FDA Guidance on Missing Data in Clinical Trials

Introduction

Missing data are unavoidable in many clinical trials. Participants may discontinue treatment, miss visits, withdraw consent, die, start rescue medication, move to another therapy, or become unable to complete an assessment.

The statistical problem is not simply that a cell in a database is blank. The more important question is what the missing outcome might have been, why it is missing, and whether the missingness could change the clinical conclusion.

FDA's current statistical perspective places missing data within a broader framework that includes trial design, data collection, estimands, prespecification, assumptions, and sensitivity analysis. FDA's Office of Biostatistics explicitly identifies prevention and treatment of missing data, along with evaluation of statistical assumptions, as part of its review of clinical-trial evidence.

The Core FDA Principle

The central principle is simple: prevent missing data whenever possible, understand why data are missing, prespecify how missingness will be handled, and demonstrate that the primary conclusion is robust to clinically plausible departures from the primary assumptions.

FDA materials repeatedly emphasize that missing data can create uncertainty and bias, that sponsors should make efforts to reduce loss to follow-up, and that the reasons for missingness should be documented. FDA also emphasizes that the statistical analysis plan should address the assumptions underlying the chosen method.

Why Missing Data Can Bias a Trial

Suppose a randomized trial has 200 patients, 100 per treatment arm. The primary endpoint is a continuous measure at Week 24.

If 5 patients in each arm are missing, the missingness may have limited impact if the reasons and outcome trajectories are reasonably balanced. But suppose 20 experimental-arm patients discontinue because they perceive the treatment as ineffective, while only 5 control patients discontinue for the same reason.

Simply analyzing observed Week 24 values can make the experimental arm look better than it would have looked had outcomes been observed for everyone. Conversely, if patients with poor outcomes remain more likely to return for assessment in the experimental arm, complete-case analysis could move the estimate in the opposite direction.

The direction of bias is not determined by the percentage of missing data alone. The relationship between missingness, treatment, observed outcomes, unobserved outcomes, and post-randomization events matters.

Missingness Is More Than a Percentage

A statement such as “10% of Week 24 observations are missing” is incomplete. A useful missing-data assessment asks at least four questions:

• How much data are missing?
• When do observations become missing?
• Why are they missing?
• Are missingness patterns different between treatment groups or important subgroups?

FDA-oriented reporting should therefore distinguish between the numerical amount of missing data and the clinical reasons that generated it.

What Counts as Missing Data?

Missingness can arise in several distinct ways.

Missing Data Versus Intercurrent Events

ICH E9(R1) introduced an essential distinction: failure to collect data should not be confused with an intercurrent event that changes the meaning or availability of the outcome.

For example, treatment discontinuation can lead to two very different situations:

1. The patient stops treatment but continues outcome assessments.
2. The patient stops treatment and consequently provides no later outcome assessments.

The first is primarily an intercurrent-event question. The second contains both an intercurrent event and a missing-data problem.

Estimands: The Missing-Data Question Starts With the Scientific Question

An estimand defines what treatment effect the trial is trying to estimate. ICH E9(R1) describes five attributes: treatment, population, variable, intercurrent-event strategy, and population-level summary.

This matters because missing-data methods cannot be selected intelligently without first knowing the target treatment effect.

For example, consider treatment discontinuation because of an adverse event. Two legitimate scientific questions might be:

• Treatment-policy question: What is the effect of assignment to treatment regardless of whether patients discontinue treatment?
• While-on-treatment question: What is the effect while patients remain on the assigned treatment?

These are different estimands. The data needed and the missing-data assumptions can therefore differ.

A Simple Estimand Example

Suppose the endpoint is change from baseline in systolic blood pressure at Week 24.

A treatment-policy estimand might target:

where outcomes after treatment discontinuation remain relevant to the question.

A hypothetical estimand could instead ask what the Week 24 difference would have been if all patients had remained on their assigned treatment.

The numerical answer to these questions can differ even when they use the same observed trial data.

MCAR, MAR, and MNAR

Missing-data mechanisms are often described using three shorthand concepts: MCAR, MAR, and MNAR.

These are assumptions about a data-generating process, not labels that can usually be proven from the observed dataset.

where \(R\) is the missingness indicator, \(Y_{\mathrm{obs}}\) represents observed outcomes, and \(Y_{\mathrm{mis}}\) represents unobserved outcomes.

MCAR: Missing Completely at Random

Under MCAR, missingness is unrelated to the outcome information after accounting for the trial design.

For example, imagine that a laboratory shipment is accidentally lost for a randomly selected group of patients, with no relationship to treatment, disease severity, or outcome.

That scenario is more compatible with MCAR than a pattern in which patients with worsening disease are more likely to miss visits.

MCAR is mathematically convenient but often implausible for clinical trial dropout.

MAR: Missing at Random

Under MAR, missingness can depend on information that has already been observed.

For example, patients with poor Week 8 outcomes may be more likely to discontinue before Week 24, but if the relationship between discontinuation and the missing Week 24 outcome is adequately explained by observed information such as Week 8 outcome, prior visits, baseline disease severity, and treatment, an MAR analysis may be plausible.

Conceptually:

The assumption is not “the data are random.” It is a conditional assumption about the relationship between missingness and the unobserved outcome.

MNAR: Missing Not at Random

MNAR is often used to describe settings in which the probability of missingness depends on the unobserved outcome even after conditioning on observed data.

For example, suppose two patients have identical observed histories, but the patient with the worse unobserved Week 24 outcome is more likely to withdraw. That dependence is not captured by a simple MAR assumption.

Visualizing the Logic of Missingness Assumptions

Why Complete-Case Analysis Is Usually Not Enough

Complete-case analysis removes participants with missing endpoint data. That can be attractive because it is simple and transparent.

However, the patients who remain may differ systematically from those who are excluded.

In a randomized trial, randomization protects the treatment comparison at baseline. It does not guarantee that the subset of completers remains comparable after post-randomization dropout.

Therefore, excluding incomplete patients can change the population being analyzed and can introduce bias.

Why Last Observation Carried Forward Is Problematic

Last observation carried forward (LOCF) replaces a missing later measurement with the patient's last observed value.

For example:

LOCF would set:

The problem is that this imposes a strong and usually scientifically unjustified assumption: the patient's unobserved outcome would have remained at the last observed value.

FDA review materials have specifically noted that LOCF is not a generally preferred modern strategy for primary analyses, and that the assumptions behind the chosen method should be explained.

Likelihood-Based Repeated-Measures Models

For continuous longitudinal endpoints, mixed models for repeated measures (MMRM) are frequently used because they can incorporate observed outcomes from participants with incomplete follow-up without requiring a single imputed value for every missing visit.

A simplified model can be written as:

where \(Y_{ij}\) is the observed outcome for patient \(i\) at visit \(j\), \(A_i\) is treatment, \(V_j\) is visit, \(X_i\) represents covariates, and the residual covariance structure accounts for within-patient correlation.

Under a standard MAR interpretation, likelihood-based estimation uses the observed history to contribute information about the treatment effect.

Multiple Imputation

Multiple imputation replaces each missing value with multiple plausible values rather than a single deterministic substitute.

For \(M\) imputed datasets, the analysis produces estimates:

These are combined using Rubin's rules. The pooled point estimate is:

and total variance combines within-imputation and between-imputation uncertainty.

where \(\bar U\) is the average within-imputation variance and \(B\) is the between-imputation variance.

What Should Go Into an Imputation Model?

An imputation model should generally contain information that predicts the outcome and/or missingness, subject to the estimand and analysis strategy. Potential variables include:

• Treatment assignment
• Baseline prognostic factors
• Observed post-baseline outcomes
• Prior treatment response
• Important disease characteristics
• Reasons for treatment discontinuation
• Other variables associated with missingness

The model should be developed before looking at the final treatment result when possible, and it should be documented in the SAP.

Pattern-Mixture Models

Pattern-mixture models stratify or model the outcome distribution according to missingness patterns.

For example, patients might be grouped as:

• Complete through Week 24
• Missing after Week 16
• Missing after Week 12
• Missing after Week 8

The observed data inform each pattern, while assumptions are needed to relate unobserved outcomes to observed outcomes.

Pattern-mixture approaches are especially useful for sensitivity analysis because they make departures from MAR explicit.

Delta Adjustment

A common sensitivity-analysis approach is to modify imputed values by a clinically justified shift:

For an endpoint where larger values are worse, a positive \(\delta\) might represent a worse-than-MAR assumption for missing patients.

For an endpoint where larger values are better, the direction would need to be interpreted accordingly.

The value of \(\delta\) should not be arbitrary. It should be motivated by clinical knowledge, observed outcome differences, historical evidence, or other credible information.

Tipping-Point Analysis

A tipping-point analysis asks: How unfavorable would the missing outcomes need to be before the clinical conclusion changes?

For example, suppose the primary analysis gives:

The sensitivity analysis might impose increasingly unfavorable assumptions for missing experimental-arm outcomes.

The exact numerical threshold depends on the endpoint, estimator, covariance structure, and inferential criterion.

Reference-Based Imputation

Reference-based methods use the observed behavior of a reference group to construct assumptions about outcomes after certain intercurrent events. Examples include jump-to-reference and copy-reference strategies.

These methods can be useful when the scientific question implies that after an event such as treatment discontinuation, the future trajectory may resemble a control or reference trajectory.

The choice must be justified in the context of the estimand and clinical setting rather than treated as a generic “conservative” option.

Control-Based Imputation Is Not Automatically Conservative

It is tempting to label any method that moves missing experimental-arm outcomes toward control as conservative.

But conservatism depends on the endpoint, treatment effect definition, direction of benefit, and estimand.

A sensitivity assumption should therefore be described in terms of the scientific behavior it represents, not merely as “conservative” or “nonconservative.”

Missing Binary Endpoints

For a binary endpoint, missingness often appears as an incomplete response status.

Suppose the objective response rate is:

If patients with missing response assessments are excluded from the denominator, the apparent response rate can increase substantially.

A simple worst-case sensitivity analysis might classify missing experimental patients as nonresponders and missing control patients as responders.

That is only one sensitivity analysis, however. More realistic methods may model response probabilities using observed covariates and response history.

Missing Time-to-Event Endpoints

Time-to-event endpoints require special care because censoring is not the same thing as ordinary missingness.

For progression-free survival, for example, a patient who has no documented progression before the last adequate assessment may be censored according to the prespecified event and censoring rules.

A patient who withdraws consent may have a different censoring implication than a patient who remains in follow-up but simply has no event by study end.

Missing Safety Data

Missing safety data create a different set of concerns from missing efficacy data.

A participant who stops treatment early may have fewer opportunities to experience and report later adverse events. The resulting exposure-time differences can affect interpretation of safety summaries.

Safety analyses therefore require careful consideration of:

• Exposure duration
• Early treatment discontinuation
• Follow-up duration
• Serious adverse events after treatment discontinuation
• Deaths and their ascertainment
• Laboratory and vital-sign assessment schedules

The missing-data framework should not be reduced to an imputation exercise. For many safety endpoints, the more important goal is complete ascertainment and transparent accounting of follow-up.

Preventing Missing Data Begins in the Protocol

The strongest missing-data strategy is often operational rather than statistical.

Protocol planning can include:

• Flexible visit windows
• Remote assessments where scientifically appropriate
• Transportation assistance
• Reminder systems
• Retention plans
• Clear instructions for missed visits
• Collection of outcome data after treatment discontinuation when relevant
• Training sites to document reasons for missingness
• Procedures for contacting participants who miss visits

The goal is not simply to increase the number of completed visits. The goal is to preserve the ability of the trial to answer its intended scientific question.

Collect the Reason for Missingness

“Lost to follow-up” is often too vague to be analytically useful.

A more informative classification might distinguish:

These categories help investigators understand whether the observed missingness pattern is clinically informative.

Missing Data Should Be Summarized by Treatment Arm

A useful clinical-trial missing-data table should show counts and percentages by treatment arm, and often by visit.

The table suggests increasing attrition and a potentially important difference between treatment groups. The numbers alone do not prove bias, but they should trigger review of reasons and patterns.

Patterns by Subgroup

Missingness can also differ by:

• Age group
• Disease severity
• Geographic region
• Site
• Baseline prognostic factors
• Prior treatment history
• Biomarker subgroup

This matters because a missing-data method can perform differently if the missingness process differs substantially across clinically important subgroups.

Observed Versus Missing Outcome Distributions

One useful exploratory strategy is to compare the observed outcome history of patients who later become missing with that of patients who remain observed.

For example, compare Week 8 outcomes between:

• Patients with a Week 24 measurement
• Patients without a Week 24 measurement

If the missing group already has markedly worse Week 8 outcomes, an assumption that their future outcomes are identical to those of completers may be questionable.

Missing Data and the Treatment Effect

Suppose the treatment effect is defined as:

If missing values are systematically different from observed values, the estimate based only on observed outcomes can be written conceptually as:

while the estimand may require:

The missing-data problem is therefore the gap between what is observed and what is required to estimate the target population-level quantity.

Primary Analysis and Sensitivity Analysis Are Different Jobs

The primary analysis should estimate the prespecified estimand using a method supported by clinically plausible assumptions.

Sensitivity analyses ask whether the conclusion survives plausible alternative assumptions.

Do Not Choose Sensitivity Analyses After Seeing the Result

If a sponsor chooses only sensitivity analyses that preserve statistical significance, the exercise is not a meaningful robustness assessment.

The SAP should therefore describe:

• Which assumptions will be examined
• Which sensitivity estimators will be used
• How missing outcomes will be shifted or modeled
• Which covariates will be included
• How intercurrent events will be treated
• How conclusions will be summarized

A Practical Sensitivity-Analysis Grid

What Is a Good Sensitivity Analysis?

A good sensitivity analysis should be:

• Clinically plausible: it represents a scenario that could reasonably occur.
• Transparent: the assumptions can be explained to a reviewer.
• Relevant: it addresses a weakness in the primary assumption.
• Prespecified: it is not selected solely because of the observed result.
• Interpretable: the reader can understand what departure from the primary assumption it represents.

Example: A Continuous Endpoint

Suppose a randomized trial has:

• 200 patients
• 100 per arm
• Week 24 continuous endpoint
• 15% missing in the experimental arm
• 8% missing in control

The primary analysis estimates:

Suppose the 95% confidence interval is:

A sensitivity analysis might progressively worsen the assumed outcomes for missing experimental-arm patients until the confidence interval includes zero.

If this requires an implausibly large departure from the observed data, confidence in robustness is strengthened.

Example: Binary Response Endpoint

Suppose:

An observed-case analysis could be:

But if missing patients differ systematically, these percentages may not represent the intended estimand.

A sensitivity analysis can examine assumptions ranging from all missing experimental patients being responders to all being nonresponders, along with more clinically realistic model-based assumptions.

Handling Treatment Discontinuation

Treatment discontinuation is one of the most important sources of missing efficacy data.

A strong analysis plan distinguishes:

• Why the patient discontinued
• Whether outcome assessments continued
• Whether rescue medication was initiated
• Whether the estimand is treatment-policy, hypothetical, while-on-treatment, or another strategy

For example, if the estimand is treatment policy, continued outcome collection after treatment discontinuation may be particularly important. If the estimand is while-on-treatment, the analysis may intentionally focus on outcomes before a defined intercurrent event.

Withdrawal of Consent

Withdrawal of consent creates an ethical and operational boundary. Once a participant withdraws consent for further research participation, the sponsor cannot simply assume that additional outcome data will be available.

The protocol and informed-consent process should therefore anticipate the possibility of early withdrawal and define what information may still be collected where legally and ethically permitted.

Death Is Not Just Another Missing Value

If a patient dies before a planned endpoint, treating the endpoint as a generic missing continuous value can be scientifically inappropriate.

Death may be an intercurrent event that changes the meaning of the endpoint. The correct strategy depends on the estimand and disease context.

For example, in some settings death may be incorporated as a failure, while in others the scientific question may require a composite endpoint or a competing-risk framework.

Rescue Medication

Rescue medication can create a similar issue.

Suppose a patient receives rescue therapy because the assigned treatment is not controlling the disease.

The post-rescue outcome is not merely “missing” if the rescue therapy changes what the outcome means.

The analysis should first define the estimand strategy for rescue treatment and then address any missing observations created by the event.

Estimand Strategy and Missing Data Should Agree

FDA-Oriented Data Review

A regulatory reviewer is likely to ask questions such as:

• How much data are missing?
• Is missingness balanced across treatment groups?
• When does missingness occur?
• Why are patients missing?
• Are missingness reasons related to treatment efficacy or safety?
• What outcome information was collected after treatment discontinuation?
• What assumptions support the primary analysis?
• What happens if those assumptions are violated?
• Does the treatment conclusion remain consistent?

Missing Data in the Statistical Analysis Plan

A strong SAP section should define, at minimum:

• The estimand for the key endpoint
• Intercurrent-event strategies
• The primary estimator
• The analysis population
• Baseline and covariate definitions
• The primary missing-data assumption
• How missing observations contribute to the primary analysis
• How missingness reasons will be summarized
• Prespecified sensitivity analyses
• Any imputation model and its variables
• Any MNAR or reference-based assumptions
• Rules for post-rescue or post-discontinuation data

Recommended SAP Structure

1. Estimand
   1.1 Treatment
   1.2 Population
   1.3 Variable
   1.4 Intercurrent-event strategies
   1.5 Population-level summary

2. Primary analysis
   2.1 Analysis population
   2.2 Statistical model
   2.3 Covariates
   2.4 Variance/covariance assumptions
   2.5 Missing-data assumption

3. Missing-data prevention
   3.1 Retention procedures
   3.2 Continued outcome collection
   3.3 Documentation of reasons

4. Sensitivity analyses
   4.1 Alternative model
   4.2 Pattern-mixture analysis
   4.3 Delta adjustment
   4.4 Reference-based analysis
   4.5 Tipping-point analysis

5. Reporting
   5.1 Missing-data summaries
   5.2 Disposition
   5.3 Reasons for missingness
   5.4 Primary and sensitivity results

Data Standards and Analysis Dataset Considerations

In a CDISC-oriented development program, the missing-data strategy should remain traceable from collected observations through analysis datasets and tables, listings, and figures.

Useful analysis variables may include:

The exact variables and derivations depend on the study and analysis datasets. The key requirement is traceability: the reviewer should be able to understand how the final analysis was constructed from the observed clinical-trial data.

R Example: Missingness by Visit

A simple exploratory summary can calculate the percentage of patients without an observed endpoint at each visit.

library(dplyr)
library(ggplot2)

missing_by_visit <- adsl %>%
  group_by(TRT01P, AVISITN, AVISIT) %>%
  summarise(
    N = n(),
    Missing = sum(is.na(AVAL)),
    PctMissing = 100 * Missing / N,
    .groups = "drop"
  )

ggplot(
  missing_by_visit,
  aes(
    x = AVISITN,
    y = PctMissing,
    group = TRT01P,
    color = TRT01P
  )
) +
  geom_line(linewidth = 0.9) +
  geom_point(size = 2) +
  labs(
    x = "Analysis Visit",
    y = "Missing Endpoint (%)",
    color = "Treatment"
  ) +
  theme_classic()

R Example: Multiple Imputation

For a continuous endpoint, a simple multiple-imputation workflow can be implemented with packages such as mice. The actual imputation model must be aligned with the study estimand and SAP.

library(mice)

imp <- mice(
  analysis_data,
  m = 50,
  method = "pmm",
  seed = 20260911
)

fit <- with(
  imp,
  lm(
    CHANGE24 ~ TRT01P + BASELINE + AGE
  )
)

pooled <- pool(fit)

summary(pooled, conf.int = TRUE)

This example illustrates software mechanics only. It should not be copied into a regulated analysis without specifying the estimand, imputation model, auxiliary variables, treatment of intercurrent events, diagnostics, and sensitivity analyses.

R Example: A Simple Delta-Adjustment Sensitivity Analysis

A delta adjustment can be conceptualized by shifting imputed outcomes after a primary MAR imputation.

# Conceptual example only

imputed_data$Y24_MNAR <-
  imputed_data$Y24_MAR + delta

fit_mnar <- lm(
  Y24_MNAR ~ TRT01P + BASELINE,
  data = imputed_data
)

summary(fit_mnar)

In a real analysis, the shift may differ by treatment, missingness pattern, visit, or event history, and the imputation and analysis should be combined using an appropriate multiple-imputation framework.

R Example: Tipping-Point Grid

deltas <- seq(
  -5, 5,
  by = 0.5
)

results <- lapply(
  deltas,
  function(delta) {

    # Re-impute under the specified delta
    # Fit the prespecified analysis model
    # Return the treatment estimate and CI

    data.frame(
      delta = delta,
      estimate = NA_real_,
      lower = NA_real_,
      upper = NA_real_
    )
  }
)

results <- do.call(
  rbind,
  results
)

The missing portion of the example is intentional: the correct code depends on the endpoint, imputation model, estimand, and primary analysis model.

Common Statistical Mistakes

1. Using complete cases without assessing why patients are missing.
2. Using LOCF as a default simply because it is easy to implement.
3. Calling an analysis “MAR” without describing the variables and model supporting the assumption.
4. Calling an MNAR analysis “conservative” without explaining the clinical meaning of the assumption.
5. Ignoring treatment discontinuation when defining the estimand.
6. Confusing an intercurrent event with missing data.
7. Performing sensitivity analyses that are not clinically plausible.
8. Changing the missing-data method after seeing the primary result without justification.
9. Failing to report differential missingness between treatment groups.
10. Failing to collect reasons for missed visits and discontinuations.

Common Regulatory-Review Mistakes

1. Reporting only the percentage missing. The reasons and timing matter.
2. Claiming randomization solves missing-data bias. It does not guarantee random dropout.
3. Presenting one sensitivity analysis as proof of robustness. Robustness is stronger when clinically plausible alternatives have been considered.
4. Ignoring post-treatment data that are relevant to the estimand.
5. Using inconsistent estimands across primary and sensitivity analyses.
6. Failing to explain the clinical interpretation of the missing-data assumption.

Quality-Control Checklist

How to Present Missing Data in a Clinical Study Report

A clear CSR presentation should include both disposition and statistical summaries.

Example CSR Language

When the Primary and Sensitivity Analyses Disagree

Disagreement is not necessarily a failure of the trial.

It is information about the uncertainty created by missing data.

Suppose the primary MAR analysis favors treatment, but a clinically plausible MNAR sensitivity analysis removes the effect.

The appropriate conclusion is not that one analysis is automatically correct. Instead, the report should explain:

• What assumptions differ
• Why the primary assumption was selected
• Why the sensitivity assumption is plausible
• How much missing data are involved
• Whether missingness is differential by treatment
• How the overall evidence should be interpreted

How Much Missing Data Is Too Much?

There is no universal percentage at which missing data suddenly become unacceptable.

A 5% missing-data rate can be problematic if all missing patients are concentrated in one clinically important subgroup for a reason related to treatment efficacy.

Conversely, a larger percentage may be less damaging if the missingness is well understood, balanced, and appropriately addressed under a credible estimand and analysis strategy.

Missing Data and Sample-Size Planning

Missingness should be considered when planning sample size, but simply inflating the sample size does not solve bias.

If the expected completer fraction is \(c\), a simple planning adjustment may be:

For example, if 90% of randomized participants are expected to provide the information required for the primary analysis:

This protects precision against attrition but does not guarantee unbiased inference.

Operational Monitoring of Missing Data

Missing-data prevention should continue during the trial rather than being left to the statistician at database lock.

Useful monitoring metrics include:

• Missed visits per site
• Visit completion rates by treatment
• Time to recovery after missed visits
• Early discontinuation rates
• Reasons for discontinuation
• Assessment completion by region
• Assessment completion by important subgroup

Monitoring should be designed so that operational interventions do not compromise trial integrity or introduce differential treatment of participants.

Missing Data and Blinding

The collection and review of missing-data information should be handled in a way that respects the trial's blinding and operational controls.

For example, site-level retention monitoring can often be performed without revealing treatment assignment.

If unblinded access is required for a particular monitoring purpose, the process should be prespecified and governed appropriately.

What FDA Reviewers May Consider When Evaluating Robustness

A robust package generally connects:

Practical Decision Tree

A useful way to approach a new missing-data problem is:

Three Questions to Ask Before Choosing an Imputation Method

1. What estimand am I estimating?
2. What assumption about the unobserved outcomes am I making?
3. How will I demonstrate robustness if that assumption is wrong?

If those questions cannot be answered clearly, choosing an imputation method is premature.

What Not to Say

Avoid vague statements such as:

• “Missing values were imputed conservatively.”
• “LOCF was used to avoid bias.”
• “The missing data were assumed to be random.”
• “The results were robust because the sensitivity analysis was significant.”

Instead, describe the assumption mathematically or clinically.

Better Reporting Language

For example:

Relationship to ICH E9(R1)

The modern regulatory framework is best understood as an integrated estimand-and-sensitivity-analysis framework.

ICH E9(R1) emphasizes that the estimand defines the treatment effect of interest, while the estimator is the method used to estimate it. Missing data and intercurrent events are therefore not isolated programming problems; they are part of the specification of the scientific question.

FDA materials have emphasized the same general direction: relevant data should be collected where possible, missingness should be minimized, assumptions should be documented, and sensitivity analyses should examine the robustness of conclusions.

FDA, NRC, and ICH: How the Pieces Fit Together

Current Regulatory Perspective

FDA's Office of Biostatistics describes prevention and treatment of missing data as part of the statistical evaluation of clinical trials and also highlights evaluation of assumptions of statistical methods and robustness of trial findings.

Recent FDA materials continue to emphasize prespecification of estimands, analysis approaches, missing-data assumptions, and sensitivity analyses. This means that missing-data planning should be treated as part of trial design rather than a database-lock cleanup exercise.

A Practical Missing-Data Analysis Plan

Final Checklist for a Regulatory Submission

The Most Important Concept

The most important lesson is that missing data are not primarily an imputation problem.

They are a trial-design, clinical, estimand, and statistical-inference problem.

A sophisticated imputation method cannot rescue a trial in which the reasons for missingness were never documented, important post-discontinuation data were not collected, or the scientific estimand was never defined.

Conversely, a trial with carefully designed retention procedures, continued outcome collection, transparent missingness documentation, a coherent primary estimator, and clinically credible sensitivity analyses can provide much stronger evidence even when some data remain missing.

References and Regulatory Resources

U.S. Food and Drug Administration. Office of Biostatistics. Statistical review principles include prevention and treatment of missing data and evaluation of statistical assumptions.

National Research Council. (2010). The Prevention and Treatment of Missing Data in Clinical Trials. National Academies Press. This report was commissioned by FDA and remains a foundational reference for the prevention and treatment of missing data.

ICH. E9(R1) Addendum on Estimands and Sensitivity Analysis in Clinical Trials. The addendum provides the modern framework linking estimands, intercurrent events, estimators, and sensitivity analyses.

ICH. E9 Statistical Principles for Clinical Trials.

FDA. Endpoint- and indication-specific guidances should also be consulted because recommendations for missing data can differ according to the clinical setting, endpoint, and trial design.

Useful official resources:

• FDA Office of Biostatistics
• FDA Guidance Documents
• FDA patient-focused drug development material discussing missing data
• ICH E9(R1) Estimands and Sensitivity Analysis