Designing Dose-Combination Studies for Interaction Estimation

Designing for an Interaction, Not Just a Combination Effect

A combination that produces a larger response than either drug alone is not necessarily interacting: two independent drug effects can add to produce that result. A useful study must therefore distinguish the observed combination response from a defined reference for what the response would be without interaction. The reference depends on the response scale, and the chosen doses, sampling times, and allocation determine how well the study can estimate the departure from it.

For continuous responses, this tutorial focuses on additivity on the response scale. That is one possible definition of no interaction, not a scale-free property of the drugs. The design principles apply to a planned response-surface analysis: include enough monotherapy and combination information to estimate individual dose effects as well as departures from the chosen reference.

Definition: Let $Y_{ab}$ be the mean response at doses $a$ and $b$, $Y_{a0}$ and $Y_{0b}$ the corresponding monotherapy means, and $Y_{00}$ the mean with neither drug. The response-scale interaction contrast at that dose pair is $$ \Delta(a,b)=Y_{ab}-Y_{a0}-Y_{0b}+Y_{00}. $$ Under response-scale additivity, $\Delta(a,b)=0$. A nonzero contrast is a departure from that reference at the specified doses and response time. Its sign describes the direction of departure only after the response’s clinical direction is stated.

The contrast clarifies why monotherapy arms matter. A combination arm alone cannot show whether its response exceeds the additive expectation, because that expectation requires the two single-drug responses and the untreated or placebo response. In a dose-response model, a series of such contrasts across dose pairs can describe how interaction varies over the surface.

Choose Doses That Inform the Surface

Dose selection should be based on the response range expected in the intended population, informed where possible by prior pharmacokinetic and pharmacodynamic knowledge. When concentrations are the model predictors, the corresponding concentrations—not merely nominal doses—should cover informative parts of each drug’s exposure range. A design using only very low exposures may have little measurable response; one using only saturating exposures may reveal little about the shape of either drug’s individual effect.

One useful planning representation is a saturating response surface:

E(Y\mid C_A,C_B)=E_0+E_{\max,A}f_A+E_{\max,B}f_B+\gamma f_Af_B, \qquad f_j=\frac{C_j}{EC_{50,j}+C_j}.

Here $C_A$ and $C_B$ are concentrations, $E_0$ is the response at zero concentrations, and $E_{\max,j}$ and $EC_{50,j}$ describe each drug’s single-agent contribution in this illustrative model. The product term allows the combination to depart from response-scale additivity; $\gamma=0$ is the additive reference in this parameterization. This model is a planning aid, not a guarantee that the true surface has this shape.

For each drug, include zero exposure and multiple nonzero levels. Levels around the rising part of the response curve are often especially informative: there, response changes as exposure changes. Include combinations that pair informative levels of both drugs, rather than concentrating all combinations at one corner of the surface. If a clinically important region is known, ensure it is represented even if other regions are included to estimate the broader shape.

Worked Example: Selecting Levels to Reveal an Interaction Term

Hypothetical planning example. Suppose prior work suggests $EC_{50,A}=EC_{50,B}=10$ concentration units. The planned levels for each drug are 0, 5, and 20 concentration units. Calculate the normalized effects $f_j$, then the interaction basis $f_Af_B$ for two combination cells.

At 5 units, $f=5/(10+5)=0.3333$. At 20 units, $f=20/(10+20)=0.6667$. The dose matrix therefore includes zero, a level below the estimated $EC_{50}$, and a level above it for each drug.

For the 5-by-5 combination, $f_Af_B=0.3333\times0.3333=0.1111$. For the 20-by-20 combination, $f_Af_B=0.6667\times0.6667=0.4444$. In the stated planning model, the interaction contributions to mean response are therefore $0.1111\gamma$ and $0.4444\gamma$, respectively.

The higher combination has a larger interaction basis in this model, but it is not automatically the best design point. If responses there are near a clinical ceiling, measurement may be less informative or the model may be less credible. Including both levels, the monotherapy arms, and intermediate combinations helps show whether the proposed surface describes the data rather than relying on one high-dose contrast.

Balance Monotherapy, Combination, and Control Information

A complete factorial grid is a direct design for a response surface: each selected dose of A is paired with each selected dose of B, including zero-dose levels. It supports comparisons across both dose dimensions and makes the single-agent responses observable within the same study. A full grid may be impractical when there are many dose levels, so a smaller design can prioritize clinically important combinations and include enough monotherapy points to estimate each individual response curve.

The trade-off is between coverage and precision. A single combination dose can estimate an interaction contrast at that dose pair, but cannot establish how the contrast changes elsewhere. Conversely, many sparsely allocated cells can leave each cell mean imprecise. Prespecify which dose pairs are primary for interaction estimation and allocate adequate observations to their corresponding placebo, monotherapy, and combination cells.

Randomize participants to the planned regimens so that differences in baseline characteristics do not systematically track dose assignment. In a parallel-group design with one independent continuous endpoint per participant, a simple precision calculation can guide allocation to a primary four-cell contrast. This calculation assumes independent observations, a common within-cell standard deviation, and the same endpoint definition and sampling time in all four arms. If variability differs by arm, or if repeated measurements are used, the calculation must reflect that design instead.

Worked Example: Allocating Participants for a Primary Contrast

Hypothetical planning example. Consider four randomized arms: placebo, A alone at a selected dose, B alone at a selected dose, and the A-plus-B combination. Suppose the response standard deviation is expected to be 8 response units in every arm. The target interaction contrast is 10 response units. Estimate the equal sample size per arm for approximately 80% power at a two-sided 5% significance level, using a normal approximation.

The contrast estimator is $\hat{\Delta}=\bar{Y}_{AB}-\bar{Y}_{A0}-\bar{Y}_{0B}+\bar{Y}_{00}$. With $n$ independent participants per arm and common standard deviation $\sigma=8$, its variance is the sum of the four mean variances:

\operatorname{Var}(\hat{\Delta})= \frac{\sigma^2}{n}+\frac{\sigma^2}{n}+ \frac{\sigma^2}{n}+\frac{\sigma^2}{n} =\frac{4\sigma^2}{n}, \qquad SE(\hat{\Delta})=\frac{2\sigma}{\sqrt{n}}.

For the stated approximation, the signal-to-standard-error target is $z_{0.975}+z_{0.80}\approx1.96+0.84=2.80$. Thus $10/(16/\sqrt{n})\geq2.80$, or $n\geq(2.80\times16/10)^2=20.07$. Rounding up gives 21 participants per arm, or 84 in total.

This is a planning approximation, not a guarantee of the achieved power. It relies on the assumed standard deviation, equal allocation, independent endpoint errors, and a normal approximation. Allowances for attrition, multiple primary contrasts, unequal variances, or a need to estimate several surface parameters may increase the required sample size.

Choose Sampling Times to Match the Interaction Question

A dose-combination study must specify when response is measured. Drug concentrations and pharmacodynamic effects can change over time; a contrast at one sampling time need not equal a contrast at another. If the target is an interaction at a particular clinical time, measure all four arms at comparable times and define that estimand in advance. If the target concerns the response profile over time, collect measurements at times that describe the relevant onset, peak, and decline rather than treating one convenient observation as the whole profile.

When feasible, measure concentrations close enough to response observations to characterize the exposures that drive the response. Sampling only at a time when both drugs are nearly absent can make a concentration-dependent interaction difficult to detect; sampling only near a peak may miss a delayed effect or a clinically important later interaction. For delayed or persistent pharmacodynamic effects, include response measurements after concentrations change, as well as concentration measurements that can help interpret the delay. The timing should be based on expected pharmacokinetics and pharmacodynamics, not chosen solely for convenience.

Worked Example: Why a Single Sampling Time Can Mislead

Hypothetical illustration. Suppose the same four arms are measured at 2 and 8 hours. At 2 hours, their mean responses are $Y_{00}=10$, $Y_{A0}=22$, $Y_{0B}=18$, and $Y_{AB}=40$ response units. At 8 hours, the respective means are 10, 15, 14, and 24 response units.

At 2 hours, the contrast is $40-22-18+10=10$ response units. At 8 hours, it is $24-15-14+10=5$ response units. These calculations indicate different departures from response-scale additivity at the two observed times; neither value by itself establishes the shape of the interaction over the full time course.

If only the 8-hour sample had been collected, the 2-hour contrast of 10 units would not have been observed. If only the 2-hour sample had been collected, the later contrast would be unknown. A design intended to estimate time-varying interaction therefore needs sampling times that span the relevant response period, with the same time definitions across treatment arms.

Design Diagnostics, Alternatives, and Limitations

Several design features can warn that an apparent interaction estimate will be difficult to interpret. Before data collection, check whether every primary combination has the corresponding monotherapy and control information, whether selected levels span the expected response range, and whether planned sampling overlaps with meaningful exposure and response. After data collection, inspect observed concentrations and responses by dose and time. Sparse cells, concentrations outside the planned range, or responses clustered at a floor or ceiling can undermine estimation even when the nominal dose matrix looks complete.

The main design alternatives serve different purposes:

Design approach	Strength	Limitation
Full factorial dose grid	Broad coverage of both dose dimensions, with monotherapies included	Can require many arms and participants
Focused set of combinations plus monotherapy levels	Can concentrate resources on clinically important dose pairs	Provides less information about unselected regions of the surface
One fixed combination versus the two monotherapies	Directly estimates a contrast at one selected dose pair	Cannot describe how interaction varies over dose

A full factorial design is not always necessary, and a smaller design is not automatically inadequate. The choice depends on whether the scientific goal is a contrast at one dose pair or a response surface across a region. Designs that select combinations without the needed single-agent comparisons cannot estimate the stated additive reference directly. Similarly, varying only one drug while holding the other at one dose does not fully characterize how the response changes across both dimensions.

Model-based design calculations depend on assumptions about the response surface, residual variability, and exposure distribution. If the assumed $EC_{50}$ values or response ranges are inaccurate, nominally well-spaced doses may produce little useful variation in exposure or response. Where feasible, use prior evidence or a staged design to update dose choices, while preserving the planned comparisons needed for interpretation. A design can estimate a departure from additivity on its chosen scale; it does not make that scale uniquely correct or prove a biological mechanism.

Common Mistakes

Calling any larger combination response synergy. A combination can exceed either monotherapy under additivity. Compare it with the explicitly defined additive expectation, using the chosen response scale.
Omitting monotherapy arms. Without the single-agent means at the relevant doses, the response-scale interaction contrast cannot be calculated directly.
Using only very low or saturating levels. Such levels can provide little information about the changing response surface. Select levels expected to cover informative portions of each drug’s effect range.
Measuring only at an arbitrary time. The estimated contrast applies to the measured time and may not describe the response at other times. Align sampling with the scientific question and expected time course.

Key takeaway: Design around the estimand. State the additivity reference and response time, include the dose-specific monotherapy and control information needed to calculate it, and choose dose levels and allocation that provide useful information across the region of interest.

Check Your Understanding

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

At a selected dose pair, the hypothetical means are $Y_{00}=5$, $Y_{A0}=12$, $Y_{0B}=9$, and $Y_{AB}=20$. What is the response-scale interaction contrast?
Why is a combination arm and placebo arm alone insufficient to estimate the response-scale interaction contrast?
For the four-arm contrast, suppose the common standard deviation is 6 response units and there are 36 participants per arm. What is the standard error under the independent, equal-variance assumptions?
Give one reason to include both an intermediate and a high dose level for each drug rather than testing only high doses.

Solutions. (1) $\Delta=20-12-9+5=4$ response units, a positive departure from response-scale additivity. (2) The contrast also requires the A-alone and B-alone means at the selected doses to form the additive expectation. (3) $SE=2\sigma/\sqrt{n}=2(6)/\sqrt{36}=2$ response units. (4) Intermediate levels can characterize the rising response region, while high levels show behavior nearer the upper range; high-dose-only observations may be saturated and provide limited information about dose-response shape.