to the treatment arm, has an observed outcome history up to a particular point, and then has one missing outcome. A single-visit jump-to-reference (J2R) imputation predicts that missing outcome using a reference-arm mean restriction for the missing visit, while allowing the participant’s observed history to inform the prediction through the covariance model. The result is therefore not generally just the reference-arm average.

The target here is the conditional mean of one missing outcome under J2R. It is a model-based prediction for the participant given their observed history. It is not an observed outcome, and it is not by itself a complete procedure for estimating a treatment effect or its uncertainty.

Definition: Let $Y_O$ be a participant’s vector of observed outcomes and $Y_m$ the one missing outcome to be imputed. Under the single-visit J2R mean restriction, the observed outcomes retain their treatment-arm mean vector $\mu_{T,O}$, while the missing outcome has the reference-arm mean $\mu_{R,m}$. With covariance blocks $\Sigma_{OO}$ and $\Sigma_{mO}$, the J2R conditional mean is $$ E(Y_m\mid Y_O=y_O) = \mu_{R,m} + \Sigma_{mO}\Sigma_{OO}^{-1}(y_O-\mu_{T,O}). $$

Here $T$ denotes the treatment arm and $R$ the reference arm. The vector $y_O-\mu_{T,O}$ is the participant’s observed history expressed as deviations from the treatment-arm means for those observed outcomes. The covariance multiplier translates those deviations into their expected association with the missing outcome. In the common-covariance version used in the examples, the same covariance structure is used for both arms; the mean restriction distinguishes the treatment and reference components.

The formula requires a specified mean model and covariance model, with $\Sigma_{OO}$ invertible. It is commonly used under a multivariate normal model for the outcomes, conditional on any covariates included in the model. The J2R restriction is a sensitivity assumption about outcomes after treatment discontinuation or dropout: their mean is set to the reference-arm mean. It is not established merely by observing the participant’s data.

Why the History Adjusts the Reference Mean

A missing outcome can be associated with earlier outcomes. If a participant’s observed outcomes are unusually high or low relative to the treatment-arm means, that information can shift their conditional prediction. J2R sets the missing visit’s starting mean to the reference-arm mean, but does not necessarily discard the association between observed history and the missing outcome.

The adjustment follows from the conditional mean of a partitioned multivariate normal vector. Write $X=Y_O$, $Z=Y_m$, and let their means and covariance blocks be those in the definition. Define $B=\Sigma_{mO}\Sigma_{OO}^{-1}$ and $\varepsilon=Z-\mu_m-B(X-\mu_O)$. Then $\operatorname{Cov}(\varepsilon,X)=\Sigma_{mO}-B\Sigma_{OO}=0$. Because $(\varepsilon,X)$ is jointly normal, zero covariance implies independence. Also $E(\varepsilon)=0$, so conditioning on $X=x$ gives $E(Z\mid X=x)=\mu_m+B(x-\mu_O)$. For J2R, the observed-history mean $\mu_O$ is $\mu_{T,O}$ and the missing-outcome mean $\mu_m$ is $\mu_{R,m}$, yielding the stated calculation.

Worked Example: One Observed Visit and One Missing Visit

Hypothetical data. A clinical outcome is measured in points at two visits. A treatment-arm participant has an observed first-visit score of 9 points and a missing second-visit score. The fitted first-visit treatment mean is 7 points; the reference-arm mean at the missing visit is 5 points. Assume the common covariance model gives first-visit variance 4 points squared and covariance between visits 3 points squared. The observed covariance block is invertible because its variance is positive.

With one observed value, $\Sigma_{mO}\Sigma_{OO}^{-1}$ is the covariance divided by the observed variance:

\widehat{Y}_{m,\mathrm{J2R}} = 5+\frac{3}{4}(9-7) = 5+0.75(2) = 6.50\text{ points}.

The participant’s observed score is 2 points above the treatment mean at the first visit. The modelled covariance carries some of that positive deviation to the missing visit, adding 1.50 points to the reference mean. The J2R conditional mean is therefore 6.50 points, not 5 points. The arithmetic uses a reference mean at the missing visit and a treatment mean for centring the observed treatment history.

For completeness, if the missing-visit variance is 9 points squared, the conditional variance under this covariance model is $9-3^2/4=6.75$ points squared. A stochastic imputation would need to account for this conditional variation; 6.50 points is the conditional mean, not a claim that the unobserved score must equal 6.50.

Using More Than One Observed Outcome

When a participant has multiple observed outcomes, the adjustment uses the full observed-history covariance matrix. It is not generally correct to calculate a separate adjustment from each visit and add those adjustments without accounting for correlations among observed outcomes. The matrix multiplication handles their joint information.

Worked Example: Two Observed Visits

Hypothetical data. A participant has two observed treatment-arm scores, $y_O=(7,6)^\mathsf{T}$, in points, and one later score is missing. The treatment-arm means for the observed visits are $\mu_{T,O}=(5,7)^\mathsf{T}$, and the reference-arm mean for the missing visit is $\mu_{R,m}=4$ points. Suppose $\Sigma_{OO}=\begin{pmatrix}4&1\\1&4\end{pmatrix}$ and $\Sigma_{mO}=(1,2)$, with all covariance entries in points squared. The determinant of $\Sigma_{OO}$ is $4(4)-1(1)=15>0$, so this observed-history covariance block is invertible.

First calculate the observed deviations: $y_O-\mu_{T,O}=(7-5,6-7)^\mathsf{T}=(2,-1)^\mathsf{T}$. The inverse observed covariance matrix is $\Sigma_{OO}^{-1}=\frac{1}{15}\begin{pmatrix}4&-1\\-1&4\end{pmatrix}$. Applying it to the deviations gives

\Sigma_{OO}^{-1}(y_O-\mu_{T,O}) = \frac{1}{15} \begin{pmatrix} 4&-1\\ -1&4 \end{pmatrix} \begin{pmatrix} 2\\ -1 \end{pmatrix} = \begin{pmatrix} 0.6000\\ -0.4000 \end{pmatrix}.

Now multiply by the covariance between the missing outcome and the observed history, then add the reference mean:

\widehat{Y}_{m,\mathrm{J2R}} = 4+ \begin{pmatrix}1&2\end{pmatrix} \begin{pmatrix}0.6000\\-0.4000\end{pmatrix} = 4+(0.6000-0.8000) = 3.80\text{ points}.

The first observed score is above its treatment mean and the second is below its treatment mean. Their joint, covariance-adjusted contribution is $-0.20$ points, so the conditional mean is slightly below the missing visit’s reference mean. The direction and size of the adjustment depend on both the deviations and their covariance with the missing outcome.

Checks Before Reporting the Imputed Value

A numerical result is meaningful only under the model used to produce it. Before interpreting a J2R conditional mean, check the following:

Visit and arm alignment: $\mu_{T,O}$ must correspond, in order, to the participant’s observed visits; $\mu_{R,m}$ must be the reference mean for the visit being imputed.
Covariance alignment: the rows and columns of $\Sigma_{OO}$ and $\Sigma_{mO}$ must refer to those same visits, in compatible order and units. The observed covariance block must be invertible.
Mean-model conditioning: if the model conditions on baseline covariates or other design variables, the arm means used must be appropriate for the participant’s covariate values or strata.
Model plausibility: assess whether the fitted means and covariances reasonably describe the observed data. Implausible covariance estimates or unstable inverses can produce unreliable adjustments.

The result is a conditional mean, so it can fall outside a bounded outcome’s permitted range even when every input mean is within that range. Such a prediction is a warning to inspect the model and outcome scale; it is not, by itself, proof of an arithmetic error. Conversely, a plausible-looking value does not validate the J2R assumption. The post-dropout reference-mean restriction concerns unobserved outcomes and cannot be directly checked against those outcomes.

Worked Example: A Below-Reference Prediction

Hypothetical data. A participant has an observed treatment-arm score of 16 points at one visit, compared with a treatment-arm mean of 20 points. The reference mean at the missing visit is 12 points. Suppose the observed-visit variance is 25 points squared and the covariance between the observed and missing visits is 10 points squared. The observed variance is positive, so the scalar covariance block is invertible.

The conditional-mean adjustment is the covariance-to-variance ratio, $10/25=0.40$, multiplied by the observed deviation, $16-20=-4$ points:

\widehat{Y}_{m,\mathrm{J2R}} = 12+\frac{10}{25}(16-20) = 12+0.40(-4) = 10.40\text{ points}.

The prediction is 1.60 points below the reference mean because this participant’s observed score is below the treatment-arm mean and the fitted covariance is positive. This does not mean J2R has been abandoned: the missing-visit mean restriction is still the reference mean, while the conditional prediction also uses observed history.

How This Calculation Differs From Alternatives

The defining distinction in this calculation is the mean assigned to the missing visit. A treatment-based prediction would instead use a treatment-arm mean for that visit, along with its corresponding conditioning model. J2R uses the reference-arm mean for the missing visit while retaining the treatment-arm mean for centring the participant’s observed treatment history. The copy-reference approach is a separate reference-based restriction; it is not the calculation being performed here.

A simple substitution of the reference-arm mean, with no history adjustment, is generally not the same as the J2R conditional mean. It agrees with the formula only when the covariance adjustment is zero, for example when the observed history has no modelled association with the missing outcome, or when the relevant observed deviations combine to zero after covariance adjustment. Likewise, using a treatment-arm mean for the missing visit would change the J2R restriction itself.

J2R is a sensitivity assumption, not a fact that follows from the observed data. Results can depend on the selected reference arm, fitted means, covariance structure, and assumptions about post-dropout outcomes. The calculation for one participant does not by itself account for uncertainty in estimated model parameters or supply valid population-level inference.

Common Mistakes

Using only the reference mean. This omits the conditional adjustment when observed history is associated with the missing outcome. Use the covariance-weighted deviations unless the model implies that adjustment is zero.
Centring observed treatment outcomes at reference means. In the J2R setup stated here, the observed components retain treatment-arm means. The missing component uses its reference-arm mean.
Ignoring correlations between observed visits. With multiple observed outcomes, use the full $\Sigma_{OO}^{-1}$; treating visits as unrelated can change the result.
Calling the conditional mean the realised missing score. It is a model-based average over possible missing values given observed history. A stochastic imputation also reflects conditional variability.
Treating J2R as empirically verified. The reference-mean restriction concerns outcomes after dropout that are unobserved. It should be interpreted as an assumption for sensitivity analysis.

Key takeaway: For one missing outcome under J2R, start with the reference-arm mean at that visit, then add the covariance-weighted deviation of the participant’s observed treatment history from its treatment-arm means. Check the visit alignment, covariance model, and assumptions before interpreting the result.

Check Your Understanding

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

A participant’s observed treatment outcome is 8, its treatment-arm mean is 6, the missing visit’s reference mean is 10, and the covariance-to-variance ratio is 0.25. What is the J2R conditional mean?
Why does a J2R calculation use a treatment-arm mean to centre an observed treatment outcome but a reference-arm mean for the missing outcome?
With two observed visits, why is it generally insufficient to adjust separately for each visit without using their covariance?
If the covariance between the observed outcome and the missing outcome is zero, what does the scalar J2R formula predict?
What does it mean if the J2R conditional mean differs from the reference-arm marginal mean?

Solutions. (1) The observed deviation is $8-6=2$, so the conditional mean is $10+0.25(2)=10.50$. (2) The J2R restriction assigns the reference-arm mean to the post-dropout missing outcome, while the observed history remains modelled as treatment-arm data; centring at the treatment mean measures that history’s deviation from its assigned mean. (3) The observed outcomes may be correlated with one another, so their joint information is represented by $\Sigma_{OO}^{-1}$; separate adjustments can count or weight information incorrectly. (4) The covariance ratio is zero, so the adjustment is zero and the conditional mean equals the missing visit’s reference mean. (5) The difference is the history-based conditional adjustment: observed deviations inform the prediction through their covariance with the missing outcome. It does not change the reference-mean restriction or establish what the unobserved score actually was.