What a Confirmatory Dose-Response Test Establishes
A confirmatory trial can provide evidence that outcomes change systematically across randomized doses, or that a particular dose differs from control. These are related but distinct claims. A trend test compresses the dose groups into a prespecified ordered pattern; a dose-specific contrast directly compares selected groups. Neither test, by itself, establishes that the observed difference is clinically important, that every dose differs from the next, or that a particular dose is optimal.
The earlier material on separating statistical significance from clinical dose relevance is important here: inferential evidence concerns compatibility with a statistical null under specified assumptions, whereas clinical relevance requires interpretation against a meaningful effect threshold, endpoint properties, safety, and the intended treatment context. This tutorial focuses on what trend tests and dose-specific contrasts establish in a confirmatory trial—and where their evidential reach stops.
Model, Assumptions, and the Trend Test
Suppose a parallel-group trial randomizes participants to \(J\) dose groups. Let \(Y_{ij}\) be the endpoint for participant \(i\) in group \(j\), and let \(x_j\) be the prespecified dose score for that group. For a simple analysis of a continuous endpoint, assume
The classical pooled-variance trend analysis additionally assumes approximately normal errors within groups, a common variance across groups, independent observations, and a score chosen before examining the outcomes. Under these conditions, fit the group-mean linear trend using the scores \(x_j\). With equal group size \(n\), let \(\bar{x}\) be the mean of the group scores, and let \(\bar{Y}_j\) be the observed mean in group \(j\). The estimated slope is
Its units are endpoint units per score unit. The pooled within-group variance estimate is \(s^2\), with \(df=N-J\), where \(N\) is the total number of participants. Under the null hypothesis of no linear trend, the test statistic is \(t=\hat{\beta}/SE\), compared with a \(t\) distribution with \(N-J\) degrees of freedom.
A trend alternative can instead be one-sided if that direction was justified and specified in advance. The test is most interpretable when the scores encode the scientific question. Equally spaced scores test a linear trend over that score scale; ordinal scores test a linear pattern over the assigned ranks, not necessarily a linear effect per milligram of dose.
Why the Standard Error Has This Form
The slope is a weighted sum of independent group means. Under the equal-size and common-variance assumptions, each \(\bar{Y}_j\) has variance \(\sigma^2/n\). Therefore, the variance of the numerator is
Dividing by the squared denominator of the slope gives
Replacing the unknown \(\sigma\) with the pooled estimate \(s\) yields the stated standard error. This derivation also shows why the score spread matters: if the assigned scores are tightly clustered, the denominator is smaller and the estimated slope is less precise.
Worked Example: Testing a Prespecified Trend Across Four Doses
Consider a hypothetical confirmatory trial with four equally sized groups: control and three increasing doses, with 100 participants per group. The prespecified scores are \(-1.5,-0.5,0.5,1.5\), so \(\bar{x}=0\) and \(\sum(x_j-\bar{x})^2=5\). Suppose the group means, in endpoint points, are 17.45, 19.15, 20.85, and 22.55, and the pooled within-group standard deviation is \(s=10\) points. These are hypothetical data; the assumptions of independent randomized groups, a common within-group variance, and an adequately normal within-group endpoint distribution are taken as reasonable for this worked calculation.
The equally spaced means increase by 1.70 points for each one-unit increase in score. The slope calculation confirms this:
The standard error is
There are \(N=400\) participants and \(J=4\) groups, so \(df=400-4=396\). The test statistic is \(t=1.70/0.4472=3.801\). The two-sided p-value is approximately 0.0001666. Thus, under the model and null hypothesis, a trend statistic at least this extreme in either direction would be uncommon. The data support a nonzero linear trend on the prespecified score scale; they do not establish that the increase is clinically important or that the relationship is linear at every dose.
A 95% confidence interval uses \(t_{0.975,396}\), approximately 1.966:
The interval gives a range of values compatible with the data and model at the 95% confidence level. It is not a 95% probability statement about this fixed slope.
Dose-Specific Contrasts Answer a Different Question
A contrast can be written \(C=\sum_j c_j\mu_j\), where the fixed coefficients \(c_j\) sum to zero. For example, a dose-versus-control comparison has coefficient 1 for that dose, \(-1\) for control, and 0 for other groups. Its estimate is the same weighted combination of observed group means. With equal group sizes and a common variance, the standard error is \(s\sqrt{\sum_j c_j^2/n}\). A test of \(H_0:C=0\) asks whether that particular mean contrast is zero; it does not test the overall shape of the curve.
Worked Example: Interpreting a Dose-versus-Control Contrast
In a second hypothetical four-group trial, each group has 80 participants. The prespecified scores are again \(-1.5,-0.5,0.5,1.5\), and the estimated trend slope is 1.70 endpoint points per score unit. Suppose the pooled standard deviation is 8 points. Randomization, independence, approximate within-group normality, and a common variance are assumed for this calculation.
The sum of squared centered scores remains 5. Therefore,
The total sample size is \(N=320\), so \(df=320-4=316\). The statistic is \(t=1.70/0.400=4.25\), with a two-sided p-value less than 0.0001. This is strong evidence against a zero linear trend under the stated model. It remains evidence about the scored trend, not evidence that every dose-control contrast is statistically significant, or that the effect size is clinically worthwhile.
Now suppose a planned contrast compares the highest dose with control and estimates a difference of 5 endpoint points. With 80 participants in each group, its standard error is \(8\sqrt{1/80+1/80}=1.265\) points, and its test statistic is \(5/1.265=3.953\). This comparison directly estimates the highest-dose difference; the trend test does not substitute for it. The contrast’s p-value and interval should be interpreted with the trial’s prespecified multiplicity strategy.
Worked Example: Several Dose-versus-Control Comparisons
Suppose a hypothetical trial has four groups with 100 participants each and pooled standard deviation 10 endpoint points. The observed mean differences from control are 2, 4, and 5 points for the low, middle, and high doses. Assume the group independence, common variance, and approximate normality conditions used for the pooled two-sample comparisons. For each comparison, the standard error is
The three unadjusted two-sided tests, each using \(df=396\), give:
- Low dose: \(t=2/1.414=1.414\), with \(p\approx0.158\).
- Middle dose: \(t=4/1.414=2.828\), with \(p\approx0.0049\).
- High dose: \(t=5/1.414=3.536\), with \(p\approx0.0004551\).
These are unadjusted p-values. If the confirmatory claim requires testing all three dose comparisons while controlling the familywise type I error rate, the prespecified multiplicity procedure matters. For example, a simple Bonferroni rule at familywise \(\alpha=0.05\) compares each p-value with \(0.05/3=0.0167\): the middle- and high-dose comparisons meet that threshold, while the low-dose comparison does not. This does not make the two significant differences clinically meaningful; it only addresses the stated statistical error-rate criterion for this family of tests.
Diagnostics, Limitations, and the Choice of Evidence
The assumptions are not merely formal. Examine group sizes, within-group variability, influential observations, and residual patterns. Strongly unequal variances can make a pooled standard error unreliable; severe skewness or outliers can undermine small-sample t-based inference. Missing outcomes, protocol deviations, or departures from the planned analysis can also affect the estimand and interpretation. In a confirmatory setting, the primary endpoint, analysis population, score assignment, contrast family, and multiplicity procedure should be prespecified. A post hoc trend score chosen because it produces a small p-value does not retain the confirmatory interpretation of a prespecified test.
The trend test and contrasts are complementary, not interchangeable. A trend test can gain precision by using the order of all groups, but it may conceal departures from the assumed pattern. A contrast is directly relevant to a specific dose decision, but testing many contrasts creates a multiplicity issue and does not summarize the entire response pattern. A categorical group analysis avoids imposing a linear score relationship but may use more parameters and answer a less compact question. The choice should follow the intended claim, not whichever analysis yields the smallest p-value.
| Analysis | Question it addresses | What it does not establish by itself |
|---|---|---|
| Prespecified trend test | Is there a linear change in mean outcome over the assigned score scale? | That the curve is linear at every dose, or that a dose difference is clinically important |
| Dose-versus-control contrast | Does this specified dose differ from control on the endpoint? | That other doses differ, or that this dose is optimal |
| Separate unadjusted contrasts | What are the nominal results for each specified comparison? | Control of the overall false-positive rate across the comparison family |
| Categorical group comparison | Do any group means differ without imposing a linear score trend? | A compact dose slope or a particular clinically meaningful comparison |
A statistically significant trend can coexist with a small, clinically unimportant change, especially in a large trial. Conversely, a nonsignificant trend does not prove the absence of dose response: the study may have limited precision, the trend may be nonlinear, or the chosen score may poorly represent the response. Use confidence intervals to assess the range of effects compatible with the data and compare them with a clinically justified threshold in the endpoint’s units. Clinical relevance cannot be inferred from a p-value alone.
Common Pitfall
Check Your Understanding
Answer each question using the distinctions and calculations developed above. Brief solutions follow the questions.
- A trial’s prespecified linear trend test is significant, but the low-dose versus control contrast is not. What is established, and what is not?
- Why should dose scores be chosen before examining the endpoint results?
- For three planned dose-control comparisons and familywise \(\alpha=0.05\), what per-comparison threshold does the simple Bonferroni rule use?
- Can a nonsignificant trend test establish that the response is flat? Explain.
- What additional interpretation is needed before calling a statistically significant 2-point difference clinically relevant?
Solutions. (1) The test supports a nonzero linear pattern on its prespecified score scale under its assumptions; it does not establish that the low dose differs from control or that every dose is effective. (2) Choosing scores after seeing outcomes can make the test data-dependent and invalidate its planned confirmatory interpretation. (3) The threshold is \(0.05/3=0.0167\) per comparison. (4) No. A nonsignificant result is not proof of no trend; precision, power, nonlinearity, and score choice matter. (5) Interpret the difference in endpoint units against a clinically justified threshold and relevant uncertainty interval, alongside the trial context; the p-value alone does not establish clinical importance.