J2R Under Time-Varying Control Response

Why a Changing Reference Trajectory Matters

A reference arm does not necessarily have a flat mean outcome over follow-up. Control participants may improve, worsen, or change direction between visits. A J2R prediction must use the reference-arm mean for the visit being predicted, together with the reference trajectory’s covariance structure. Reusing a single control mean at every visit can distort the predicted path even when the covariance calculation is otherwise correct.

This tutorial focuses on calculating those predictions when the reference-arm mean changes over time. The calculations concern hypothetical outcomes for participants assigned to treatment who discontinue treatment or otherwise have missing follow-up outcomes. They are model-based predictions under an assumption, not observed outcomes or guaranteed descriptions of what happened.

In the examples, the outcome is measured in points, \(R\) denotes the reference arm, and \(A\) denotes the treatment arm. A dropout time \(k\) means outcomes through visit \(k\) are observed and later outcomes are missing. For clarity, numerical examples treat the estimated means and covariance parameters as fixed. In multiple imputation, parameter uncertainty must also be represented when drawing imputations.

Conditional Prediction with Visit-Specific Means

Suppose the reference-arm outcome vector across visits follows a multivariate normal model with mean vector \(\boldsymbol{\mu}_R\) and covariance matrix \(\boldsymbol{\Sigma}_R\). Partition the visits into the observed history \(P\) and the future visits \(F\). The reference model’s conditional mean for the future outcomes, given the observed history \(\mathbf y_P\), is

The first term, \(\boldsymbol{\mu}_{R,F}\), contains the reference mean at each future visit. The second term adjusts that mean for how the observed history differs from the reference-arm history, with the adjustment determined by the reference covariance. Thus, a J2R conditional prediction is not generally just the unconditional reference mean at the future visit. It uses the time-varying reference means and the conditional relationship between past and future outcomes.

For a useful special case, suppose the reference outcomes have equal variance and an AR(1) correlation structure: the correlation between adjacent visits is \(\rho\), and the correlation between visits \(i\) and \(j\) is \(\rho^{|i-j|}\). This model is Markov, so the next visit’s conditional mean given the full history depends only on the most recent outcome:

For sequential predictions, replace \(y_{j-1}\) with the observed outcome if visit \(j-1\) is observed, or with the preceding conditional prediction when calculating the next expected value. The reference mean increment is therefore present through the two visit-specific means; the covariance term adjusts for the preceding outcome’s deviation from its own visit’s reference mean.

Worked Calculations for Rising and Falling Means

The AR(1) recursion is convenient, but it is not a substitute for the general conditional formula when the covariance structure is not AR(1), when the observed history is irregular, or when the full pattern of correlations matters. The next example calculates a conditional prediction directly from a partitioned covariance matrix.

Justification of the Conditional Formula

The conditional-mean expression follows from the joint normal model. Define the future residual after its linear adjustment for the observed history as

This residual has mean zero. Its covariance with the observed history is

Because \(\mathbf U\) and \(\mathbf Y_P\) are jointly normal, zero covariance implies independence. Hence the conditional expectation of \(\mathbf U\) given \(\mathbf Y_P=\mathbf y_P\) is still zero. Rearranging the definition of \(\mathbf U\) gives the stated conditional mean. This argument also shows why the covariance matrix must match the modeled visits and why the observed-history block must be invertible.

Checks, Alternatives, and Limitations

Several practical checks help catch errors before interpreting a J2R trajectory:

• Match the visit index. For a prediction at visit \(j\), use \(\mu_{R,j}\), not a baseline or dropout-visit mean by default.
• Check the residual reference visit. In the AR(1) recursion, subtract \(\mu_{R,j-1}\) from the preceding outcome or prediction, not \(\mu_{R,j}\).
• Confirm the covariance scale and ordering. The rows and columns must correspond to the same visits as the mean vector, and the covariance matrix must be valid for the modeled outcomes.
• Inspect the resulting path. Plot predictions alongside the visit-specific reference means and observed outcomes. Unexpected jumps, reversals, or extreme values can indicate a coding, indexing, or model-fit problem; they are not automatically proof that the assumption is wrong.
• Check support for estimation. Sparse reference-arm data or an inadequate covariance model can make conditional predictions unstable. A mathematically computable prediction is not necessarily a reliable one.

J2R is one assumption for missing outcomes, not a consequence of the observed data alone. The contrast with common alternatives is useful:

A common mistake is to set every missing outcome equal to the control mean at the dropout visit, or to repeat that one mean at all later visits. That ignores a changing reference trajectory and, in the conditional model, the covariance adjustment. The correction is to use each future visit’s reference mean and apply the appropriate conditional adjustment. Conversely, using a future reference mean without checking its visit index, outcome scale, and covariance model does not make a calculation valid.

The predictions also depend on the chosen reference arm, mean model, covariance structure, and J2R assumption. They do not establish that a participant’s unobserved outcomes truly followed the reference-arm process. Sensitivity analysis may be needed to assess how conclusions change under alternative assumptions; the J2R calculation itself does not remove that uncertainty.