Pooling Dose-Response Evidence Across Phase 2 Trials

Why Pool Dose-Response Evidence Across Trials?

Phase 2 trials can study different dose ranges, enroll different populations, or be conducted at different times. Their average outcomes may therefore differ even when the drug’s treatment effect is similar. A joint analysis can use the dose-response information from all trials while allowing each trial its own baseline level. The central modeling decision is which parameters to share: pooling can improve precision when effects are genuinely common, but can obscure meaningful differences when they are not.

The aim here is to estimate a common dose-response relationship while representing baseline differences explicitly. “Common” does not mean that every trial has the same outcome at every dose. It means that, after accounting for the specified trial-specific baseline, the dose-related parameters are shared. The earlier material on choosing dose-response models supplies candidate curve forms; the pooling task is to combine their evidence without treating trial baselines as treatment effects.

Joint Model and Assumptions

Definition: Let $y_{ij}$ be the estimated mean outcome in trial $j$ at dose $d_{ij}$, and let $f(d;\theta)$ be a dose-response function with $f(0;\theta)=0$. A joint model with trial-specific baselines and shared treatment-effect parameters is $$ y_{ij}=\alpha_j+f(d_{ij};\theta)+\epsilon_{ij}, $$ where $\alpha_j$ is the expected outcome at zero dose in trial $j$, $\theta$ contains common treatment-effect parameters, and $\epsilon_{ij}$ represents sampling error.

For arm-level estimates, a common working assumption is that the vector of estimates from trial $j$ is approximately multivariate normal with the model mean and covariance matrix $V_j$, and that estimates from separate trials are independent. The covariance matrix should reflect the precision of the arm estimates and any within-trial correlation—for example, correlation caused by a shared comparator. If the covariance is diagonal, the arm estimates are treated as independent. A weighted fit then gives more influence to more precise estimates.

The model also assumes that the trials measure a sufficiently comparable estimand: the endpoint, its scale and direction, and the assessment time should align, or differences must be handled in the analysis plan. Doses must have a common meaning and units. A trial-specific intercept can represent a shift in outcome level, but it cannot correct for different dose-response shapes, incompatible endpoint scales, or treatment-effect modification by trial characteristics. A causal interpretation of the within-trial dose contrast also relies on appropriate trial conduct, including valid randomization and analysis of the assigned dose as defined by the estimand.

In practice, fit the model by maximizing the joint likelihood or, for normally distributed estimates with known covariance matrices, minimizing the weighted residual sum of squares:

Q(\alpha_1,\ldots,\alpha_J,\theta) = \sum_{j=1}^{J} \left(\mathbf{y}_j-\boldsymbol{\mu}_j\right)^\mathsf{T} V_j^{-1} \left(\mathbf{y}_j-\boldsymbol{\mu}_j\right), \qquad \boldsymbol{\mu}_j=\alpha_j\mathbf{1}+\mathbf{f}_j(\theta).

Here, $\mathbf{y}_j$ contains the arm estimates for trial $j$; $\mathbf{f}_j(\theta)$ contains the dose-response contributions at those arms. The form of $f$ and the parameters allowed to vary across trials should be chosen before interpreting a pooled estimate. If residual variance must be estimated, it must be included in the likelihood or estimated using an appropriate method rather than assumed known.

How Baseline Adjustment Enables Pooling

The role of trial-specific baselines is easiest to see for a linear treatment effect, $f(d;\theta)=\beta d$. With independent arm estimates having equal known variance within each trial, least squares estimates each $\alpha_j$ as the mean of $y_{ij}-\beta d_{ij}$ in that trial. Substituting these fitted intercepts into the objective is equivalent to centering dose and outcome separately within each trial. The resulting common-slope estimate is

\widehat{\beta} = \frac{ \sum_j\sum_i(d_{ij}-\bar d_j)(y_{ij}-\bar y_j) }{ \sum_j\sum_i(d_{ij}-\bar d_j)^2 }.

The centering removes each trial’s level difference: adding a constant to every outcome in one trial changes its intercept, but not its contribution to the slope numerator. This is why a joint model can use information from trials with different baselines without forcing those baselines to be equal. With unequal precision or correlated arm estimates, the corresponding calculation is weighted using the covariance matrices; ordinary unweighted centering is then not generally the correct fit.

Worked Example: A Common Linear Slope With Different Baselines

Suppose two hypothetical Phase 2 trials each report mean outcome at doses 0, 50, and 100 mg. Higher outcomes are better. The arm means are 10, 14, and 18 units in Trial A, and 15, 19, and 23 units in Trial B. Assume each arm mean has known standard error 1 unit, estimates are independent, and a common linear dose effect is appropriate over this dose range.

Within each trial, the centered doses are $-50,0,50$ mg. The centered outcomes are $-4,0,4$ units in both trials. Thus,

\widehat{\beta} = \frac{2\{(-50)(-4)+(50)(4)\}} {2\{(-50)^2+50^2\}} = \frac{800}{10000} = 0.0800\ \text{units/mg}.

The trial-specific intercepts are the fitted means at 0 mg: $\widehat{\alpha}_A=14-(0.0800)(50)=10$ units and $\widehat{\alpha}_B=19-(0.0800)(50)=15$ units. The trials therefore have different baseline levels but the same fitted increase of 0.0800 outcome units per mg. Under the stated known-error model, the slope standard error is $1/\sqrt{10000}=0.0100$ units/mg. A normal-approximation 95% confidence interval is $0.0800\pm1.96(0.0100)$, or 0.0604 to 0.0996 units/mg. Under repeated sampling and the model assumptions, this procedure would cover the common slope 95% of the time; the interval does not assign a 95% probability to the fixed slope.

Pooling a Nonlinear Dose-Response Curve

For a saturating response, one possible model is the Emax form $f(d;\theta)=E_{\max}d/(ED_{50}+d)$, where $E_{\max}$ is the maximum modeled change from zero dose and $ED_{50}$ is the dose producing half that change. A joint model can share both parameters while allowing each trial a separate $\alpha_j$. Nonlinear likelihood fitting generally requires numerical optimization. The following calculation illustrates the pooling logic with $ED_{50}$ fixed in advance; it does not account for uncertainty in that fixed value.

Worked Example: Combining Trials With Different Dose Grids

Consider hypothetical arm means with independent standard errors of 1 outcome unit. Trial A has means 10, 14.5, and 16 units at 0, 50, and 100 mg. Trial B has means 14, 17, and 20 units at 0, 25, and 100 mg. Fit separate baselines and a shared Emax parameter, taking $ED_{50}=50$ mg as fixed. Define $q(d)=d/(50+d)$, so the model is $y_{ij}=\alpha_j+E_{\max}q(d_{ij})+\epsilon_{ij}$.

The $q$ values are 0, 0.5, and 0.6667 in Trial A, and 0, 0.3333, and 0.6667 in Trial B. With equal arm-mean variances, profiling out each trial intercept is equivalent to centering $q$ and $y$ within each trial. The shared estimate is the centered least-squares slope:

\widehat E_{\max} = \frac{\sum_j\sum_i(q_{ij}-\bar q_j)(y_{ij}-\bar y_j)} {\sum_j\sum_i(q_{ij}-\bar q_j)^2}.

For Trial A, $\bar q_A=7/18$, and its contribution to the denominator is $13/54$. For Trial B, $\bar q_B=1/3$, and its denominator contribution is $2/9$. The total denominator is $13/54+2/9=25/54$. The outcomes in both trials lie exactly on the specified curve with $E_{\max}=9$, so the numerator is $9(25/54)=25/6$. Therefore,

\widehat E_{\max} = \frac{25/6}{25/54} = 9.00\ \text{outcome units}.

The fitted baselines are 10 units for Trial A and 14 units for Trial B. Under the assumed independent, known unit variances, the standard error of $\widehat E_{\max}$ is $1/\sqrt{25/54}=1.47$ outcome units. A normal-approximation 95% interval is $9.00\pm1.96(1.47)$, or approximately 6.12 to 11.88 outcome units. This uncertainty is conditional on the chosen Emax form and fixed $ED_{50}$; estimating $ED_{50}$ jointly would add uncertainty and could produce a different interval. The different dose grids pose no problem in themselves because each observation contributes its curve value at its own dose.

When Is a Shared Treatment Effect Defensible?

Sharing parameters is a substantive assumption, not a consequence of putting trials in one data set. It is more credible when the trials address compatible populations and estimands, use comparable endpoint definitions and assessment times, and have plausible reasons to share the same dose-response relationship. The model should also be appropriate over the combined dose range. A single shared slope, for example, imposes a common change per dose unit; a shared Emax model imposes common curve parameters, not merely similar observed means at one dose.

A direct way to examine this assumption is to compare the shared-effect model with a model that allows trial-specific treatment effects, such as trial-specific slopes. The comparison asks whether the data show evidence against the imposed equality; it does not establish that the effects are identical. This check is informative only if the trials contain enough dose-response information to estimate their effects. A non-significant comparison may reflect imprecision rather than close agreement.

Worked Example: A Non-Significant Difference Does Not Prove Equality

Suppose hypothetical trials each have arm means at 0, 50, and 100 mg, with known standard error 1 unit per arm mean. Trial A has means 10, 14, and 18; Trial B has means 15, 18, and 21. Under separate linear fits with an intercept per trial, the estimated slopes are 0.0800 and 0.0600 units/mg. Each slope has standard error $1/\sqrt{5000}=0.01414$ units/mg, because the within-trial centered dose sum of squares is 5000 mg$^2$.

The estimated difference is $0.0800-0.0600=0.0200$ units/mg, with standard error $\sqrt{0.01414^2+0.01414^2}=0.0200$ units/mg. The test statistic for equal slopes is $z=0.0200/0.0200=1.00$, giving a two-sided $p$-value of about 0.317. At a 0.05 significance level, this comparison does not reject equal slopes. But the result does not demonstrate equality: the data remain compatible with a range of slope differences, and the trial-specific estimates are imprecise. A shared slope would need scientific justification and suitable precision, not just a $p$-value above 0.05.

Diagnostics, Alternatives, and Limitations

Inspect fitted values and residuals by dose and trial, not only across the pooled data. A systematic within-trial residual pattern can indicate that the selected curve does not describe the dose response. A trial-specific sequence of residuals that consistently trends with dose can indicate that a common treatment effect is inadequate. Also check whether a trial has doses that overlap the others or whether it contributes information only in a region where the curve form is influential. These checks help reveal when a pooled estimate depends heavily on assumptions rather than shared direct evidence.

The main alternatives clarify the trade-off. Separate trial fits avoid imposing common treatment-effect parameters, but may be imprecise when each trial has sparse dose-ranging data and do not directly provide one common estimate. A model with trial-by-dose interactions estimates differences in treatment effects explicitly, at the cost of additional parameters and often wider uncertainty. The shared model is more efficient when its equality assumptions are reasonable; it can be misleading when they are not. More elaborate models that allow effects to vary across trials are another option, but their assumptions and interpretation must be stated rather than treated as an automatic compromise.

Pooling also cannot repair poor alignment of endpoints, doses, or estimands, nor can it create information about doses absent from all trials. Shared-parameter uncertainty may be understated if covariance is ignored, if within-trial estimates are treated as independent when they are not, or if estimated variances are treated as known. Finally, a good numerical fit does not by itself establish that the model is clinically useful. The pooled curve should be interpreted in the context of the dose-selection objective and the evidence supporting its assumptions.

Common mistake: “The trial means differ, so their treatment effects must differ.” Different baselines can produce different means at every dose. Estimate or adjust for trial-specific baselines before comparing dose effects.
Common mistake: “The interaction test was not significant, so the effects are the same.” A non-significant result is not proof of equivalence. Consider precision, scientific comparability, and whether the observed differences could matter for the decision.
Common mistake: “Pooling more arms automatically improves the answer.” More data help only under a suitable model and correctly represented uncertainty; a misspecified shared curve can give a precise but misleading estimate.

Key takeaway: A pooled dose-response analysis can separate study-specific outcome levels from shared treatment-effect parameters. Its credibility depends on compatible estimands, an adequate common curve, correctly represented sampling uncertainty, and a defensible scientific reason to share the effects.

Check Your Understanding

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

In a shared linear dose-response model, what does a trial-specific intercept allow to differ, and what does it not allow to differ?
Two trials use the same numerical dose units but assess different endpoints on different scales. Is a common treatment-effect parameter automatically interpretable? Explain.
For two trials with estimated slopes of 0.05 and 0.07 units/mg and standard errors of 0.01 units/mg each, calculate the standard error of their difference and the z statistic for equal slopes.
Why can a pooled nonlinear model use trials with different dose grids?

Solutions. (1) The intercept allows the zero-dose expected outcome to differ by trial; it does not allow the dose-response slope or other shared treatment-effect parameters to differ. (2) No. A common numerical parameter may not represent the same estimand across incompatible endpoint scales; endpoint alignment or a justified transformation is needed. (3) The difference is $0.07-0.05=0.02$ units/mg. Its standard error is $\sqrt{0.01^2+0.01^2}=0.01414$ units/mg, so $z=0.02/0.01414=1.41$. This alone does not demonstrate that the slopes are equal. (4) Each arm contributes the modeled response at its own dose, so a joint curve can use different dose locations as long as the dose definition and curve assumptions are appropriate.