What Copy-Reference Copies
When a participant assigned to an active treatment stops treatment, the outcomes that would have followed are unobserved. A reference-based imputation rule specifies an assumed distribution for those outcomes. Copy-reference says that, after dropout, the participant’s outcome distribution is borrowed from the reference arm. This is a distributional restriction: it concerns the whole post-dropout distribution, not just its average.
Let \(G=T\) denote assignment to the active treatment arm and \(G=R\) assignment to the reference arm. Write \(Y_1,\ldots,Y_K\) for outcomes at successive visits. For a dropout pattern \(d\), let \(H_d\) denote the participant’s observed outcome history up to dropout, and let \(Z_d\) denote the vector of outcomes after dropout whose distribution is to be imputed. The reference distribution used for borrowing describes \(Z_d\) among reference-arm participants with a comparable observed history.
The equality is conditional on the history used to make the prediction. If a model also conditions on baseline covariates, those covariates must be included on both sides of the restriction, at the same values. The reference distribution therefore needs to be defined for histories and covariate values relevant to the active-arm dropout pattern. An unconditional equality of distributions is a different restriction and should not be substituted silently for this conditional one.
Copying a distribution copies its mean and covariance, as well as other features specified by that distribution. For example, if the reference predictive law is multivariate normal,
then copy-reference assigns the active-arm dropout pattern the same predictive mean vector \(m_R(h)\) and covariance matrix \(\Sigma_R(h)\). In a multivariate normal model, these parameters determine the conditional distribution. Without a normal assumption, matching means and covariances alone does not in general guarantee equality of the full distributions.
Mean and Covariance Restrictions
For a post-dropout vector with \(q\) visits, the copied mean has \(q\) components and the copied covariance is a \(q\times q\) matrix. Its diagonal entries are conditional variances at individual visits; its off-diagonal entries describe conditional covariances between visits. Thus copy-reference constrains both the uncertainty at each post-dropout visit and the dependence across those visits, according to the chosen reference predictive distribution.
The copied covariance is the covariance of \(Z_d\) conditional on the history in the restriction. It is not, by itself, a statement that every covariance involving pre-dropout outcomes is copied. Those cross-covariances depend on the joint model and on how the predictive distribution is constructed. Keeping that distinction clear prevents a conditional distribution restriction from being misread as an unrestricted equality of all covariance components in the trial.
Worked Example: Copying a Two-Visit Distribution
Suppose a hypothetical trial has two post-dropout outcomes, measured in score points. At a specified observed history \(h\), the reference-arm predictive distribution is normal with mean vector \((8,7)\) and covariance matrix $$ \Sigma_R(h)= \begin{pmatrix} 4 & 1.2\\ 1.2 & 9 \end{pmatrix}. $$ What distribution does copy-reference assign to the active-arm participants with this dropout pattern and history?
The restriction copies the whole predictive distribution. The imputed vector has mean \((8,7)\) score points and covariance matrix equal to the reference matrix: $$ Z_d\mid H_d=h,G=T,D=d \sim N\!\left( \begin{pmatrix}8\\7\end{pmatrix}, \begin{pmatrix}4&1.2\\1.2&9\end{pmatrix} \right). $$
The first and second conditional variances are \(4\) and \(9\) score-points-squared, so the conditional standard deviations are \(\sqrt{4}=2\) and \(\sqrt{9}=3\) score points. The correlation is $$ \frac{1.2}{\sqrt{4}\sqrt{9}}=\frac{1.2}{2(3)}=0.2000. $$ Thus the copy includes a positive association between the two post-dropout outcomes. A rule that copied only the two means would not, on its own, specify either variance or this correlation.
The distributional rule can be stated without requiring normality. Normal models are common because they give a compact description through means and covariance matrices and support conditional prediction. The modeling choice must nevertheless be stated: equality of mean and covariance is equivalent to equality of distributions only under additional distributional assumptions, such as the specified multivariate normal form.
How the Restriction Is Applied
The main steps are to identify the dropout pattern, define its post-dropout outcomes and observed history, specify the reference predictive law, and impose that law on the active-arm pattern. This is an identifying assumption about unobserved outcomes, not a conclusion established merely by the observed trial data.
State which outcomes are observed at dropout and which later outcomes make up \(Z_d\).
Record the history \(H_d\) and any baseline covariates used to predict post-dropout outcomes.
Use the stated reference-arm model to determine the distribution at the relevant history and covariate values.
Assign that predictive distribution, including its mean and covariance, to the active-arm pattern after dropout.
The key assumption is that, after conditioning on the information specified in the model, the reference-arm predictive distribution is an appropriate distribution for the active-arm participants’ unobserved post-dropout outcomes. This assumption may be motivated by a clinical interpretation of what happens after treatment discontinuation, but it is not verified simply because the reference arm’s observed data fit the model.
Worked Example: Copying a Conditional Predictive Distribution
In a hypothetical study, suppose that for a particular observed history \(h\), a fitted reference-arm model gives $$ Z_d\mid H_d=h,G=R \sim N\!\left( \begin{pmatrix}6.4\\6.0\end{pmatrix}, \begin{pmatrix}2.25&0.90\\0.90&4.00\end{pmatrix} \right), $$ where the outcomes are measured in score points. Apply copy-reference to active-arm participants with that pattern and history, and summarize the covariance.
By the restriction, the active-arm pattern receives the same conditional normal distribution: $$ Z_d\mid H_d=h,G=T,D=d \sim N\!\left( \begin{pmatrix}6.4\\6.0\end{pmatrix}, \begin{pmatrix}2.25&0.90\\0.90&4.00\end{pmatrix} \right). $$ The copied conditional standard deviations are \(\sqrt{2.25}=1.5\) and \(\sqrt{4.00}=2.0\) score points. The copied conditional correlation is $$ \frac{0.90}{\sqrt{2.25}\sqrt{4.00}} = \frac{0.90}{1.5(2.0)} = 0.3000. $$
These quantities are conditional on \(H_d=h\). They are not automatically the marginal means, variances, or correlation across all participants in the pattern, because those marginal quantities also depend on how histories vary across participants.
Conditional Borrowing Is Not Marginal Equality
A copied predictive distribution can depend on the observed history. Consequently, its conditional mean at a particular history need not equal the reference arm’s overall marginal mean. This is a direct consequence of conditioning, rather than a contradiction in the copy-reference rule.
Worked Example: The Copied Mean Depends on History
Suppose, hypothetically, the reference-arm conditional mean of one post-dropout score is $$ E_R(Z_d\mid H_d=h)=10+0.5(h-8), $$ and the conditional variance is \(4\) score-points-squared. Assume the active-arm participant’s observed history is \(h=12\). Under copy-reference, what conditional mean and variance are assigned?
Substituting the observed history into the reference conditional mean gives $$ 10+0.5(12-8)=10+0.5(4)=12 $$ score points. The copied conditional variance is \(4\) score-points-squared, so the conditional standard deviation is \(\sqrt{4}=2\) score points. The active-arm pattern is assigned the reference predictive distribution at \(h=12\), not the reference mean averaged over all possible histories.
For comparison, at \(h=8\), the same conditional-mean model gives \(10+0.5(8-8)=10\) score points. The two values differ because they refer to different histories. This example assumes the stated reference conditional-mean and variance model is appropriate; it does not establish that model from the formula alone.
In applying a restriction, check whether the reference model can provide predictions for the active-arm histories of interest. If the relevant histories or covariate values lie outside the reference arm’s observed range, the restriction may require extrapolation. Estimates can then be sensitive to the model’s functional form and to covariance assumptions, even if the numerical procedure returns predictions.
Comparison, Diagnostics, and Limitations
| Restriction | What it constrains after dropout | What remains to be specified |
|---|---|---|
| Copy-reference distribution | The reference predictive distribution, including its mean and covariance, conditional on the specified information | The reference model, conditioning set, and any assumptions needed to identify its predictions for the target histories |
| Mean-only restriction | A specified post-dropout mean or mean trajectory | Variances, covariances, and the remaining distributional form |
| Other pattern-mixture restriction | The particular features stated for each missingness pattern | Any distributional features not constrained by that restriction |
Copy-reference is more restrictive than a mean-only rule: it supplies a distributional target rather than leaving all dispersion and dependence features open. That stronger specification can make imputation operational, but it also means that conclusions depend on more assumptions. Earlier material on distributional restrictions versus mean restrictions provides the general distinction; here the defining choice is to use the reference arm’s predictive law for the post-dropout vector.
Useful checks include verifying that the reference predictive covariance matrix is symmetric and positive semidefinite, inspecting the fitted model’s predictions across the histories being borrowed, and checking whether the model gives plausible marginal summaries after the restrictions are applied. Extreme means, implausible variances, or unstable predictions across nearby histories can signal model misspecification or weak support. Such checks assess coherence and model behavior; they do not test whether the unobserved active-arm outcomes truly follow the reference distribution.
- Limited overlap: If few reference participants have histories comparable to a treatment-arm dropout pattern, the predictive law may be driven by extrapolation.
- Covariance misspecification: An implausible reference covariance structure can distort simulated trajectories even when the means appear reasonable.
- Unstated conditioning: Borrowing an unconditional reference distribution when the intended rule is conditional on history can create a different restriction.
- Overinterpretation: Copy-reference is an assumption about missing outcomes, not evidence that treatment cessation causes reference-arm behavior.
The rule is therefore best reported with the dropout pattern, conditioning variables, reference model, and distributional features copied made explicit. Its estimates should be understood as conditional on these choices, alongside the observed-data model and other analysis assumptions.
Check Your Understanding
Answer each question using the stated conditioning information and distributional assumptions.
- A reference predictive distribution for two post-dropout visits has mean vector \((5,4)\) and covariance matrix \(\begin{pmatrix}1&0.3\\0.3&4\end{pmatrix}\). Under copy-reference, what mean vector and covariance matrix are assigned to the active-arm pattern?
- For the covariance matrix in question 1, calculate the conditional correlation.
- Explain why matching the two conditional means alone does not establish copy-reference when the distribution is not specified to be multivariate normal.
- A copied conditional mean is \(9\) score points at one observed history and \(11\) at another. Does this necessarily violate copy-reference? Explain.
- Name one diagnostic that can identify a concern with the reference predictive distribution and one limitation of that diagnostic.
Solutions. (1) The assigned mean is \((5,4)\), and the assigned covariance matrix is \(\begin{pmatrix}1&0.3\\0.3&4\end{pmatrix}\), because copy-reference borrows the full stated predictive distribution. (2) The standard deviations are \(1\) and \(2\), so the correlation is \(0.3/(1\cdot2)=0.1500\). (3) Different distributions can have the same means but different variances, dependence, or shapes; matching means alone does not determine the full distribution. (4) No. Conditional means can vary with observed history, so different values are compatible with copying the reference conditional distribution at each history. (5) One can inspect whether predicted means and variances are plausible across relevant histories. This can reveal model instability or implausible predictions, but it cannot establish that the unobserved active-arm outcomes truly follow the reference distribution.