Combining Dose-Response Evidence Across Trials

Why Combine Dose-Response Evidence?

A dose-response pattern observed in one trial may be imprecise, limited to a narrow dose range, or specific to the population and regimen studied. Other trials can add information, but their estimates are not automatically comparable. One study may enroll patients with more severe disease, use a different dosing interval, or measure improvement on a different endpoint. Pooling numbers before addressing those differences can produce a precise estimate of the wrong quantity.

The central task is therefore to define what is being combined. For dose selection, that usually means estimating a treatment effect at a clinically relevant dose or exposure, on a clearly defined endpoint and in a target population. Pooling is useful only to the extent that trial-specific estimates meaningfully inform that shared target.

Definition: A common dose-response estimand is the treatment effect or response quantity that each trial is intended to estimate at a specified dose, exposure, or regimen, for a defined population, endpoint, and time point. A harmonized trial estimate is a trial-specific estimate expressed on a scale and in a direction that represent this same estimand, together with its standard error.

A positive effect will be used here to mean greater benefit. The effect scale may be a difference in the original clinical units when endpoints are the same, or a standardized effect when different continuous endpoints measure a sufficiently similar construct. Standardization makes units comparable; it does not make unlike clinical concepts equivalent.

Establish Comparability Before Pooling

Start by checking the clinical meaning of each estimate, not by choosing a meta-analytic formula. The dose must refer to comparable treatment exposure or a defensible mapping between regimens. The endpoint definitions, direction, assessment time, comparator, intercurrent-event handling, and analysis population must also be examined. A nominal dose in milligrams is not necessarily comparable when bioavailability, dosing frequency, or population pharmacokinetics differ.

Population differences require particular care. If the dose effect varies with disease severity, body size, organ function, or another covariate, trial averages may differ simply because the enrolled populations differ. A pooled average can still be meaningful for a specified target population, but the target and any required standardization must be stated. The earlier material on covariate-stratified dosing distinguishes population-wide dosing from subgroup-specific dosing; cross-trial pooling does not remove the need to make that distinction.

Different endpoint definitions may sometimes be aligned by coding all effects in the beneficial direction and using a standardized mean difference. This is defensible only when the endpoints measure sufficiently similar constructs and the standard deviations are interpretable. If one endpoint measures symptom relief and another measures functional capacity, a shared standardized scale alone does not establish a shared estimand.

Worked Example: Harmonizing Two Continuous Endpoints

Suppose two hypothetical randomized trials compare a dose with placebo. Trial A measures symptom reduction, for which larger reductions are beneficial. Mean reductions are 8 points with treatment and 3 points with placebo; the pooled within-trial standard deviation is 10 points. Trial B measures improvement on a different scale, where higher values are better. Mean improvements are 4 points with treatment and 1 point with placebo; the pooled standard deviation is 8 points. Assume, for this illustration, that both endpoints measure the same underlying clinical construct at the same follow-up time.

In Trial A the treatment-placebo benefit is \(8-3=5\) symptom points. Its standardized estimate is \(5/10=0.500\). In Trial B the benefit is \(4-1=3\) points, giving \(3/8=0.375\). Both estimates are positive because each has been oriented so that positive values indicate benefit.

Suppose the standard errors of the standardized estimates are 0.200 and 0.250. Under a common-effect assumption, their inverse-variance weights are \(1/0.200^2=25\) and \(1/0.250^2=16\). The pooled estimate is

\hat{\theta} =\frac{25(0.500)+16(0.375)}{25+16} =\frac{18.5}{41} =0.4512.

Its standard error is \(1/\sqrt{41}=0.1562\). This calculation is meaningful only under the stated comparability assumption: the endpoints must represent a sufficiently similar construct, and the standardized effects must target comparable populations, doses, and times. If those conditions fail, the arithmetic remains possible but the pooled result may not answer a coherent clinical question.

Fixed-Effect and Random-Effects Summaries

Let \(\hat{\theta}_i\) be the harmonized effect estimate from trial \(i\), with standard error \(SE_i\), for \(i=1,\ldots,k\). A fixed-effect model assumes that all trials estimate one common true effect \(\theta\), and that observed differences arise from sampling error. Define \(w_i=1/SE_i^2\). The inverse-variance pooled estimate and its standard error are

\hat{\theta}_{FE} =\frac{\sum_{i=1}^{k}w_i\hat{\theta}_i}{\sum_{i=1}^{k}w_i}, \qquad SE(\hat{\theta}_{FE})=\frac{1}{\sqrt{\sum_{i=1}^{k}w_i}}.

A random-effects model instead allows the true effects to vary across trials. A simple formulation is \(\hat{\theta}_i\mid\theta_i\sim N(\theta_i,SE_i^2)\) and \(\theta_i\sim N(\mu,\tau^2)\), where \(\mu\) is the mean of the distribution of trial effects and \(\tau^2\) is the between-trial variance. Its weights are \(w_i^*=1/(SE_i^2+\tau^2)\), giving \(\hat{\mu}=\sum w_i^*\hat{\theta}_i/\sum w_i^*\) and \(SE(\hat{\mu})=1/\sqrt{\sum w_i^*}\), when \(\tau^2\) is treated as known.

In practice, \(\tau^2\) is estimated. One common moment estimator is the DerSimonian–Laird estimator. First calculate the fixed-effect heterogeneity statistic \(Q=\sum w_i(\hat{\theta}_i-\hat{\theta}_{FE})^2\), with \(df=k-1\). Then define \(C=\sum w_i-\sum w_i^2/\sum w_i\), and estimate

\hat{\tau}^2_{DL} =\max\left(0,\frac{Q-df}{C}\right).

The estimator compares observed dispersion around the common-effect estimate with the dispersion expected from sampling error. Values of \(Q\) above \(df\) indicate excess dispersion by this measure. The truncation at zero prevents a negative variance estimate. The random-effects model does not explain why trial effects differ; it summarizes their distribution under its assumptions.

Worked Example: Quantifying Heterogeneity

Consider three hypothetical trials estimating benefit at a comparable target exposure and on a harmonized standardized endpoint. Their estimates are 0.40, 0.80, and 0.20, each with standard error 0.20. Assume independent trial estimates and, for the moment estimator, approximately normal sampling errors.

Each fixed-effect weight is \(1/0.20^2=25\). Thus \(\hat{\theta}_{FE}=(0.40+0.80+0.20)/3=0.4667\). The heterogeneity statistic is

Q=25\left[(0.40-0.4667)^2+(0.80-0.4667)^2+(0.20-0.4667)^2\right] =4.667.

Here \(df=3-1=2\). For a chi-square distribution with two degrees of freedom, the upper-tail probability is approximately \(P(\chi^2_2\geq4.667)=0.0970\). This is not strong evidence against the common-effect model at a 0.05 threshold, but it does not establish that trial effects are identical.

For the moment estimator, \(\sum w_i=75\), \(\sum w_i^2=3(25^2)=1875\), and \(C=75-1875/75=50\). Therefore \(\hat{\tau}^2_{DL}=(4.667-2)/50=0.05333\), in squared standardized-effect units. Each random-effects weight is \(1/(0.04000+0.05333)=10.71\). Because the standard errors are equal, the random-effects mean remains \(0.4667\), and its standard error is \(1/\sqrt{3(10.71)}=0.1764\). A normal-approximation 95% confidence interval is \(0.4667\pm1.96(0.1764)=(0.121,0.812)\).

An approximate prediction interval for the true effect in a new, comparable trial is \(0.4667\pm1.96\sqrt{0.05333+0.1764^2}\), or approximately \((-0.103,1.036)\). It is wider because it includes between-trial variation as well as uncertainty in the pooled mean; it is only approximate and does not account fully for uncertainty in estimating \(\tau^2\).

What Heterogeneity Means for a Dose Recommendation

A pooled estimate can conceal clinically important variation. If every trial supports benefit at a dose, moderate variation may have little effect on the decision. If some trials suggest little benefit or harm while others suggest substantial benefit, the mean may describe no individual trial well. In dose selection, the relevant question is not only whether \(Q\) is statistically notable, but whether plausible trial-specific effects would change the balance of efficacy, safety, and burden that supports the recommendation.

Consider the following brief derivation of why the random-effects standard error differs from the fixed-effect standard error. Under the random-effects model, the sampling variance of \(\hat{\theta}_i\) around \(\mu\) is \(SE_i^2+\tau^2\). Independent estimates therefore have inverse-variance weights \(w_i^*=1/(SE_i^2+\tau^2)\). The variance of their weighted mean is

\operatorname{Var}\left(\frac{\sum_i w_i^*\hat{\theta}_i}{\sum_i w_i^*}\right) =\frac{\sum_i (w_i^*)^2(SE_i^2+\tau^2)}{(\sum_i w_i^*)^2} =\frac{1}{\sum_i w_i^*}.

The second equality follows because \(w_i^*(SE_i^2+\tau^2)=1\), so the numerator becomes \(\sum_i w_i^*\). As \(\tau^2\) increases, the most precise studies lose some of their relative dominance, and the pooled mean becomes less precise.

Worked Example: Heterogeneity and the Preferred Dose

Suppose two hypothetical trials estimate standardized benefit at two clinically comparable exposure targets. At the lower target, estimates are 0.30 and 0.25, each with standard error 0.15. At the higher target, estimates are 0.55 and 0.40, each with standard error 0.18. Assume the endpoint, direction, follow-up, and target population have been harmonized. For this illustration, use fixed-effect summaries at each target and assume no detectable between-trial variance, while recognizing that only two studies provide little information about heterogeneity.

At the lower target, each weight is \(1/0.15^2=44.44\); the pooled estimate is \((0.30+0.25)/2=0.275\), and its standard error is \(1/\sqrt{88.89}=0.1061\). The approximate 95% confidence interval is \(0.275\pm1.96(0.1061)=(0.0671,0.483)\).

At the higher target, each weight is \(1/0.18^2=30.86\); the pooled estimate is \((0.55+0.40)/2=0.475\), and its standard error is \(1/\sqrt{61.73}=0.1273\). Its approximate 95% confidence interval is \(0.475\pm1.96(0.1273)=(0.225,0.725)\). The higher target has a larger estimated benefit, but these intervals alone do not test the difference between doses; the two estimates may be correlated within each trial.

Suppose a prespecified safety criterion allows a serious adverse-event rate no greater than 5%, and the pooled observed rates are 3% at the lower target and 9% at the higher target. The higher target does not meet that criterion. Under these hypothetical inputs, the lower target is the defensible choice among the two, although the uncertainty in efficacy and the limited safety evidence should be communicated. Pooling has informed the comparison, but it has not overridden the decision rule or resolved uncertainty about the target population.

Diagnostics, Alternatives, and Limitations

A forest plot is a useful first diagnostic: inspect the direction and size of each trial estimate and its uncertainty, rather than viewing only the pooled diamond. Investigate influential trials, unexpected endpoint conversions, and whether the estimates overlap in dose or exposure. Compare results under fixed-effect and random-effects assumptions, and examine whether conclusions change when a clinically dissimilar study is omitted. Such checks are diagnostics, not permission to discard an inconvenient result.

The \(Q\) test has limited power when few trials are available and can be overly sensitive with many precise trials. \(I^2=\max(0,(Q-df)/Q)\times100\%\) describes the share of observed dispersion attributed to between-trial variation under its assumptions; it is not an absolute measure of clinical importance. Always interpret it alongside \(\hat{\tau}^2\), the range of effects, study context, and dose-decision consequences. With few trials, estimates of \(\tau^2\), confidence intervals, and prediction intervals can be unstable.

A two-stage approach first estimates a dose-response curve within each trial and then combines comparable parameters or predictions. A one-stage approach models participant-level data jointly, often allowing trial-specific baselines, covariate distributions, and dose-response deviations. Individual participant data can support more direct covariate standardization, but access and model complexity may limit this option. Neither approach makes noncomparable endpoints or unsupported dose extrapolations safe to combine. Combining p-values is another alternative, but it tests evidence against a null rather than estimating a shared dose-response effect; it is generally insufficient for choosing a dose.

A numerical correction is worth keeping distinct from the statistical definitions: the arithmetic \(2-(0.25^2+0.25^2)/(0.25+0.25)\) equals \(2-0.125/0.5=1.75\), not 0.25. For the DerSimonian–Laird calculation, however, use the defined \(C=\sum w_i-\sum w_i^2/\sum w_i\) with one consistent set of trial weights, as in the worked calculation above.

Do not pool unlike quantities. Reversing endpoint direction or standardizing units does not make distinct clinical constructs equivalent.
Do not treat a nonsignificant \(Q\) test as proof of homogeneity. Few trials may provide little power to detect meaningful differences.
Do not confuse a mean effect with a guaranteed effect in a new trial. Between-trial variation is relevant to transportability and is reflected more directly in a prediction interval.
Do not let a pooled efficacy estimate erase safety or regimen constraints. Dose selection is a decision across benefits, harms, population, and dosing practicality.

Key takeaway: Integrate trial evidence only after defining a shared dose-response estimand. Quantify heterogeneity, investigate its clinical sources, and ask whether the range of plausible trial-specific effects changes the dose decision—not merely whether a pooled estimate is statistically significant.

Check Your Understanding

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

Why does orienting all treatment effects so positive values mean benefit fail, by itself, to justify pooling endpoints?
Two independent trials estimate the same effect as 0.30 and 0.50, with standard errors 0.20 and 0.20. Calculate the fixed-effect pooled estimate and its standard error.
What does a nonsignificant heterogeneity \(Q\) test establish when only a few trials are available?
How does a prediction interval differ in interpretation from a confidence interval for the pooled mean?
A higher dose has the largest pooled efficacy estimate but exceeds a prespecified safety limit. What is the appropriate role of the pooled estimate in the dose decision?

Solutions. (1) Direction alignment only makes the signs consistent. The endpoints must still represent a comparable construct, time point, population, and estimand. (2) Both weights are \(1/0.20^2=25\), so the pooled estimate is \([25(0.30)+25(0.50)]/50=0.400\); its standard error is \(1/\sqrt{50}=0.1414\). (3) It means the data did not provide sufficient evidence to reject the common-effect assumption at the chosen threshold; low power means meaningful heterogeneity may remain. (4) A confidence interval quantifies uncertainty about the mean effect across the modeled trial distribution, whereas a prediction interval describes a range for the true effect in a new comparable trial and includes between-trial variation. (5) The efficacy estimate informs the benefit side of the decision, but it does not negate the safety constraint; if the limit is prespecified and the higher dose exceeds it, that dose does not satisfy the decision rule.