Sequential J2R Imputation Across Follow-Up Visits

Why Generate Post-Dropout Visits Sequentially?

When a participant leaves the treatment arm before the end of follow-up, several later outcomes may be missing. Under jump-to-reference (J2R), the post-dropout behavior is governed by a reference-arm distribution, while the participant’s observed treatment-arm history remains part of the information used to predict missing outcomes. If more than one visit is missing, each imputed value becomes part of the history for predicting the next one. This is the key feature of sequential J2R imputation.

Sequential imputation is not a rule to copy the reference-arm mean at every visit. It generates a conditional draw at each visit, using the reference model and the participant’s observed and previously generated outcomes. The resulting imputed trajectory can vary from one imputation to another, and values at later visits remain statistically related to earlier ones.

Definition: Suppose a participant has observed outcomes through visit \(k\) and missing outcomes at visits \(k+1,\ldots,T\). Under sequential J2R, at each missing visit \(t\), draw \(Y_t\) from the reference-arm conditional distribution given (1) the participant’s observed outcomes and (2) the values already generated at visits \(k+1,\ldots,t-1\). Use the reference-arm mean and covariance model for this conditional distribution.

The Conditional Distribution at Each Step

Write \(O\) for the indices of the participant’s observed outcome visits, and let \(H_t\) denote the available history just before imputing visit \(t\): the observed outcomes \(Y_O\) together with the generated outcomes \(Y_{k+1},\ldots,Y_{t-1}\). For example, when outcomes are observed through visit 2, the history for imputing visit 4 is \((Y_1,Y_2,Y_3)\), where \(Y_3\) has already been generated.

Assume the reference-arm outcomes have a multivariate normal distribution with mean vector \(\mu_R\) and covariance matrix \(\Sigma_R\). Assume the relevant covariance matrices are nonsingular. Conditional on the history, the reference-arm distribution for \(Y_t\) is normal. Its conditional mean and variance are:

m_t(H_t)=\mu_{R,t}+\Sigma_{R,t,H_t}\Sigma_{R,H_t,H_t}^{-1}(H_t-\mu_{R,H_t}), \qquad v_t(H_t)=\Sigma_{R,t,t}-\Sigma_{R,t,H_t}\Sigma_{R,H_t,H_t}^{-1}\Sigma_{R,H_t,t}.

Here \(\mu_{R,t}\) is the reference-arm mean at visit \(t\); \(\Sigma_{R,t,H_t}\) is the row of reference-arm covariances between \(Y_t\) and the history; and \(\Sigma_{R,H_t,H_t}\) is the history’s reference-arm covariance matrix. The conditional standard deviation is \(\sqrt{v_t(H_t)}\). A draw can be written as:

Y_t^{\mathrm{imp}}=m_t(H_t)+\sqrt{v_t(H_t)}Z_t,\qquad Z_t\sim N(0,1).

After this draw, include \(Y_t^{\mathrm{imp}}\) in the history for the next missing visit. In a full multiple-imputation procedure the unknown model parameters may also be drawn; here, the formulas and examples treat the reference-arm parameters as fixed and focus on sequential outcome generation.

Why the Sequential Formula Gives the Joint Imputation

The rationale follows from two standard properties of the multivariate normal distribution. First, any component of a multivariate normal vector conditional on another subset of its components has a normal conditional distribution, with the mean and variance shown above. Second, the chain rule for conditional probabilities factors the joint distribution of all missing outcomes into successive conditional distributions.

f_R(Y_{k+1},\ldots,Y_T\mid Y_O) = \prod_{t=k+1}^{T} f_R(Y_t\mid Y_O,Y_{k+1},\ldots,Y_{t-1}).

To see why this matters, start with the joint reference distribution for observed and missing visits. Its conditional density for the missing vector given the observed history can be factored by the chain rule: first the density of \(Y_{k+1}\) given \(Y_O\), then the density of \(Y_{k+2}\) given \(Y_O,Y_{k+1}\), and so on. Each factor is exactly the conditional distribution used at one sequential step. Therefore drawing from these successive conditionals generates a draw from the joint reference conditional distribution for the entire missing sequence. The simulated innovations \(Z_t\) can be independent standard normal variables; the imputed outcomes are not generally independent because each step conditions on prior values.

Worked Example: Two Missing Visits Under an AR(1) Reference Model

Suppose a participant in the treatment arm has observed outcomes \(Y_1=11\) and \(Y_2=13\), then drops out. For illustration, assume the reference-arm outcomes follow a Gaussian AR(1) model with mean vector \((10,12,14,15)\), marginal variance \(16\) at each visit, and correlation \(\rho=0.5\) between adjacent visits. Under this model, conditional on the immediately preceding outcome and the full earlier history, the next outcome has conditional mean \(\mu_{R,t}+\rho(Y_{t-1}-\mu_{R,t-1})\) and conditional variance \(16(1-\rho^2)=12\). This Markov property justifies using the immediately preceding outcome in the calculations below.

For visit 3, the preceding outcome \(Y_2=13\) is observed. The conditional mean is \(14+0.5(13-12)=14.5\), and the conditional standard deviation is \(\sqrt{12}\approx3.464\). If the chosen standard-normal draw is \(Z_3=-0.5\), then:

Y_3^{\mathrm{imp}}=14.5+\sqrt{12}(-0.5)\approx14.5-1.732=12.768.

For visit 4, update the history with the generated value \(Y_3^{\mathrm{imp}}\). The conditional mean is \(15+0.5(12.768-14)\approx14.384\), with the same conditional standard deviation \(\sqrt{12}\). If \(Z_4=0.25\), then:

Y_4^{\mathrm{imp}}=14.384+\sqrt{12}(0.25)\approx14.384+0.866=15.250.

The imputed sequence is therefore approximately \(12.768,15.250\). The second mean uses the first imputed value, not the participant’s last observed value alone. These are hypothetical values from a specified reference model and random draws, not reported clinical results.

Observed History Can Contribute Beyond the Last Visit

In some covariance structures, earlier observed outcomes continue to predict a missing visit even after conditioning on the most recent outcome. The general matrix formula automatically accounts for all available history. The next example uses conditional regression summaries that are assumed to have been calculated from a reference-arm multivariate normal model.

Worked Example: Using the Full Available History

A participant has observed outcomes \(Y_1=20\) and \(Y_2=21\), with visits 3 and 4 missing. Suppose the reference model gives the following conditional distributions, where the means and variances are conditional on the history shown:

Y_3\mid(Y_1,Y_2)\sim N\big(20+0.3(Y_1-18)+0.6(Y_2-19),25\big),

Y_4\mid(Y_1,Y_2,Y_3)\sim N\big(22+0.1(Y_1-18)+0.2(Y_2-19)+0.5(Y_3-20),20\big).

For visit 3, the conditional mean is \(20+0.3(2)+0.6(2)=21.8\). Its standard deviation is \(\sqrt{25}=5\). With \(Z_3=0.4\), the imputed outcome is \(21.8+5(0.4)=23.8\).

For visit 4, use \(Y_3^{\mathrm{imp}}=23.8\) as well as both observed outcomes. The conditional mean is \(22+0.1(2)+0.2(2)+0.5(3.8)=24.5\), and the standard deviation is \(\sqrt{20}\approx4.472\). With \(Z_4=-0.25\):

Y_4^{\mathrm{imp}}=24.5+\sqrt{20}(-0.25)\approx24.5-1.118=23.382.

The numerical conditional distributions here are stipulated summaries of the reference model; they are not inferred from the two observed outcomes alone. The calculation illustrates the general rule: use the conditional model’s entire specified history, including the newly generated visit.

Sequential Simulation as a Practical Procedure

For each participant with a monotone post-dropout sequence, apply the following steps. The reference model, observed visit values, and parameter values must be specified before generating outcomes.

Set the reference model and observed history.
Identify the reference-arm mean vector and covariance matrix, and list the participant’s observed outcomes through dropout.

Calculate the first conditional distribution.
For the first missing visit, condition on the observed history and calculate its reference-based conditional mean and variance.

Draw and append the value.
Generate a normal draw from that conditional distribution and append it to the history.

Repeat through the final visit.
At each later visit, condition on all observed values and all values generated so far; then draw from the corresponding conditional normal distribution.

Check the completed trajectory.
Confirm that observed values were retained, every missing visit was generated once, and the later conditional calculations used the updated history.

Worked Example: Dropout Before Three Remaining Visits

Suppose a participant has \(Y_1=14\) observed and visits 2 through 4 missing. Assume a reference-arm Gaussian AR(1) model with mean vector \((10,11,13,14)\), marginal variance \(9\), and adjacent correlation \(\rho=0.4\). As in the first example, the AR(1) Markov structure means the conditional mean at each step uses the immediately preceding outcome, while the conditional variance is \(9(1-0.4^2)=7.56\). Thus the conditional standard deviation is \(\sqrt{7.56}\approx2.749\).

At visit 2, the mean is \(11+0.4(14-10)=12.6\). With \(Z_2=1\), \(Y_2^{\mathrm{imp}}=12.6+2.749(1)\approx15.349\).

At visit 3, update the history with \(15.349\). The mean is \(13+0.4(15.349-11)\approx14.740\). With \(Z_3=-1\), \(Y_3^{\mathrm{imp}}=14.740-2.749\approx11.991\).

At visit 4, use the generated visit 3 value. The mean is \(14+0.4(11.991-13)\approx13.596\). With \(Z_4=0.5\), \(Y_4^{\mathrm{imp}}=13.596+2.749(0.5)\approx14.971\). This non-monotone sequence is possible: sequential J2R draws follow the specified conditional distribution, rather than enforcing a smooth or steadily changing trajectory.

Diagnostics, Alternatives, and Limitations

A sequential implementation should be checked for both model and coding errors. Confirm that the reference covariance matrix is positive definite and that each conditional variance is nonnegative; a negative value usually signals a covariance or indexing error. Check that the observed treatment-arm history is not inadvertently replaced by reference-arm means or simulated reference outcomes. For every later missing visit, verify that the conditioning set contains all required observed outcomes and the previously generated outcomes. In a multivariate model, omitting an earlier observed visit can change both the conditional mean and variance.

A useful computational check is to generate many complete missing sequences under a fixed model and compare their empirical means and covariances with the corresponding reference conditional distribution given the observed history. Large discrepancies can reveal a faulty update, a visit-order mismatch, or an incorrect covariance submatrix. Individual trajectories may look unusual by chance, so diagnostics should assess collections of imputations as well as single paths.

Mean-only imputation replaces a missing outcome by a conditional mean and does not represent residual uncertainty in that outcome. Drawing each missing visit independently from its marginal reference distribution also ignores the observed history and dependence across visits. Sequential J2R instead uses conditional distributions, so both the observed history and prior generated values can affect subsequent predictions. Single-visit J2R uses the same conditional-normal logic for one missing outcome, but it does not require the repeated history update needed for several missing visits.

J2R is a modeling assumption about post-dropout outcomes, not a fact established by the observed data. Results can depend on the choice of reference arm, the adequacy of the reference mean and covariance model, and the amount of information available in that arm. The calculation also relies on a suitable conditional distribution; the Gaussian formulas above do not automatically apply to outcomes or models that are not multivariate normal. Sequential generation specifies how to draw outcomes under the chosen model; by itself, it does not account for uncertainty in estimated model parameters or complete the inferential analysis.

Common pitfall: Treating later missing visits as independent draws, or predicting each one from the observed history alone, leaves out the generated outcomes that should update the history. The correction is to append each draw immediately and use the enlarged history at the next step. This is what makes the sequence follow the joint conditional model rather than a set of disconnected predictions.

Check Your Understanding

Answer each question, then compare with the solutions below.

A participant has observed outcomes through visit 2 and visits 3 and 4 are missing. What information belongs in the history when imputing visit 4?
Under a Gaussian reference model, what are the conditional mean and variance of a missing outcome given a history?
Why are sequentially generated outcomes generally dependent even if independent standard-normal innovations are used?
In an AR(1) reference model with marginal variance \(16\) and adjacent correlation \(0.5\), what is the conditional variance of the next visit given the immediately preceding visit?
Name one diagnostic for detecting an error in sequential imputation and one limitation of the J2R assumption.

Solutions. (1) Use the observed outcomes through visit 2 and the generated outcome at visit 3. (2) The conditional mean is \(\mu_{R,t}+\Sigma_{R,t,H}\Sigma_{R,H,H}^{-1}(H-\mu_{R,H})\), and the conditional variance is \(\Sigma_{R,t,t}-\Sigma_{R,t,H}\Sigma_{R,H,H}^{-1}\Sigma_{R,H,t}\). (3) Each later outcome is conditioned on earlier outcomes, so the outcomes share dependence through the sequential conditional distributions. (4) The variance is \(16(1-0.5^2)=12\). (5) For example, compare empirical means and covariances from many generated sequences with the model’s reference conditional distribution. One limitation is that J2R depends on the assumption that post-dropout outcomes follow the selected reference-arm distribution.