Why a rise after withdrawal is not enough
An outcome can worsen after a drug is stopped for several different reasons. The disease may have progressed while the patient was treated; the drug's benefit may simply have faded; or the outcome may have risen beyond the level expected without treatment. Only the last pattern is evidence of rebound beyond the untreated disease trajectory. The distinction matters because a post-withdrawal rise, by itself, does not show that stopping treatment caused an excess worsening.
The comparison must be made at the same follow-up time. A value above the patient's pretreatment measurement may still be ordinary disease progression if the untreated disease would also have worsened. Conversely, an outcome can be rising from its on-treatment value yet remain below the untreated trajectory, which is consistent with a fading benefit rather than rebound.
The untreated trajectory is a counterfactual: for a person who received treatment, the outcome that same person would have had without treatment cannot be observed at the same time. In practice, it must be estimated from an appropriate untreated control group, a randomized design, or a justified disease-course model. The definition is about the comparison to that trajectory, not about whether the post-withdrawal value exceeds baseline.
Constructing the comparison
A useful descriptive contrast compares changes from baseline in a withdrawal group with changes over the same period in an untreated control group. Write the group means as \(\bar{Y}_{W,0}\) and \(\bar{Y}_{W,t}\) for the withdrawal group at baseline and follow-up, and \(\bar{Y}_{C,0}\) and \(\bar{Y}_{C,t}\) for untreated controls. The baseline-adjusted contrast is
A positive \(\widehat{\Delta}(t)\) means the withdrawal group worsened more than the controls over that interval. It estimates excess worsening relative to the control trajectory; it does not, on its own, establish that the excess is statistically distinguishable from zero or that withdrawal caused it. If the groups differ at baseline, the change contrast removes a fixed baseline offset, but it does not correct for different underlying time trends.
The key assumption behind a simple baseline-adjusted contrast is that, without withdrawal, the groups would have followed comparable trajectories over the observation period. Random assignment can support this assumption, but it does not guarantee it in a small sample or when follow-up, co-interventions, or outcome measurement differ. With nonlinear progression, the untreated trajectory need not be a straight line: time-matched controls can capture curvature that a linear extrapolation would miss.
Worked examples
Worked Example: A Rise That Matches Disease Progression
Suppose a symptom score is measured in points, with larger scores indicating worse disease. A hypothetical withdrawal group has a mean score of 8.2 at baseline and 9.1 two months later. Untreated controls have means of 8.0 at baseline and 8.8 at two months. Is the rise in the withdrawal group evidence of rebound?
The withdrawal group's mean increase is \(9.1-8.2=0.9\) points. The untreated controls' mean increase is \(8.8-8.0=0.8\) points. Thus,
The withdrawal group's score rose, but nearly the same increase occurred in untreated controls. The estimated excess worsening is only 0.1 points, so these means do not show a substantial departure from the untreated trajectory. Without uncertainty estimates, we also cannot determine whether this small contrast is distinguishable from ordinary sampling variation. Calling the full 0.9-point rise “rebound” would wrongly count expected disease progression as an effect of withdrawal.
Worked Example: Excess Worsening Beyond the Control Trajectory
In a second hypothetical study, a disease-activity score is higher when disease is worse. At withdrawal baseline, the mean is 10.2 points in the withdrawal group and 10.0 points in untreated controls. At month 4, the corresponding means are 13.1 and 12.0 points. Assume a prespecified model estimates the baseline-adjusted contrast as 0.9 points with \(SE=0.30\) points, and that a normal-approximation 95% confidence interval is suitable for this illustration. Does the pattern support rebound?
The group's changes are \(13.1-10.2=2.9\) points and \(12.0-10.0=2.0\) points. The difference in changes is
For the stated normal approximation, the 95% confidence interval is
Under the model and assumptions used to compute this interval, the estimated excess worsening is positive, and the interval excludes zero. This supports an average outcome above the untreated trajectory at month 4. It is not a 95% probability statement about the true contrast, nor does the interval alone establish that withdrawal caused the difference. That causal interpretation also depends on the design and the credibility of the untreated comparison.
Worked Example: Timing Reveals a Temporary Excess
Consider hypothetical weekly symptom-score means for a withdrawal group and untreated controls. Both groups have a baseline mean of 10 points. At week 1, the withdrawal group mean is 14 and the control mean is 12. At week 4, the means are 12.4 and 12.8, respectively. Larger scores indicate worse symptoms. Calculate the baseline-adjusted contrasts and interpret the pattern.
At week 1, the contrast is
At week 4, it is
The week-1 result is consistent with a transient excess beyond the control trajectory, whereas the week-4 contrast shows no excess in these means: the withdrawal group is 0.4 points below the controls. The pattern could reflect a short-lived rebound followed by convergence, but these values alone do not establish a mechanism or statistical certainty. Interpretation also depends on whether week 1 is a meaningful post-withdrawal time point and whether residual drug exposure or delayed pharmacodynamic effects could still influence either group.
Justifying the baseline-adjusted contrast
The change contrast is useful because it removes a fixed difference in starting levels while retaining differences in change. To see this, suppose each group's observed mean at time \(t\) can be written as its baseline level plus a time-related change. Let \(B_W\) and \(B_C\) be the group baselines, and let \(g_W(t)\) and \(g_C(t)\) be their changes. Then
Subtracting each group's own baseline gives
Thus, the baseline levels cancel, and the contrast measures how much more the withdrawal group changed. This algebra does not make the groups' changes comparable by itself. Interpreting \(g_W(t)-g_C(t)\) as rebound requires the control change to represent what the withdrawal group would have experienced without treatment-related effects. In particular, if the groups would naturally progress at different rates, the contrast can be biased even though baseline differences cancel.
Withdrawal timing, diagnostics, and alternatives
The time origin and withdrawal date must be stated clearly. A control comparison made at a different disease stage, or at a different interval since baseline, may not represent the relevant untreated course. Drug concentrations may also persist after the last dose, and pharmacodynamic effects can lag behind concentration changes. An outcome measured immediately after the final dose may therefore reflect ongoing exposure or delayed response rather than the post-washout state. Report time since last dose and, where relevant, exposure or pharmacodynamic evidence; avoid treating “dose stopped” as equivalent to “drug effect ended.”
Useful diagnostics include plotting individual and group trajectories from before treatment through follow-up, examining whether the groups had comparable pre-withdrawal trends, and checking that outcome measurement and follow-up timing were consistent. A sudden divergence at the withdrawal point can be informative, but is not conclusive: differential dropout, rescue medication, changes in assessment, or selection of patients for withdrawal may also cause divergence. If there are few controls or sparse measurements, the untreated trajectory may be too uncertain to classify the pattern confidently.
| Observed pattern after withdrawal | Most appropriate interpretation |
|---|---|
| Outcome rises but follows the time-matched untreated trajectory | Consistent with ordinary progression or return toward untreated disease |
| Outcome rises but remains below the untreated trajectory | Consistent with loss of some benefit while retaining a lower disease level than untreated |
| Outcome exceeds the untreated trajectory | Consistent with rebound, if the excess is credible and not explained by bias or other changes |
| Untreated trajectory is unavailable or poorly estimated | Insufficient information to distinguish rebound from progression reliably |
A pretreatment baseline is a simple reference but cannot represent what untreated disease would have done later when progression is expected. A within-person pre/post comparison accounts for each person's starting level but still cannot separate withdrawal-associated change from time trends, regression to the mean, or other changes occurring over time. Concurrent untreated controls provide a direct time-matched comparison, while a continued-treatment group can help distinguish withdrawal from changes occurring in patients who remain treated. Neither comparison is automatically decisive: the control condition must answer the counterfactual question of interest, and group comparability and follow-up must be credible.
Common pitfall and limitations
Wrong reasoning: “The patient's score was 9 before treatment and 11 after stopping, so stopping caused rebound.” This compares two observations but says nothing about the untreated score expected at the later time. If untreated disease would have reached 12, a score of 11 is below the untreated trajectory, despite being higher than baseline.
Correction: Compare the post-withdrawal outcome with a defensible estimate of untreated disease at the same time. Use the disease context, control observations, baseline trends, and withdrawal timing to judge that estimate. When evidence is insufficient, report that rebound cannot be distinguished from progression rather than assigning a mechanism.
Even a well-designed comparison has limits. A group average can conceal patients with different trajectories; a positive average contrast does not mean every individual experienced rebound. A model-based untreated trajectory depends on its form and assumptions, and estimates can be unstable when follow-up is short or controls are few. Finally, an excess beyond untreated levels is a description of the outcome pattern. It does not by itself identify the biological process that produced the pattern.
Check Your Understanding
Answer each question using the principle that comparisons must be made against the time-matched untreated trajectory.
- A disease score is higher when disease is worse. A withdrawal group rises by 1.5 points over six weeks, while untreated controls rise by 1.4 points. What does the difference in changes suggest?
- Why does an outcome above pretreatment baseline not, by itself, establish rebound?
- At baseline, a withdrawal group has a mean score of 7.0 and controls have 7.5. At follow-up, their means are 9.0 and 10.0. Calculate the baseline-adjusted contrast and interpret it.
- Name two reasons why a measurement taken immediately after the last dose might not represent the outcome after drug effects have ended.
- What can be concluded if the post-withdrawal trajectory appears above the estimated untreated trajectory but there is no concurrent control group and the disease course is uncertain?
Solutions. (1) The estimated difference in changes is \(1.5-1.4=0.1\) points. The rises are similar, so the means do not show substantial excess worsening beyond the control trajectory; uncertainty would be needed to assess precision. (2) Untreated disease may naturally worsen over time, so the relevant comparison is the expected untreated outcome at follow-up, not the earlier baseline value. (3) The contrast is \((9.0-7.0)-(10.0-7.5)=2.0-2.5=-0.5\) points. The withdrawal group worsened 0.5 points less than controls, so these means do not indicate rebound. (4) Drug concentration may persist after the final dose, and pharmacodynamic effects may be delayed. (5) The pattern is suggestive but cannot be classified reliably without a credible estimate of the untreated trajectory; other differences or model assumptions may explain the apparent excess.