Why the Comparison Must Be Specified
A statement such as “the perpetrator increased victim-drug exposure” or “the combination changed response” is incomplete until the comparison is defined. The result depends on the victim-drug dose, whether the perpetrator is present, the population being described, and the exposure or response summary being measured. For response, it also matters whether the comparison holds doses fixed or instead compares conditions at a matched victim-drug concentration.
An estimand is the precise target quantity an analysis is intended to estimate. For a drug interaction, it specifies the treatment conditions being compared, the target population, the variable or endpoint, and the summary measure. Defining it before fitting a model helps prevent a numerical contrast from being given a broader interpretation than it supports.
The exposure summary might be AUC over a specified interval, \(C_{\max}\), or trough concentration. The response endpoint might be a value at a specified time, a time-averaged response, or an area under the response-time curve. These are different estimands; they should not be treated as interchangeable simply because they come from the same study.
Fixed-Dose Total Interaction Effects
For an exposure contrast, hold the victim-drug dose and regimen constant and compare exposure with versus without the perpetrator. For example, a population geometric mean ratio for AUC can be written as follows, where \(AUC(d,z)\) is the AUC under the stated treatment condition:
The target population and AUC interval must be specified. A ratio of 1 indicates equal geometric mean AUC under the two conditions; a ratio above 1 indicates higher geometric mean AUC with the perpetrator. This is a total exposure effect at the specified dose: it describes the overall exposure difference associated with the two treatment conditions, not the contribution of any single mechanism.
A fixed-dose response contrast likewise compares the treatment conditions at the same victim and perpetrator doses. If \(Y(d,z)\) is a defined response endpoint, the population mean difference is:
This contrast includes any response change that occurs under the combination treatment at those doses, including changes that accompany altered victim-drug concentrations. It does not, by itself, say whether the response difference arose from changed exposure, an effect of the perpetrator on response at a given victim concentration, or both. Those are separate questions.
Causal interpretation requires a well-defined treatment regimen and endpoint, valid measurement, and a design or analysis that supports exchangeability between conditions. Random assignment or a suitable crossover design can support that comparison; in a crossover, adequate washout and attention to period effects are important. There must also be a meaningful chance of observing each condition in the target population. In observational data, confounding by factors associated with treatment and outcome can make a treatment contrast differ from the causal effect.
Worked Example: Victim-Drug AUC Ratio at a Fixed Dose
In a hypothetical study, all participants receive a victim-drug dose of 100 mg. The estimated population geometric mean AUC over 24 hours is 80 mg·h/L without the perpetrator and 120 mg·h/L with it. Define the exposure estimand as the geometric mean AUC ratio at this dose.
Formula and substitution:
The AUC units cancel. The estimated geometric mean AUC with the perpetrator is 1.50 times the value without it, or 50% higher, because \((1.50-1)\times100\%=50\%\). This conclusion applies to the specified 100 mg regimen, 24-hour interval, and target population. It does not establish the same ratio at other doses or identify the mechanism behind the difference.
The comparison is interpretable as a fixed-dose exposure contrast only if the dose and regimen are comparable between conditions and the AUC summaries refer to the same interval. If the conditions are measured in different populations, differences in their composition could also contribute unless the design or analysis addresses them.
Worked Example: Total Response Contrast at Fixed Doses
In a separate hypothetical comparison, the victim drug is given at 100 mg and the perpetrator at 50 mg in the combination condition. The endpoint is each participant’s mean response over 0–24 hours, in response units. The estimated population mean is 42 units with the combination and 30 units with the victim drug alone at 100 mg.
Formula and arithmetic:
At these fixed doses, the estimated mean response is 12 units higher with the perpetrator present. This is the total fixed-dose response contrast for the stated 0–24-hour endpoint. It is not a response contrast at equal victim-drug concentrations: the concentration profiles may differ between the treatment conditions.
A causal interpretation would require comparable treatment groups or a suitable adjustment for differences that affect both treatment condition and response. The response definition, observation interval, and dose conditions must also be the same in the two arms. Without those conditions, 12 units remains a descriptive difference between the reported means rather than a justified causal effect.
Response Contrasts at Matched Concentrations
A matched-concentration contrast asks a different question: how does response compare with and without the perpetrator under the same specified victim-drug concentration condition? In principle, the condition should describe the relevant concentration over time, not merely one summary number. Two profiles can have the same AUC or \(C_{\max}\) while differing in timing, and those differences can matter for response.
Write \(\mathbf{c}\) for a specified victim-drug concentration-time profile, and let \(Y(\mathbf{c},z)\) denote response under a condition that sets that profile and sets the perpetrator state. A controlled matched-profile contrast is:
This contrast concerns response when the victim-drug concentration profile is held fixed. It is not the total response difference at the same doses, because maintaining the same profile across perpetrator conditions may require different victim-drug doses. It is also not automatically a causal “direct effect”: that interpretation requires a well-defined way to set the profile and assumptions that support the comparison.
In practice, a matched-profile contrast may be estimated from a design that titrates victim-drug doses to target similar profiles, or from a model that predicts response under a common concentration profile. Both approaches require adequate overlap in the data and appropriate control of other differences between conditions. Simply selecting participants whose observed concentrations happen to be similar does not create a randomized concentration match; the reasons their concentrations differ may also be related to response.
Worked Example: Response Difference at a Matched Concentration
Suppose a hypothetical dose-titration study targets a constant victim-drug concentration of 4 mg/L in both conditions. The victim-drug dose is adjusted as needed to achieve that target profile. The estimated mean response over 0–24 hours is 38 units with the perpetrator and 32 units without it.
Formula and arithmetic:
Under the specified matched-profile comparison, the estimated response is 6 units higher with the perpetrator. Because the victim-drug dose may differ between conditions, this does not estimate the fixed-dose total response effect. The interpretation also depends on whether the target profile was actually achieved comparably and whether other relevant conditions were controlled.
Matching only AUC at 4 mg·h/L would be a different specification from matching a constant concentration of 4 mg/L. The units themselves indicate the distinction: concentration is measured in mg/L, whereas AUC is measured in mg·h/L. Equal AUC alone need not imply equal concentration at every time.
How the Estimands Differ in Practice
The main distinction is the condition held fixed. A fixed-dose comparison holds the administered doses constant and allows victim-drug exposure to differ. A matched-concentration comparison holds a defined concentration condition constant, potentially requiring different doses. Neither is universally preferable: each answers a different scientific question.
| Estimand | Held fixed | Question answered | Key caution |
|---|---|---|---|
| Exposure ratio | Victim-drug dose and regimen | How does exposure differ with versus without the perpetrator at this dose? | Does not identify the mechanism or guarantee the same effect at other doses. |
| Fixed-dose response contrast | Victim- and perpetrator-drug doses | How does the response endpoint differ under these dose regimens? | Exposure may differ, so the contrast is not at matched concentrations. |
| Matched-concentration response contrast | Specified victim-drug concentration condition | How does response differ with versus without the perpetrator at that concentration condition? | Requires a meaningful match and sufficient data support; matching a summary may not match the whole profile. |
The estimand should also state how the population summary is calculated. A mean difference is sensitive to the response scale and can be difficult to interpret if the response is highly skewed. A geometric mean ratio is appropriate for positive exposure values analyzed on a log scale, but it is not the same quantity as an arithmetic mean difference. Units should be retained for differences and stated for exposure summaries, even though they cancel in a ratio.
Assumptions, Diagnostics, and Limitations
Before estimating any of these quantities, verify that the treatment conditions are defined precisely. Specify the victim and perpetrator doses, timing, duration, exposure interval, response endpoint, and target population. A contrast between “with” and “without” the perpetrator is ambiguous if its dose or timing is unspecified. For repeated-measures or crossover data, account for the design rather than treating repeated observations as independent.
For a matched-concentration analysis, inspect whether both treatment conditions have observations near the target concentration profile. If one condition has little or no support at that profile, model-based predictions may rely heavily on extrapolation. Compare concentration-time profiles, not just one metric, when the response depends on concentration timing. Also check for meaningful differences in dose, adherence, measurement schedules, and other factors that could influence response.
A useful quality check is to compute the estimand directly from its definition before interpreting a model coefficient. Confirm that the coefficient corresponds to the chosen population, endpoint, contrast, and scale. For example, a coefficient on log AUC may need exponentiation to become a geometric mean ratio; it is not an AUC difference in mg·h/L. Likewise, a model prediction at a common concentration is not the fixed-dose response contrast unless the predicted dose conditions are also the same.
These estimands summarize specified comparisons; they do not by themselves show that an interaction is clinically important, explain its mechanism, or establish that results generalize to other doses, populations, and regimens. Estimates also carry uncertainty, which should be reported with an appropriate interval or other uncertainty measure. The examples above give hypothetical point estimates only; they do not imply that the underlying population quantities are known without error.
Common Mistake: Calling a Matched-Concentration Contrast the Total Effect
Wrong reasoning: “The response difference is 6 units after matching concentration, so the combination has a total 6-unit effect at the original fixed doses.”
Correction: The 6-unit contrast compares conditions at a matched concentration profile, potentially using different victim-drug doses. The total fixed-dose response contrast instead compares the same victim- and perpetrator-drug doses and permits concentrations to differ. These are distinct estimands; one cannot be substituted for the other without additional assumptions and a clearly defined analysis.
Check Your Understanding
Answer each question using the distinction between fixed-dose and matched-concentration estimands.
- At a fixed victim-drug dose, geometric mean AUC is 90 mg·h/L without the perpetrator and 135 mg·h/L with it. What is the exposure ratio, and what is the percentage increase?
- At fixed doses, mean response is 24 units without the perpetrator and 19 units with it. What is the response difference, expressed as with minus without?
- Why does matching victim-drug AUC not necessarily mean that the concentration-time profiles are matched?
- A researcher selects participants with similar observed concentrations in two treatment groups and calls the resulting response contrast a causal matched-concentration effect. Identify one key problem with this reasoning.
Solutions. (1) \(135/90=1.50\); the geometric mean AUC is 50% higher with the perpetrator. (2) \(19-24=-5\) response units, so mean response is 5 units lower with the perpetrator at the specified fixed doses. (3) AUC is an integrated summary: profiles can have the same area but different peaks, troughs, or timing. (4) Selection on observed concentration does not set concentration by intervention; factors affecting both concentration and response can differ between the selected groups, so the contrast may be confounded and may not represent the intended controlled comparison.