Why Science? | ICH M14 Real-World Data for Non-Interventional Safety Studies

The guidance outlines general principles for planning, designing, analysing and reporting studies that assess medicine safety without protocol-assigned treatment. Drugs, vaccines and other biological products fall within its stated medicine scope.

The scientific value of m14 covers non-interventional studies using real-world data appears when evidence is connected to a defined decision. The guidance outlines general principles for planning, designing, analysing and reporting studies that assess medicine safety without protocol-assigned treatment. Drugs, vaccines and other biological products fall within its stated medicine scope. Keep this boundary visible: Real-world data record healthcare processes as well as health, and those processes can change over time.

A strong science note separates observation from interpretation. Write the population, exposure, comparator, outcome and time frame for m14 covers non-interventional studies using real-world data before inspecting results. 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

The final replaces the July 2024 draft. FDA states that the recommendations complement its other real-world-evidence guidances and withdrew the May 2013 electronic-healthcare-data safety-study guidance.

For fda lists m14 as final level 1 guidance from march 2026, a useful explanation shows the data, assumptions and operational response together. The final replaces the July 2024 draft. FDA states that the recommendations complement its other real-world-evidence guidances and withdrew the May 2013 electronic-healthcare-data safety-study guidance. Preserve this guardrail: Adjustment controls measured confounders only to the extent that variables are valid and models have support.

A strong science note separates observation from interpretation. Trace every variable in fda lists m14 as final level 1 guidance from march 2026 to its operational source and validation evidence. 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

Population, exposure, comparator, outcome, time horizon and causal or descriptive objective are stated before data analysis. A protocol explains the regulatory or safety decision the study is meant to inform.

A trustworthy account of a precise research question anchors every later choice keeps the protocol, implementation and lifecycle record aligned. Population, exposure, comparator, outcome, time horizon and causal or descriptive objective are stated before data analysis. A protocol explains the regulatory or safety decision the study is meant to inform. Remember: A diagnosis code can be a useful signal without being a perfectly timed or confirmed clinical event.

A strong science note separates observation from interpretation. Challenge a precise research question anchors every later choice with a negative control, alternate definition or quantitative bias analysis. 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, treatment strategies, assignment, follow-up, outcome, causal contrast and analysis can be described as if a pragmatic trial were being emulated. The observational implementation then identifies where real-world data depart from that ideal.

The scientific value of the target trial idea can clarify causal questions appears when evidence is connected to a defined decision. Eligibility, treatment strategies, assignment, follow-up, outcome, causal contrast and analysis can be described as if a pragmatic trial were being emulated. The observational implementation then identifies where real-world data depart from that ideal. Keep this boundary visible: A protocol written after results are known cannot provide the same protection against outcome-driven choices.

A strong science note separates observation from interpretation. Reproduce the target trial idea can clarify causal questions from versioned data, code, cohort flow and decision logs. 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

Claims, electronic health records, registries, dispensing data or linked sources differ in population coverage, clinical detail, continuity and timeliness. A famous database is not automatically suitable for a particular exposure, outcome or follow-up period.

For data-source fitness is question specific, a useful explanation shows the data, assumptions and operational response together. Claims, electronic health records, registries, dispensing data or linked sources differ in population coverage, clinical detail, continuity and timeliness. A famous database is not automatically suitable for a particular exposure, outcome or follow-up period. Preserve this guardrail: Real-world data record healthcare processes as well as health, and those processes can change over time.

A strong science note separates observation from interpretation. Write the population, exposure, comparator, outcome and time frame for data-source fitness is question specific before inspecting results. 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 study traces why a field was created, who recorded it, coding systems, transformation, linkage and refresh history. Changes in reimbursement, clinical practice or software can create apparent trends unrelated to biology.

A trustworthy account of provenance reveals how care became a research variable keeps the protocol, implementation and lifecycle record aligned. The study traces why a field was created, who recorded it, coding systems, transformation, linkage and refresh history. Changes in reimbursement, clinical practice or software can create apparent trends unrelated to biology. Remember: Adjustment controls measured confounders only to the extent that variables are valid and models have support.

A strong science note separates observation from interpretation. Trace every variable in provenance reveals how care became a research variable to its operational source and validation evidence. 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

Prescription, dispensing, administration and self-report measure different events. Grace periods, stockpiling, switching, discontinuation and dose changes affect exposed person-time. Immortal time can appear when classification uses future information.

The scientific value of exposure definitions need start, stop and adherence logic appears when evidence is connected to a defined decision. Prescription, dispensing, administration and self-report measure different events. Grace periods, stockpiling, switching, discontinuation and dose changes affect exposed person-time. Immortal time can appear when classification uses future information. Keep this boundary visible: A diagnosis code can be a useful signal without being a perfectly timed or confirmed clinical event.

A strong science note separates observation from interpretation. Challenge exposure definitions need start, stop and adherence logic with a negative control, alternate definition or quantitative bias analysis. 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

Active comparators with similar indications can improve clinical comparability. Non-users or historical controls may differ in healthcare contact and disease severity. The index date and eligibility rules should be applied symmetrically.

For comparator choice shapes confounding and interpretation, a useful explanation shows the data, assumptions and operational response together. Active comparators with similar indications can improve clinical comparability. Non-users or historical controls may differ in healthcare contact and disease severity. The index date and eligibility rules should be applied symmetrically. Preserve this guardrail: A protocol written after results are known cannot provide the same protection against outcome-driven choices.

A strong science note separates observation from interpretation. Reproduce comparator choice shapes confounding and interpretation from versioned data, code, cohort flow and decision logs. 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

Diagnosis codes, procedures, laboratory thresholds, algorithms or adjudication can identify outcomes. Positive predictive value, sensitivity, severity and timing matter. A valid algorithm in one system or era may not transfer unchanged.

A trustworthy account of outcome validation protects against code-only assumptions keeps the protocol, implementation and lifecycle record aligned. Diagnosis codes, procedures, laboratory thresholds, algorithms or adjudication can identify outcomes. Positive predictive value, sensitivity, severity and timing matter. A valid algorithm in one system or era may not transfer unchanged. Remember: Real-world data record healthcare processes as well as health, and those processes can change over time.

A strong science note separates observation from interpretation. Write the population, exposure, comparator, outcome and time frame for outcome validation protects against code-only assumptions before inspecting results. 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

Demographics, disease severity, comorbidities, prior treatment, healthcare utilisation and calendar time can reduce confounding when measured before exposure. Variables affected by treatment or selected after index can introduce bias.

The scientific value of covariates should capture causes of treatment and outcome appears when evidence is connected to a defined decision. Demographics, disease severity, comorbidities, prior treatment, healthcare utilisation and calendar time can reduce confounding when measured before exposure. Variables affected by treatment or selected after index can introduce bias. Keep this boundary visible: Adjustment controls measured confounders only to the extent that variables are valid and models have support.

A strong science note separates observation from interpretation. Trace every variable in covariates should capture causes of treatment and outcome to its operational source and validation evidence. 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

New-user cohorts, active comparators, restriction, matching, washout and appropriate time zero help align groups. Case-control, self-controlled and other designs answer different questions and rely on different assumptions.

For design choices can control bias before modelling, a useful explanation shows the data, assumptions and operational response together. New-user cohorts, active comparators, restriction, matching, washout and appropriate time zero help align groups. Case-control, self-controlled and other designs answer different questions and rely on different assumptions. Preserve this guardrail: A diagnosis code can be a useful signal without being a perfectly timed or confirmed clinical event.

A strong science note separates observation from interpretation. Challenge design choices can control bias before modelling with a negative control, alternate definition or quantitative bias analysis. 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

Stratification, regression, propensity scores, weighting or matching adjust measured differences. Balance, positivity, extreme weights and model specification are examined. No technique removes unmeasured confounding merely because software produces an estimate.

A trustworthy account of confounding methods require overlap and diagnostics keeps the protocol, implementation and lifecycle record aligned. Stratification, regression, propensity scores, weighting or matching adjust measured differences. Balance, positivity, extreme weights and model specification are examined. No technique removes unmeasured confounding merely because software produces an estimate. Remember: A protocol written after results are known cannot provide the same protection against outcome-driven choices.

A strong science note separates observation from interpretation. Reproduce confounding methods require overlap and diagnostics from versioned data, code, cohort flow and decision logs. 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

Absence can mean no event, no test, care outside the system or failed capture. Complete-case analysis, imputation or missing categories rest on assumptions. Sensitivity analyses show how conclusions change under plausible departures.

The scientific value of missing data need a mechanism and sensitivity plan appears when evidence is connected to a defined decision. Absence can mean no event, no test, care outside the system or failed capture. Complete-case analysis, imputation or missing categories rest on assumptions. Sensitivity analyses show how conclusions change under plausible departures. Keep this boundary visible: Real-world data record healthcare processes as well as health, and those processes can change over time.

A strong science note separates observation from interpretation. Write the population, exposure, comparator, outcome and time frame for missing data need a mechanism and sensitivity plan before inspecting results. 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

Misclassification may be differential or nondifferential and can bias results in several directions. Validation subsets, probabilistic bias analysis, alternative definitions and negative controls can illuminate robustness.

For measurement error can affect exposure, outcome and covariates, a useful explanation shows the data, assumptions and operational response together. Misclassification may be differential or nondifferential and can bias results in several directions. Validation subsets, probabilistic bias analysis, alternative definitions and negative controls can illuminate robustness. Preserve this guardrail: Adjustment controls measured confounders only to the extent that variables are valid and models have support.

A strong science note separates observation from interpretation. Trace every variable in measurement error can affect exposure, outcome and covariates to its operational source and validation evidence. 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

Cohort construction, code lists, models, censoring, subgroups, multiplicity and sensitivity analyses are prespecified. Revisions are timestamped with reasons and whether results were known. Reproducible code and versioned data preserve the actual execution.

A trustworthy account of the analysis plan should precede outcome-driven choices keeps the protocol, implementation and lifecycle record aligned. Cohort construction, code lists, models, censoring, subgroups, multiplicity and sensitivity analyses are prespecified. Revisions are timestamped with reasons and whether results were known. Reproducible code and versioned data preserve the actual execution. Remember: A diagnosis code can be a useful signal without being a perfectly timed or confirmed clinical event.

A strong science note separates observation from interpretation. Challenge the analysis plan should precede outcome-driven choices with a negative control, alternate definition or quantitative bias analysis. 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

Alternative exposure gaps, lag periods, outcome definitions, comparators, confounding methods, negative controls and quantitative bias analyses challenge the primary result. A long list of nearly identical models is less useful than tests of the assumptions most likely to fail.

The scientific value of sensitivity analyses test rival explanations appears when evidence is connected to a defined decision. Alternative exposure gaps, lag periods, outcome definitions, comparators, confounding methods, negative controls and quantitative bias analyses challenge the primary result. A long list of nearly identical models is less useful than tests of the assumptions most likely to fail. Keep this boundary visible: A protocol written after results are known cannot provide the same protection against outcome-driven choices.

A strong science note separates observation from interpretation. Reproduce sensitivity analyses test rival explanations from versioned data, code, cohort flow and decision logs. 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 record includes question, protocol, target population, data sources and dates, provenance, linkage, code lists, exposure and outcome validation, covariates, design, cohort flow, missingness, confounding method, diagnostics, software, code, sensitivity results, limitations, reporting tables and regulatory submission files.

For a complete m14 receipt makes the study auditable, a useful explanation shows the data, assumptions and operational response together. The record includes question, protocol, target population, data sources and dates, provenance, linkage, code lists, exposure and outcome validation, covariates, design, cohort flow, missingness, confounding method, diagnostics, software, code, sensitivity results, limitations, reporting tables and regulatory submission files. Preserve this guardrail: Real-world data record healthcare processes as well as health, and those processes can change over time.

A strong science note separates observation from interpretation. Write the population, exposure, comparator, outcome and time frame for a complete m14 receipt makes the study auditable before inspecting results. 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