Assessing Durability of Dose-Response Effects

Why Durability Is a Dose-Response Question

A dose may produce a useful effect shortly after treatment begins but a smaller effect later; alternatively, a dose difference may take time to emerge. A single average over the full treatment period can obscure either pattern. Assessing durability therefore means examining how dose effects change over follow-up, not merely asking whether treatment has an effect at one visit.

The analysis should distinguish three questions: Is the dose effect still clinically relevant late in follow-up? Has the effect changed by an amount that matters? And do the dose contrasts themselves persist, diminish, or emerge? These questions are related but not interchangeable. A statistically detectable change over time does not by itself establish that the late effect is inadequate, and an undetected change does not prove that the effect has been maintained.

The introductory material on accounting for treatment duration motivates specifying the treatment period represented by a dose justification. Here the focus is the time pattern within that period. Conclusions depend on the endpoint, follow-up visits, population, and effect definition being clinically meaningful and specified clearly.

Definition: Let $\mu_d(t)$ be the mean change from baseline at follow-up time $t$ for the group assigned dose $d$, and let dose $0$ denote placebo. Define the time-specific dose effect as $E_d(t)=\mu_d(t)-\mu_0(t)$. When positive change is beneficial, positive $E_d(t)$ favors dose $d$ over placebo. A dose contrast between doses $d_1$ and $d_2$ is $\Delta_{d_1,d_2}(t)=\mu_{d_1}(t)-\mu_{d_2}(t)$. Durability concerns the trajectory of these effects or contrasts over the specified follow-up.

The sign convention must match the endpoint. For an endpoint where lower values are better, either reverse the contrast or state explicitly that a negative value favors treatment. Likewise, define whether the outcome is change from baseline, an observed value adjusted for baseline, or another prespecified summary. The examples below use change from baseline on a scale where larger values are better; effects are measured in score points.

Modeling the Time Pattern

For visits at which outcomes are measured repeatedly, one common approach is a longitudinal model with categorical visit, dose group, and dose-by-visit terms. Categorical visit allows the dose effect to take a different value at each visit without imposing a straight-line trend. The model also accounts for the within-participant correlation among repeated measurements. A dose effect at a specified visit is obtained from the model-estimated difference between that dose and placebo at the visit.

This approach requires an appropriate mean structure and a defensible model for within-participant covariance. Randomization supports comparisons between assigned groups, subject to the trial’s conduct and the chosen estimand. Likelihood-based inference with incomplete outcomes commonly relies on a missing-at-random assumption conditional on observed information included in the analysis; this assumption is not guaranteed by randomization. The analysis should also address the consequences of treatment discontinuation or other intercurrent events according to the estimand being targeted.

For a direct durability assessment, compare the same dose effect at an early and a late visit. Define the change in effect as $C_d=E_d(t_{\mathrm{late}})-E_d(t_{\mathrm{early}})$. With the sign convention above, $C_d=0$ indicates no change between those visits, $C_d<0$ indicates attenuation, and $C_d>0$ indicates an increase. A dose-by-time interaction tests whether effects differ over time; its estimated contrasts and uncertainty are more informative than the interaction p-value alone.

Justification for the contrast uncertainty: The estimated change is $\widehat{C}_d=\widehat{E}_d(t_{\mathrm{late}})-\widehat{E}_d(t_{\mathrm{early}})$. Since both estimates usually come from the same participants, their covariance matters: $$ SE(\widehat{C}_d)=\sqrt{\operatorname{Var}(\widehat{E}_d(t_{\mathrm{late}}))+\operatorname{Var}(\widehat{E}_d(t_{\mathrm{early}}))-2\operatorname{Cov}(\widehat{E}_d(t_{\mathrm{late}}),\widehat{E}_d(t_{\mathrm{early}}))}. $$ A confidence interval is then formed using the model’s appropriate reference distribution and degrees of freedom, or a justified large-sample approximation. Treating the two visit estimates as independent would omit the covariance term and can give the wrong standard error.

A confidence interval for $E_d(t_{\mathrm{late}})$ addresses the size and precision of the late effect. A confidence interval for $C_d$ addresses change over time. If the goal is to establish that erosion is not clinically important, define a clinically acceptable erosion margin in advance and assess whether the interval for $C_d$ is sufficiently compatible with that margin. A nonsignificant test of $C_d=0$ alone is not evidence of equivalence or maintained effect.

Worked Example: An Effect That Appears Maintained

Suppose a hypothetical randomized trial compares a dose with placebo on a symptom score, where higher change from baseline is better. The estimated dose effects are 2.0 points at week 4 and 2.1 points at week 24. The model estimates $SE(\widehat{C}_d)=0.55$ points for the late-minus-early contrast. A clinically important effect has been defined as 1 point.

The estimated change is $$ \widehat{C}_d=2.1-2.0=0.1\text{ points}. $$ Using a normal approximation, its 95% confidence interval is $$ 0.1\pm1.96(0.55)=0.1\pm1.078=(-0.978,\ 1.178)\text{ points}. $$ The interval includes zero, so these data do not establish a change in either direction at the 5% level using this approximation. However, this is not proof that the effect was maintained: the interval also permits an erosion of nearly 1 point or an increase of more than 1 point.

At week 24, the estimated effect is 2.1 points. If its standard error is 0.45 points, the 95% interval is $2.1\pm1.96(0.45)=(1.218,\ 2.982)$ points. Its lower limit exceeds the 1-point clinical threshold, supporting a clinically important late effect under this analysis. The combined interpretation is more informative than “no significant interaction”: there is evidence of a meaningful late effect, while the precision of the change estimate does not rule out some erosion.

Persistence, Attenuation, and Emergence

The pattern should be described with estimates at relevant visits and contrasts that quantify change. “Persistence” is not a claim that every visit has exactly the same effect. It usually means that the late effect remains clinically relevant and that any change is not large enough to undermine the intended dose rationale. “Attenuation” means the effect decreases over follow-up; whether that decline matters depends on its magnitude and the late effect. “Emergence” means that a dose contrast is small earlier and becomes larger later. A delayed contrast can be missed if only an early endpoint is examined.

A global dose-by-time test can help identify whether the dose-response pattern varies across visits, but it does not identify which contrast changed or whether the change matters clinically. Follow-up contrasts should reflect the scientific question, with attention to multiplicity when many doses and visits are examined. If the dose-response curve is not known to be monotonic, do not impose a linear dose trend simply for convenience; estimate or contrast the relevant dose groups using a suitable prespecified model.

Worked Example: A Dose Difference That Emerges Later

In a hypothetical trial, the estimated high-dose minus low-dose contrast on a beneficial score is 0.3 points at week 4 and 1.4 points at week 24. The estimated change in that contrast from week 4 to week 24 is 1.1 points, with standard error 0.50 points.

The time change is $$ \widehat{C}_{\mathrm{high,low}}=1.4-0.3=1.1\text{ points}. $$ Its approximate 95% confidence interval is $$ 1.1\pm1.96(0.50)=1.1\pm0.98=(0.12,\ 2.08)\text{ points}. $$ The corresponding normal test statistic for no change is $z=1.1/0.50=2.20$, giving a two-sided $p$-value of approximately 0.028. This supports a change in the dose contrast over time, with the estimated difference larger at week 24.

Suppose the late high-dose versus low-dose contrast has standard error 0.55 points. Its 95% interval is $1.4\pm1.96(0.55)=(0.322,\ 2.478)$ points. If 1 point is the prespecified clinically important dose difference, this interval does not establish that the late contrast exceeds that threshold: it includes values below 1. The results suggest later separation but do not, by themselves, establish a clinically important advantage of the high dose. The comparison also does not show whether either dose has a meaningful effect versus placebo; that requires the corresponding dose-placebo estimates.

Worked Example: An Effect That Attenuates

In another hypothetical trial, a dose-placebo effect is estimated as 3.0 points at week 4 and 1.1 points at week 24. The late-effect standard error is 0.46 points, and the standard error for the late-minus-early change is 0.65 points. The clinically important threshold is 1.5 points.

The late effect has approximate 95% interval $$ 1.1\pm1.96(0.46)=1.1\pm0.902=(0.198,\ 2.002)\text{ points}. $$ The point estimate is below the 1.5-point threshold, but the interval includes values both below and above it. Thus, the late data do not precisely determine whether the effect meets that threshold.

The estimated change is $1.1-3.0=-1.9$ points, indicating attenuation. Its 95% interval is $$ -1.9\pm1.96(0.65)=-1.9\pm1.274=(-3.174,\ -0.626)\text{ points}. $$ This interval excludes zero, supporting a decrease in the estimated effect over follow-up. It does not prove that the late effect is absent: the late-effect interval is positive. The appropriate conclusion is that the effect appears to attenuate substantially, while evidence about whether it remains clinically important at week 24 is inconclusive. A sustained-dose rationale should reflect both findings rather than relying only on the early response.

Diagnostics and Common Pitfalls

A durability conclusion is only as credible as the time-specific estimates supporting it. Inspect estimated dose effects and dose contrasts by visit with uncertainty intervals, alongside observed data summaries. Check whether the fitted mean pattern tracks the data, whether residual behavior or influential observations raise concerns, and whether the covariance model is adequate for the repeated measurements. Sparse late-visit data can make apparent persistence or erosion highly uncertain.

Also examine visit attendance, discontinuation, rescue medication, and other sources of missingness or changes in treatment. If participants with poorer response are more likely to leave, a complete-case late comparison can misrepresent the randomized groups. Consider sensitivity analyses under plausible departures from the primary missing-data assumptions. Floor or ceiling effects, changes in endpoint interpretation over time, and a visit schedule that misses the expected onset or loss of effect can also distort the apparent pattern.

Mistake: “The interaction was not significant, so the effect is durable.” A nonsignificant test can reflect limited precision. Examine the interval for the change and the late effect; use a prespecified erosion margin if the aim is to rule out clinically important loss.
Mistake: “The early effect was large, so the dose is suitable for sustained use.” An early estimate does not establish a clinically relevant late effect. Report the late estimate and its uncertainty in context.
Mistake: “A significant interaction means the higher dose should be selected.” An interaction indicates time variation in a contrast, not that a particular dose is clinically preferable. Interpret the direction, magnitude, late effect, clinical threshold, and relevant safety evidence.

How the Evidence Supports a Sustained-Dose Rationale

For a sustained-dose recommendation, prioritize the planned treatment duration and the latest clinically relevant follow-up at which evidence is adequate. Summarize the late dose-placebo effects, dose contrasts, and their uncertainty; then describe how those estimates changed from earlier visits. A persuasive rationale connects the time pattern to a clinical threshold and acknowledges uncertainty. It should not claim that an effect will persist beyond the observed follow-up.

The longitudinal model is not the only possible analysis. Separate comparisons at each visit are simple to interpret but may be less efficient and can create a multiplicity problem. A continuous-time trend or slope model can be more parsimonious when the trajectory is plausibly smooth, but may miss delayed onset or nonlinear loss. A time-to-loss-of-response analysis can address a defined event but requires a defensible response threshold and does not convey the full average trajectory. These alternatives answer different questions; none removes the need to define the estimand, examine missingness, and report uncertainty.

Limitations should be explicit. Durability is bounded by the duration and quality of follow-up, the endpoint’s sensitivity, adherence and discontinuation patterns, model assumptions, and the precision of late estimates. Results from one population or regimen do not automatically establish a time pattern for another. The evidence can support a dose rationale over observed follow-up, but it cannot guarantee individual response or long-term persistence outside the studied conditions.

Check Your Understanding

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

If a dose effect is 2.4 points at week 4 and 2.0 points at week 20, what contrast represents change in effect, and what does its sign mean?
Why is a nonsignificant dose-by-time test insufficient to establish durability?
A high-minus-low dose contrast is 0.2 points early and 1.0 point late. Name one additional estimate needed before claiming a clinically important sustained dose advantage.
If the late-effect interval includes zero but the early effect is clearly positive, what should be reported?
Why can ignoring within-participant covariance produce an incorrect standard error for change over time?

Solutions. (1) The late-minus-early change is $2.0-2.4=-0.4$ points, indicating an estimated attenuation of 0.4 points. Its uncertainty is needed to judge the evidence for change. (2) The test may have low power; its failure to reject no time variation does not show that important erosion is ruled out. Examine the change interval and the late effect, ideally against a prespecified erosion margin. (3) Report the late high-dose versus low-dose estimate with its confidence interval and compare it with a prespecified clinically important difference; also examine each dose’s effect versus placebo. (4) Report that the late estimate is imprecise and compatible with no effect as well as any positive values included by the interval; do not infer that the effect disappeared or persisted from the early result alone. (5) The two estimates come from repeated measurements on the same participants, so their covariance contributes to the variance of their difference: $\operatorname{Var}(A-B)=\operatorname{Var}(A)+\operatorname{Var}(B)-2\operatorname{Cov}(A,B)$.