From a Development Concern to a Question With a Target
Drug-development questions often begin in broad language: “Will the new dose work better?”, “Is the regimen safe for patients with impaired renal function?”, or “What exposure should we expect in children?” These are useful starting points, but they do not yet say exactly what quantity a model should produce. “Work better” could refer to several outcomes, populations, and follow-up times; “safe” could mean a particular event, its frequency, or its timing.
A modelable question makes those choices explicit before model building begins. It defines who or what the answer concerns, which intervention or scenario is being considered, what outcome is measured, when it is assessed, and what numerical quantity is wanted. This framing does not guarantee that available data or a particular model can answer the question. It does make clear what an answer would mean and what information would be needed to produce one.
The components should be stated in enough detail that two analysts would identify the same target. “Adults” may need a disease stage or treatment history; “the new dose” may need dose amount, route, schedule, and duration; “response” needs an operational definition; and “at follow-up” needs a time point or interval. The amount of detail depends on the question, but ambiguity in any component can change the target itself.
A Practical Specification
Write the question in ordinary language first, then fill in each component. A comparator is not always required: a question may ask for an outcome under one regimen. When the development concern is whether one approach differs from another, specify both approaches and the direction of the contrast. Use units for measurements and define how the target population is represented.
Define the people, patients, or experimental units to which the answer is intended to apply. Include characteristics that materially delimit the group.
State the regimen or conditions being considered, and name a comparator if the question asks for a contrast.
Say what is measured, how it is defined, and in what units or categories. Avoid outcome labels that can have multiple meanings.
Specify the assessment time or interval. An outcome at week 4 is not interchangeable with the same outcome at week 24.
Choose the summary that answers the question, such as a population mean, proportion, difference between means, or predicted concentration for a specified individual profile.
A compact mathematical statement can help expose omissions. Let \(Y(t)\) denote an outcome at time \(t\), and let \(A\) and \(B\) denote two specified regimens. If the outcome is a numerical change from baseline, a population-average contrast at a fixed horizon \(T\) can be written as
This expression is not complete until the target population, the definition of \(Y\), the regimens, and \(T\) are stated. It also does not, by notation alone, establish that the contrast is causal. A causal interpretation requires an appropriate study design and assumptions; a model can calculate a contrast without those requirements being met.
Worked Translations
Worked Example: Turning “Will the Higher Dose Work Better?” Into an Efficacy Target
Suppose a development team asks whether a higher dose improves symptom scores in adults with a specified chronic condition. For this hypothetical example, the score ranges from 0 to 100, with lower scores indicating fewer symptoms. The proposed regimens are 40 mg once daily and 20 mg once daily, and the intended assessment is at week 12.
Population: adults with the specified condition who meet the study’s stated eligibility criteria. Naming the disease and eligibility criteria matters: an answer for treatment-naive adults need not represent adults previously treated with other therapies.
Intervention and comparator: 40 mg once daily versus 20 mg once daily, both taken for 12 weeks. “Higher dose” alone would not identify the regimen or its duration.
Outcome and horizon: change from baseline in the symptom score at week 12, in score points. Define change as the week-12 score minus the baseline score, so a more negative value represents greater improvement.
Target quantity: the difference between the population mean changes for the two regimens, higher dose minus lower dose. Suppose, purely hypothetically, a fitted analysis gives mean changes of \(-12.4\) points and \(-9.1\) points, respectively. Then
The model-based contrast is 3.3 points more improvement, on average, for the higher-dose regimen under this definition of change. It is not a statement about every patient, and the sign would be reversed if the contrast were defined as lower dose minus higher dose. Whether this estimate supports a development conclusion depends on the quality and relevance of the evidence, not just the question’s wording.
Worked Example: Defining a Population Safety Prediction
A team asks, “How safe is the proposed regimen in patients with reduced kidney function?” For a concrete hypothetical question, define the population as adults with moderate renal impairment receiving a specified regimen. Define the outcome as at least one protocol-defined serious adverse event by day 30, and request the proportion of this population expected to experience that outcome by day 30.
Suppose the target population contains two clinically defined renal-function strata, with target-population proportions of 70% and 30%. A model predicts day-30 event probabilities of 0.08 and 0.16 in the respective strata. If those strata exhaust the target population and the stated proportions are appropriate, the population-average probability is the weighted average:
The resulting target is a predicted 10.4% experiencing at least one such event by day 30 in the specified population under the regimen. The arithmetic depends on the stratum-specific predictions and population weights being suitable for that population; different weights imply a different target-population average. The result does not describe all adverse events, events after day 30, or severity beyond the event definition.
Worked Example: Replacing “What Exposure Will Children Have?” With a Prediction Target
A broad exposure question does not identify an age range, dose, sampling time, or summary. Suppose the intended question concerns children aged 6 to under 12 years receiving 5 mg/kg once daily, and the outcome is plasma concentration 24 hours after the fifth dose, in mg/L. The team wants the median concentration in this population, rather than a mean or a prediction for one particular child.
A hypothetical model output for the specified population and regimen gives a median 24-hour concentration of 1.35 mg/L. This is already the requested target quantity: the median predicted concentration at that sampling time for the defined population. It is not a prediction for every child, nor does it answer what the concentration will be after a different dose or at a different time.
If the development question instead asks how the target changes under a second regimen, that is a different target. For example, suppose a hypothetical model predicts medians of 1.35 mg/L under 5 mg/kg and 1.62 mg/L under 6 mg/kg. The difference, defined as the 6 mg/kg regimen minus the 5 mg/kg regimen, is
That contrast answers a regimen-comparison question, while the first median answers a single-regimen prediction question. Both are modelable, but they should not be substituted for one another.
Why the Target Quantity Matters
Once population, intervention, outcome, and time are fixed, a question may still have several valid target quantities. A mean describes an average numerical outcome; a proportion describes the frequency of a defined event; a difference compares summaries between regimens; and an individual prediction requires a specified individual profile. These quantities answer different questions, even when they use the same data and model.
For example, a population mean response does not say how many patients meet a response threshold. A proportion responding does not say how large the average score change is. A population median concentration does not describe the range of individual concentrations. Select the summary that corresponds to the development concern, and name it explicitly rather than relying on a vague word such as “effect,” “benefit,” or “exposure.”
For a population divided into mutually exclusive strata, the overall mean is the weighted mean of stratum means when the weights are the proportions in the target population. If \(w_k\) is the target-population proportion in stratum \(k\), and \(\mu_k\) is its mean outcome, then:
This follows by averaging within each stratum and then averaging across the population: each stratum contributes its mean in proportion to how much of the target population it represents. If the model’s stratum weights instead reflect a study sample with a different composition, the resulting overall average may answer a different population question. The same logic applies to event probabilities, as in the safety example.
A useful check is to read the target quantity aloud with all its qualifiers: “the mean change in symptom score, in score points, at week 12, for eligible adults under 40 mg once daily, minus the corresponding mean under 20 mg once daily.” If the sentence cannot be completed without guessing, the question is not yet specific enough.
Assumptions, Diagnostics, and Common Pitfalls
The framework is a specification discipline, not a statistical estimator. Its basic requirement is that the defined quantity be coherent: the population can be identified, the intervention or scenario is sufficiently defined, the outcome can be determined, the horizon is unambiguous, and the requested summary is mathematically meaningful. To interpret an estimated contrast as causal, further design and identification conditions are required. To interpret a model prediction for a target population, the model and its inputs must support that target; the question alone cannot supply missing evidence.
Several checks help detect a poorly framed question before analysis:
- Population check: Can a person or unit be classified as inside or outside the target group using stated criteria? If not, the population may be too vague.
- Intervention check: Are dose, route, schedule, and relevant duration clear enough to distinguish the intervention from alternatives?
- Outcome and time check: Is the outcome operationally defined and tied to a specified time or interval?
- Quantity check: Is the target a mean, proportion, contrast, distributional summary, or individual prediction? Are its units and direction clear?
- Population-composition check: If results are combined across strata, do the weights correspond to the target population?
A common mistake is to treat a development slogan as if it were already a model question: “Does the drug work?” The reasoning is that efficacy is the concern, so any efficacy-related model output should answer it. That is incorrect. The response could be a mean score change at week 4, a response proportion at week 12, or a time-to-event summary at one year; each has a different outcome definition and horizon. The correction is to specify the population, regimen and comparator if needed, outcome, time, and target quantity before selecting an analysis.
Another pitfall is to name a target precisely but overstate what its estimate establishes. A model-based difference in outcomes is not automatically evidence that the intervention caused the difference. Likewise, a predicted population mean is not a guarantee for an individual. Keep the wording aligned with the target and the evidence used to estimate it.
Choosing Among Valid Formulations
The main alternative formulations are not competing statistical methods; they are different questions that may be appropriate for different development concerns. The following distinctions help make the choice explicit.
| Formulation | Example target | Question it answers |
|---|---|---|
| Single-regimen population summary | Mean score change at week 12 under 40 mg once daily | What outcome is expected under this regimen? |
| Population contrast | Difference in mean score change at week 12 between 40 mg and 20 mg once daily | How do the specified regimen summaries differ? |
| Event probability | Proportion with a defined serious adverse event by day 30 | How frequent is this particular event by this time? |
| Individual-level prediction | Predicted 24-hour concentration for a child with stated characteristics and regimen | What outcome is predicted for this specified profile? |
A formulation becomes less useful when its summary no longer matches the question. For example, a population-average contrast can conceal meaningful differences across subgroups, while a subgroup-specific answer requires defining that subgroup and its target quantity. A numerical outcome summary also cannot substitute for a clinically defined event probability. State the target first; whether a model and evidence can support it is a separate methodological question.
Check Your Understanding
For each item, identify what is missing or state the target quantity precisely. Short solutions follow the questions.
- “Will the new regimen improve disease control?” Name at least four specifications needed to make this modelable.
- A model predicts mean change of \(-8.2\) points under regimen A and \(-5.6\) points under regimen B at week 8. Define the contrast as A minus B. Calculate and interpret it in score points.
- A target population is 40% in stratum 1 and 60% in stratum 2. Predicted event probabilities by day 14 are 0.05 and 0.10. Calculate the target-population probability, assuming the strata cover the population and the weights apply.
- Why does a median predicted concentration for a population not give the predicted concentration for every individual in that population?
- What additional definition is required to turn “adverse events” into a specific safety outcome for a model question?
Solutions: (1) Specify the target population; the new regimen and, if comparison is intended, comparator; an operational disease-control outcome; the assessment time; and a target quantity such as a mean difference or response proportion. (2) \((-8.2)-(-5.6)=-2.6\) points. Under the stated change-score convention, regimen A has 2.6 points more mean improvement at week 8; the hypothetical contrast is not by itself a causal conclusion. (3) \(0.40(0.05)+0.60(0.10)=0.02+0.06=0.08\), or 8.0% by day 14. (4) A median is a population distribution summary: it does not specify each person’s value, and individual predictions require individual characteristics and a defined prediction target. (5) State which event or events count, how they are defined, and the time horizon over which they are counted.