From Clinical Question to Statistical Question

Overview

Clinical research often begins with a concern rather than a fully formed question: too many patients are readmitted, a diagnostic test appears promising, or one treatment seems better tolerated than another. Biostatistics cannot analyse a concern until it has been translated into a question that identifies what will be observed, compared and estimated.

A clinical question expresses the decision or uncertainty that matters to patients, clinicians or health services. A research question converts that concern into a question that can be addressed using a defined population, design and set of measurements. A statistical question specifies the numerical quantity that the study intends to estimate and the uncertainty that will accompany that estimate.

This lecture explains that translation process. The central principle is that researchers should define the question and target quantity before choosing a statistical test or opening statistical software.

Why this matters

A vague question produces ambiguous data. “Does follow-up help patients with heart failure?” does not identify which patients, what follow-up consists of, what it is compared with, what “help” means or how long patients are observed.

Different interpretations can lead to different studies and different answers. One researcher may measure 30-day readmission, another may measure quality of life at six months, and another may compare mortality over one year. None is automatically wrong, but they answer different questions.

The Population, Intervention, Comparator and Outcome framework was introduced to help clinicians construct focused, answerable questions [1]. Adding a time frame is often helpful, producing the mnemonic PICOT. A focused question also guides study design, outcome measurement, sample-size planning and analysis [2].

From concern to answerable question

Step 1: identify the decision

The team first states the decision that the research should inform:

Should the hospital add structured nurse-led telephone follow-up to usual discharge care for eligible adults admitted with heart failure?

This decision is clinically meaningful, but it is not yet sufficiently precise for a protocol or analysis.

Step 2: specify PICOT

PICOT is a practical prompt rather than a statistical method. Each element must have a practical definition that can be applied consistently.

Table 1. PICOT specification for the simulated heart-failure trial

Note: The eligibility criteria, intervention procedures and outcome-ascertainment rules would require further detail in the full protocol.

The resulting clinical research question is:

Among eligible adults discharged after an admission for heart failure, does allocation to structured nurse-led telephone follow-up within 72 hours, compared with usual discharge care, reduce the risk of unplanned readmission or death within 30 days?

The words allocation to matter. They indicate that the trial evaluates the effect of being assigned to the follow-up strategy, not only the effect among participants who successfully complete the telephone call.

Step 3: define the unit of analysis

The unit of analysis is the entity represented by each independent contribution to the main comparison. Participants are individually randomised, and each participant contributes one 30-day outcome. The unit of analysis is therefore the participant.

If entire wards or hospitals were randomised, participants within the same cluster could have correlated outcomes. The statistical question and analysis would then need to reflect the cluster-randomised design rather than treating all participants as unrelated observations.

Step 4: define the target quantity

The team chooses the 30-day risk difference, calculated as:

risk in the structured-follow-up group minus risk in the usual-care group.

A negative risk difference favours structured follow-up because the adverse composite outcome is less frequent. The scale and direction are defined before viewing the results.

Research questions have different purposes

Not every clinical question asks whether an intervention works. The purpose of the question determines the quantity that should be estimated.

Table 2. Common purposes of medical research questions

Note: These examples require different designs, assumptions and analyses. A prediction question is not the same as a question about the effect of an intervention, even when both use some of the same variables.

A useful research question therefore states not only the topic but also the purpose. “Factors related to readmission” is ambiguous: it could refer to causal explanation, description of associations or prediction of future events.

From research question to statistical question

Outcome and comparison

The primary outcome must be defined precisely enough that two trained assessors would classify the same event consistently. In the scenario, the protocol should define:

• what counts as an unplanned admission;
• whether emergency observation without admission is included;
• how transfers are handled;
• how deaths are confirmed;
• when the 30-day period begins and ends; and
• how the first event is selected if more than one event occurs.

The comparator must also be explicit. “Usual care” should be described rather than treated as a universal standard, because usual discharge practices can differ across hospitals and over time.

Estimand, estimator and estimate

An estimand is the target quantity that answers the research question. It describes what is to be estimated.

An estimator is the statistical procedure used to calculate an estimate of that target quantity from the observed data.

An estimate is the numerical result produced by the estimator.

The International Council for Harmonisation E9(R1) addendum emphasises that the clinical objective, estimand and statistical analysis should be aligned [3]. Reviews of published trial protocols have shown that the precise treatment effect can often be difficult to determine when these elements are not stated clearly [4].

For the simulated trial:

• estimand: the difference in 30-day risk of unplanned readmission or death between allocation to structured follow-up and allocation to usual care among all randomised eligible adults, regardless of whether the telephone review is completed;
• estimator: the observed difference between the two group risks, with a 95% confidence interval calculated using the Newcombe score method;
• estimate: the numerical risk difference obtained from the simulated trial data.

The estimand is not the same as a database column or a statistical test. It is the population-level effect that the analysis is intended to learn about.

Hypotheses are secondary to the target effect

For the risk difference, a null hypothesis may state that the population risk difference is zero. A two-sided alternative may state that it is not zero.

These hypotheses can support a formal test, but they do not replace the clinical question. A result can be statistically incompatible with a zero difference yet too small to justify implementation. Conversely, a clinically important effect may remain uncertain if the confidence interval is wide.

Data structure

Simulated participant-level trial data

simulated data

Table 3. Example records from the heart-failure trial

participant_id	allocated_group	call_completed	event_30d
HF001	Structured follow-up	Yes	No
HF002	Usual care	Not applicable	Yes
HF003	Structured follow-up	No	No
HF004	Usual care	Not applicable	No
HF005	Structured follow-up	Yes	Yes

Note: event_30d indicates the first unplanned readmission or all-cause death within 30 days. Each row represents one randomised participant.

The primary analysis retains participant HF003 in the structured-follow-up group even though the call was not completed. Excluding this participant would change the question from the effect of allocation to the effect among a selected subgroup whose follow-up was completed.

The variable call_completed remains useful for describing implementation, but it is measured after randomisation. A comparison restricted to participants with completed calls would no longer preserve the original randomised comparison and could be biased because completion may be related to health status or social circumstances.

The five attributes of a treatment estimand MMed extension

For a clinical trial, a treatment estimand is commonly described using five connected attributes [3]:

1. Population: the patients to whom the treatment effect applies.
2. Treatment conditions: the strategies being compared.
3. Outcome variable: the event or measurement used to represent benefit or harm.
4. Intercurrent events: events occurring after treatment starts that affect the interpretation or existence of the outcome, such as discontinuation, additional treatment or death.
5. Population-level summary: the contrast used to compare treatment conditions, such as a risk difference or mean difference.

In the simulated trial, failure to complete the scheduled call is an intercurrent event. The primary question uses a treatment-policy strategy: outcomes are compared according to randomised allocation regardless of call completion. A different strategy would answer a different clinical question and might require different data and assumptions.

Worked example

Suggested reporting

Current protocol and reporting guidance for randomised trials emphasises clear objectives, prespecified outcomes, transparent methods and complete presentation of results [6], [7].

The Methods wording identifies the question, population, treatment conditions, outcome, time frame, effect measure and estimator. The Results wording reports group denominators, the direction of subtraction, the effect estimate and its uncertainty.

Common mistakes

Knowledge check

Why is 'Does telephone follow-up improve heart failure outcomes?' not yet an adequate statistical question?

Answer and reasoning

It does not define the target population, the exact follow-up strategy, the comparator, the outcome, the time frame or the numerical effect to be estimated. Several valid studies could answer different versions of this broad question. The elements must be specified before an analysis can be selected.

What is the estimand in the simulated trial?

Answer and reasoning

It is the population-level difference in the 30-day risk of unplanned readmission or death between allocation to structured follow-up and allocation to usual care among all randomised eligible adults, regardless of call completion. It is the target effect, not the statistical calculation itself.

A researcher proposes analysing only participants who completed the telephone call. Why does this change the question?

Answer and reasoning

The randomised question concerns the effect of allocation to the follow-up strategy. Restricting the intervention group to participants who completed the call selects people after randomisation, and completion may be related to prognosis. The restricted comparison no longer estimates the same allocation effect and may be biased.

What does the estimated risk difference of −5.0 percentage points mean?

Answer and reasoning

The 30-day event risk in the structured-follow-up group was estimated to be 5.0 percentage points lower than in the usual-care group. The negative sign favours structured follow-up because the adverse outcome was defined first and the contrast direction was structured follow-up minus usual care.

Why should the research team not conclude that the intervention worked solely because the reported p-value was below 0.05?

Answer and reasoning

The p-value does not describe the size or clinical importance of the effect. The risk difference and 95% confidence interval show that the estimated reduction was 5.0 percentage points and convey the remaining uncertainty. Implementation also requires consideration of validity, harms, costs and applicability.

Key take-home messages

References

1. Richardson WS, Wilson MC, Nishikawa J, Hayward RS. The well-built clinical question: a key to evidence-based decisions. ACP J Club. 1995;123(3):A12–A13.
2. Browner WS, Newman TB, Cummings SR, Grady DG, Huang AJ, Kanaya AM, et al. Designing clinical research. 5th ed. Philadelphia: Wolters Kluwer; 2022.
3. International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use. Addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials E9(R1). Geneva: ICH; 2019. https://database.ich.org/sites/default/files/E9-R1_Step4_Guideline_2019_1203.pdf
4. Kahan BC, Morris TP, White IR, Tweed CD, Cro S, Dahly D, et al. Estimands in published protocols of randomised trials: urgent improvement needed. Trials. 2021;22:686. https://doi.org/10.1186/s13063-021-05644-4
5. Newcombe RG. Interval estimation for the difference between independent proportions: comparison of eleven methods. Stat Med. 1998;17(8):873–890. https://doi.org/10.1002/%28SICI%291097-0258%2819980430%2917%3A8%3C873%3A%3AAID-SIM779%3E3.0.CO%3B2-I
6. Chan AW, Boutron I, Hopewell S, Moher D, Schulz KF, Collins GS, et al. SPIRIT 2025 statement: updated guideline for protocols of randomised trials. BMJ. 2025;389:e081477. https://doi.org/10.1136/bmj-2024-081477
7. Hopewell S, Chan AW, Collins GS, Hróbjartsson A, Moher D, Schulz KF, et al. CONSORT 2025 statement: updated guideline for reporting randomised trials. BMJ. 2025;389:e081123. https://doi.org/10.1136/bmj-2024-081123