Why Science? | ICH E22 Patient Preference Studies and Benefit-Risk

The ICH Assembly endorsed E22 at Steps 2a and 2b on 19 November 2025. Training and explanatory materials were released in January 2026, and the guideline entered Step 3 consultation. Its recommendations should be identified as draft principles until Step 4 adoption and regional implementation occur.

The listening in e22 is a step 3 draft becomes evidence when the task is explicit. The ICH Assembly endorsed E22 at Steps 2a and 2b on 19 November 2025. Training and explanatory materials were released in January 2026, and the guideline entered Step 3 consultation. Its recommendations should be identified as draft principles until Step 4 adoption and regional implementation occur. Keep this boundary visible: E22 remained a Step 3 draft on 27 October 2026.

A strong science note separates observation from interpretation. Trace e22 is a step 3 draft from regulatory question through patient input, instrument item, response and estimate. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Next chapter

Stated-preference studies use surveys or interviews in which participants express choices, rankings or acceptable trade-offs for hypothetical treatment scenarios. E22 distinguishes these from revealed-preference methods based on observed behaviour. That scope boundary matters because the two evidence types have different biases, designs and interpretations.

For the draft focuses on stated-preference methods, design choices can change the answer. Stated-preference studies use surveys or interviews in which participants express choices, rankings or acceptable trade-offs for hypothetical treatment scenarios. E22 distinguishes these from revealed-preference methods based on observed behaviour. That scope boundary matters because the two evidence types have different biases, designs and interpretations. Preserve this guardrail: The draft scope is stated preference and does not cover every form of observed choice.

A strong science note separates observation from interpretation. Ask whether a participant could understand and realistically imagine the draft focuses on stated-preference methods. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Patient-reported outcomes capture a patient’s report of health, symptoms, function or quality of life. A patient preference study estimates the value placed on treatment characteristics or outcomes. The same person may answer both, but the measurements serve different scientific questions and should not be merged by convenient terminology.

A useful account of a pps is not a patient-reported outcome measure links what patients saw to what analysts estimated. Patient-reported outcomes capture a patient’s report of health, symptoms, function or quality of life. A patient preference study estimates the value placed on treatment characteristics or outcomes. The same person may answer both, but the measurements serve different scientific questions and should not be merged by convenient terminology. Remember: Caregiver and professional preferences cannot be labelled patient preferences.

A strong science note separates observation from interpretation. Stress-test a pps is not a patient-reported outcome measure for framing, order, missing attributes and sample selection. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The research objective should identify the treatment or development decision, population, alternatives, outcomes and expected use of results. Preference evidence might inform target product design, endpoint selection, clinical-trial features, subgroup understanding or benefit-risk assessment. A generic question about what patients like is too loose for regulatory interpretation.

The listening in the decision context comes before the questionnaire becomes evidence when the task is explicit. The research objective should identify the treatment or development decision, population, alternatives, outcomes and expected use of results. Preference evidence might inform target product design, endpoint selection, clinical-trial features, subgroup understanding or benefit-risk assessment. A generic question about what patients like is too loose for regulatory interpretation. Keep this boundary visible: A preference estimate is conditional on the scenario, sample, method and model.

A strong science note separates observation from interpretation. Translate the result for the decision context comes before the questionnaire into a decision use without exceeding the study population. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Patients can help identify relevant concepts, confusing wording, unrealistic scenarios, burden and missing attributes before the instrument is fixed. Interviews, focus groups or advisory work can improve content validity. Patient involvement in designing a study is valuable, but it is different from being a research participant whose preferences enter the analysis.

For patient input should shape the study early, design choices can change the answer. Patients can help identify relevant concepts, confusing wording, unrealistic scenarios, burden and missing attributes before the instrument is fixed. Interviews, focus groups or advisory work can improve content validity. Patient involvement in designing a study is valuable, but it is different from being a research participant whose preferences enter the analysis. Preserve this guardrail: E22 remained a Step 3 draft on 27 October 2026.

A strong science note separates observation from interpretation. Trace patient input should shape the study early from regulatory question through patient input, instrument item, response and estimate. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Literature, clinical expertise and qualitative research can identify candidate benefits, risks, convenience features and process burdens. Researchers should document how concepts were included, combined or excluded. An omitted attribute can distort choices if participants use it mentally while the model pretends it is absent.

A useful account of preliminary qualitative research builds the attribute set links what patients saw to what analysts estimated. Literature, clinical expertise and qualitative research can identify candidate benefits, risks, convenience features and process burdens. Researchers should document how concepts were included, combined or excluded. An omitted attribute can distort choices if participants use it mentally while the model pretends it is absent. Remember: The draft scope is stated preference and does not cover every form of observed choice.

A strong science note separates observation from interpretation. Ask whether a participant could understand and realistically imagine preliminary qualitative research builds the attribute set. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

An attribute represents a treatment feature such as effectiveness, adverse-event probability, route or frequency. Attributes should be distinct enough that participants can understand and trade them independently. Overlap between symptom improvement and quality of life, for example, may lead to double counting or inconsistent interpretation.

The listening in attributes need clear, non-overlapping meanings becomes evidence when the task is explicit. An attribute represents a treatment feature such as effectiveness, adverse-event probability, route or frequency. Attributes should be distinct enough that participants can understand and trade them independently. Overlap between symptom improvement and quality of life, for example, may lead to double counting or inconsistent interpretation. Keep this boundary visible: Caregiver and professional preferences cannot be labelled patient preferences.

A strong science note separates observation from interpretation. Stress-test attributes need clear, non-overlapping meanings for framing, order, missing attributes and sample selection. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Levels describe possible values of each attribute. They should span realistic differences, use understandable units and avoid framing that makes one option obviously dominant. Risk levels need time horizons and denominators. A 2% lifetime risk and a 2% one-year risk are not interchangeable even though the number is the same.

For levels must be plausible and decision-relevant, design choices can change the answer. Levels describe possible values of each attribute. They should span realistic differences, use understandable units and avoid framing that makes one option obviously dominant. Risk levels need time horizons and denominators. A 2% lifetime risk and a 2% one-year risk are not interchangeable even though the number is the same. Preserve this guardrail: A preference estimate is conditional on the scenario, sample, method and model.

A strong science note separates observation from interpretation. Translate the result for levels must be plausible and decision-relevant into a decision use without exceeding the study population. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Discrete-choice experiments, threshold questions, best-worst scaling, ranking and other stated-preference approaches answer different questions and impose different cognitive tasks. The method should match the decision, population, attribute count and desired estimand. Fashion or software availability is not a sufficient reason for choosing a design.

A useful account of method choice should fit the research question links what patients saw to what analysts estimated. Discrete-choice experiments, threshold questions, best-worst scaling, ranking and other stated-preference approaches answer different questions and impose different cognitive tasks. The method should match the decision, population, attribute count and desired estimand. Fashion or software availability is not a sufficient reason for choosing a design. Remember: E22 remained a Step 3 draft on 27 October 2026.

A strong science note separates observation from interpretation. Trace method choice should fit the research question from regulatory question through patient input, instrument item, response and estimate. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Choice sets are constructed so attribute levels vary in a way that permits estimation of preference weights and interactions. Balance, overlap, blocking and number of tasks affect precision and burden. Pilot testing and simulation can reveal dominant alternatives, weak level contrasts or parameters the design cannot identify.

The listening in experimental design determines what can be estimated becomes evidence when the task is explicit. Choice sets are constructed so attribute levels vary in a way that permits estimation of preference weights and interactions. Balance, overlap, blocking and number of tasks affect precision and burden. Pilot testing and simulation can reveal dominant alternatives, weak level contrasts or parameters the design cannot identify. Keep this boundary visible: The draft scope is stated preference and does not cover every form of observed choice.

A strong science note separates observation from interpretation. Ask whether a participant could understand and realistically imagine experimental design determines what can be estimated. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Descriptions, graphics, ordering, reference points and risk formats influence understanding. Presenting both natural frequencies and consistent time horizons may help. Materials should be accurate and balanced, then tested for comprehension. A preference elicited after a misunderstood scenario is a response, but it is not evidence about the intended trade-off.

For information presentation can change preferences, design choices can change the answer. Descriptions, graphics, ordering, reference points and risk formats influence understanding. Presenting both natural frequencies and consistent time horizons may help. Materials should be accurate and balanced, then tested for comprehension. A preference elicited after a misunderstood scenario is a response, but it is not evidence about the intended trade-off. Preserve this guardrail: Caregiver and professional preferences cannot be labelled patient preferences.

A strong science note separates observation from interpretation. Stress-test information presentation can change preferences for framing, order, missing attributes and sample selection. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Eligibility, recruitment channel, geography, disease stage, treatment experience, literacy and access can shape preferences. A convenience sample may still answer a bounded exploratory question, but it should not be labelled representative without evidence. Caregiver or healthcare-professional preferences are informative yet cannot silently replace patient preferences.

A useful account of the study sample sets the population boundary links what patients saw to what analysts estimated. Eligibility, recruitment channel, geography, disease stage, treatment experience, literacy and access can shape preferences. A convenience sample may still answer a bounded exploratory question, but it should not be labelled representative without evidence. Caregiver or healthcare-professional preferences are informative yet cannot silently replace patient preferences. Remember: A preference estimate is conditional on the scenario, sample, method and model.

A strong science note separates observation from interpretation. Translate the result for the study sample sets the population boundary into a decision use without exceeding the study population. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

There is no single participant count suitable for every PPS. Precision depends on design efficiency, task number, attribute structure, heterogeneity and intended subgroup analysis. Planning may use simulations or formulae plus allowance for incomplete or low-quality responses. A large sample cannot repair an invalid instrument.

The listening in sample size follows the design and estimand becomes evidence when the task is explicit. There is no single participant count suitable for every PPS. Precision depends on design efficiency, task number, attribute structure, heterogeneity and intended subgroup analysis. Planning may use simulations or formulae plus allowance for incomplete or low-quality responses. A large sample cannot repair an invalid instrument. Keep this boundary visible: E22 remained a Step 3 draft on 27 October 2026.

A strong science note separates observation from interpretation. Trace sample size follows the design and estimand from regulatory question through patient input, instrument item, response and estimate. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Researchers may examine completion time, dominance, consistency, comprehension, straight-lining or repeated-task agreement. Excluding responses after results are known can change estimates, so criteria and sensitivity analyses should be planned. Apparent inconsistency can also reflect genuine indifference, fatigue or a model that poorly represents how people decide.

For response quality deserves planned checks, design choices can change the answer. Researchers may examine completion time, dominance, consistency, comprehension, straight-lining or repeated-task agreement. Excluding responses after results are known can change estimates, so criteria and sensitivity analyses should be planned. Apparent inconsistency can also reflect genuine indifference, fatigue or a model that poorly represents how people decide. Preserve this guardrail: The draft scope is stated preference and does not cover every form of observed choice.

A strong science note separates observation from interpretation. Ask whether a participant could understand and realistically imagine response quality deserves planned checks. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Statistical models estimate relative preference weights, trade-offs, choice probabilities or thresholds under assumptions about utility and error. Coding, reference levels, interactions and scale affect interpretation. Results should include uncertainty and avoid presenting a derived willingness-to-accept measure as a universal patient value.

A useful account of analysis translates choices into conditional estimates links what patients saw to what analysts estimated. Statistical models estimate relative preference weights, trade-offs, choice probabilities or thresholds under assumptions about utility and error. Coding, reference levels, interactions and scale affect interpretation. Results should include uncertainty and avoid presenting a derived willingness-to-accept measure as a universal patient value. Remember: Caregiver and professional preferences cannot be labelled patient preferences.

A strong science note separates observation from interpretation. Stress-test analysis translates choices into conditional estimates for framing, order, missing attributes and sample selection. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

Average estimates can hide people who prioritise benefits, risks or convenience differently. Latent-class, mixed-model or subgroup analyses may explore variation, but classes need stability, interpretability and adequate data. A data-driven cluster is a model result, not automatically a clinically recognised patient population.

The listening in preference heterogeneity may reveal important groups becomes evidence when the task is explicit. Average estimates can hide people who prioritise benefits, risks or convenience differently. Latent-class, mixed-model or subgroup analyses may explore variation, but classes need stability, interpretability and adequate data. A data-driven cluster is a model result, not automatically a clinically recognised patient population. Keep this boundary visible: A preference estimate is conditional on the scenario, sample, method and model.

A strong science note separates observation from interpretation. Translate the result for preference heterogeneity may reveal important groups into a decision use without exceeding the study population. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter · Next chapter

The receipt records draft status, decision context, population, patient involvement, qualitative work, attribute and level rationale, method, experimental design, instrument versions, risk presentation, sample and recruitment, sample-size basis, quality rules, model code, heterogeneity analysis, uncertainty, sensitivity checks, proposed regulatory use and limitations.

For a complete e22 receipt makes the study interpretable, design choices can change the answer. The receipt records draft status, decision context, population, patient involvement, qualitative work, attribute and level rationale, method, experimental design, instrument versions, risk presentation, sample and recruitment, sample-size basis, quality rules, model code, heterogeneity analysis, uncertainty, sensitivity checks, proposed regulatory use and limitations. Preserve this guardrail: E22 remained a Step 3 draft on 27 October 2026.

A strong science note separates observation from interpretation. Trace a complete e22 receipt makes the study interpretable from regulatory question through patient input, instrument item, response and estimate. Record the source, version or date, population or device context, method, result and one uncertainty. If the official record does not show a field, write “not shown” instead of filling the gap from memory.

Contents · Previous chapter