Beyond Average Effects: A Practical Framework for Tails, Treatment Heterogeneity, and Clinical Decision-Making
M. Iabluchanskyi, A Martynenko, V. Shlyakhover,
G. Raimondi, A. Yabluchanskiy
Abstract
Clinical trials are commonly summarized by a single average treatment effect, expressed as a
mean difference, relative risk, odds ratio, hazard ratio, or median outcome. These measures are
indispensable for population-level inference, but they may not fully describe the clinical reality
of treatment. An apparently favorable average can coexist with profound nonresponse, clinically
important deterioration, severe toxicity, rare catastrophic harm, extreme vulnerability, and, in
some cases, exceptional benefit. Such clinically consequential information may be concentrated
in the lower and upper tails of outcome distributions.
Existing methodological and regulatory guidance already addresses important components of this
problem, including harms reporting, estimands, sensitivity analysis, prespecified subgroup
analysis, enrichment, quality by design, and postmarketing surveillance. This article synthesizes
these principles into a practical Tail-Sensitive Evidence Framework. The framework
recommends prospective identification of clinically meaningful favorable and unfavorable tail
outcomes; measurement of patient characteristics needed to interpret them; hierarchical analysis
that distinguishes confirmatory, supportive, sensitivity, and exploratory questions; systematic
investigation of extreme observations as possible biological, safety, or data-quality signals;
transparent reporting; and explicit postmarketing plans for questions that cannot be resolved
before approval.
Tail-sensitive research creates economic and operational challenges. Broader eligibility, longer
follow-up, richer biological characterization, detailed safety surveillance, and adequately
powered subgroup analyses increase cost and complexity. Yet incomplete evidence has
substantial downstream costs, including inappropriate generalization, preventable harm,
ineffective treatment, delayed recognition of rare adverse events, and repeated corrective studies
after widespread adoption. Artificial intelligence, adaptive designs, Bayesian and hierarchical
methods, real-world data, and postmarketing surveillance can improve efficiency and direct
limited resources toward the questions that matter most. They cannot replace randomized
evidence where causal inference is required, high-quality measurement, clinical judgment, patient
protection, or regulatory accountability.
Introduction
Randomized controlled trials remain the principal method for determining whether an
intervention provides meaningful benefit with acceptable harm. Their findings guide regulatory
approval, clinical guidelines, reimbursement decisions, prescribing, and patient choice. Trial
results are usually communicated through a small set of summary measures: a mean difference,
relative risk, odds ratio, hazard ratio, median survival, or p-value.
These estimates are essential. They provide concise summaries of comparisons between
randomized groups and support evidence synthesis. Yet they can create an impression of
uniformity that does not correspond to real clinical populations.
A favorable average effect may coexist with a clinically meaningful minority of patients who
obtain no benefit, deteriorate, discontinue treatment because of intolerance, or experience rare
severe harm. Conversely, a modest average effect may conceal a smaller group with unusually
large, durable, and clinically meaningful benefit. Such information may be found in the full
outcome distribution and, particularly, in its lower and upper tails.
This is not an argument that modern trials ignore safety, heterogeneity, or extreme observations.
High-quality trials commonly use independent data-monitoring committees, prespecified
Statistical Analysis Plans, adjudicated outcomes, systematic safety reporting, sensitivity
analyses, subgroup analyses, risk-based monitoring, and postmarketing follow-up. Existing
guidance also addresses harms reporting, trial quality, estimands, enrichment, missing data, and
transparent reporting.
The remaining problem is integration and visibility. In many reports, the average primary effect
dominates the publication and decision narrative, whereas clinically important deterioration, rare
harms, extreme outcomes, site-specific anomalies, and uncertainty in vulnerable groups are
treated as secondary matters. They may be underpowered, incompletely displayed, difficult to
locate, or deferred without an explicit plan for further evaluation.
This article does not introduce a new statistical estimator. It synthesizes existing clinical,
methodological, regulatory, quality-management, and computational principles into a practical
Tail-Sensitive Evidence Framework. Its purpose is to make the planning, analysis, interpretation,
reporting, and postmarketing follow-up of clinically decisive tails more coherent, transparent,
and decision-relevant.
The Average Patient Is a Useful FictionThe average treatment effect is a foundational achievement of clinical epidemiology.
Randomization permits fairer comparison between groups, while concise summary statistics
permit evidence synthesis and clinical decision-making.
But the average patient is a statistical abstraction, not a real person.
Imagine a therapy that improves a symptom score by an average of four points compared with
placebo. The same mean difference may arise if:
• Almost every participant improves by approximately four points.
• Half improve substantially while half do not improve at all.
• Most participants improve slightly while a small group improves dramatically.
• Some participants benefit while others deteriorate or experience severe toxicity.
• Benefit is largely concentrated in a biomarker-defined subgroup.
• Average benefit in robust participants is counterbalanced by falls, hypotension, renal
injury, or treatment discontinuation in frail older adults.
These situations may yield similar average effects but have fundamentally different implications
for prescribing, dose selection, treatment monitoring, informed consent, guideline development,
reimbursement, and product labeling.
The clinical question is therefore not only whether treatment changes the average outcome. It is
also whether treatment changes the probability of meaningful benefit, nonresponse, deterioration,
severe harm, and exceptional response.
Why Tails Matter
The center of an outcome distribution describes a typical or average pattern. The clinical value of
treatment, however, may be determined by what occurs at its margins.
The unfavorable tail may include death, irreversible injury, severe toxicity, major bleeding,
serious infection, severe hypoglycemia, syncope, falls, delirium, acute kidney injury, treatment
discontinuation, complete therapeutic failure, or clinically important loss of function and quality
of life.
The favorable tail may include complete remission, major recovery of function, profound
symptom relief, unusually durable disease control, exceptional survival, or a response pattern
suggesting treatment-sensitive biology.
The lower tail may reveal that a favorable average comes at an unacceptable cost for a
meaningful minority. The upper tail may provide a pathway toward biological discovery and
more precise treatment. Neither should be ignored simply because it concerns fewer participants.
Patients differ across interacting genetic, physiological, metabolic, immunological,
developmental, clinical, psychological, social, and time-dependent spectra. These differences
influence disease trajectory, drug exposure, treatment response, adverse effects, adherence,resilience, and long-term outcome. Randomization balances such characteristics between
treatment groups on average; it does not eliminate their clinical importance.
This is especially important in aging and multimorbidity. Two people of the same chronological
age may differ greatly in frailty, body composition, renal function, cardiovascular reserve,
cognition, inflammatory state, medication burden, social support, and capacity to tolerate harm.
A trial’s average result in “older adults” may therefore be statistically valid but clinically
incomplete.
Tail Analysis Is Not Outlier Hunting
Tail-sensitive analysis is not a call to discard inconvenient data, celebrate anecdotal success, or
conduct unlimited post-hoc searches for favorable subgroups.
It is a structured assessment of clinically meaningful extremes and of how treatment changes the
full distribution of outcomes.
An exceptional outcome in one participant does not prove an exceptional individual causal
effect. A patient who achieves complete remission on therapy may indeed have benefited, but the
observed outcome may also reflect favorable disease biology, baseline prognosis, concurrent
treatment, measurement variation, regression to the mean, or an outcome that might have
occurred under the comparator.
Similarly, a severe event in a treated participant does not automatically establish that treatment
caused it.
These limitations are reasons for disciplined investigation, not dismissal. Clinically important
extreme observations should be evaluated for:
• Technical validity and completeness
• Reliability and timing of measurement
• Treatment dose, exposure, adherence, and concomitant therapy
• Biological plausibility
• Occurrence in the comparator group
• Clustering by site, investigator, region, calendar time, treatment batch, or device
• Association with age, frailty, organ dysfunction, genotype, biomarkers, baseline risk, or
comorbidity
• Robustness in sensitivity analyses
• Replication in other trials, registries, or pooled individual-participant datasets
Tail analysis is therefore an extension of clinical reasoning and causal inference, not a substitute
for either.
Tails as a Test of Study Quality
Tails do not only describe unusual patients. They also test study quality.An extreme laboratory value may represent severe toxicity or unusual drug metabolism. It may
also indicate assay failure, incorrect units, sample misidentification, or data-entry error. A cluster
of serious events at one center may reveal a genuine local risk factor, but it may also reflect
inconsistent outcome ascertainment, protocol nonadherence, poor follow-up, or data-quality
failure.
A study that cannot account for its extreme observations cannot fully claim to understand its own
results.
Tail pattern Extreme
laboratory
Possible clinical
interpretation
Severe toxicity, altered
metabolism, unexpected
Possible quality
Required response
concern
Assay failure, wrong
units, specimen error
Source verification,
laboratory review, repeat
result
Cluster of
serious events
pharmacodynamics
Local vulnerability,
environmental exposure,
Protocol
nonadherence,
testing where feasible
Site audit, central review,
cross-site comparison
at one site
Excess harm in
a subgroup
distinct practice pattern
Susceptibility due to
frailty, organ
inconsistent
Differential
ascertainment or co-
Review baseline risk,
exposure, follow-up, and
Exceptional
response at one
dysfunction, genotype,
Genuine phenotype or
unusually effective care
intervention imbalance
Selection bias,
unblinding, assessment
interaction estimates
Blinded review, source
verification, site-level
center
Heavy
unfavorable tail
context
Real harm in a
vulnerable minority
bias
Missing-event bias,
incomplete follow-up,
analysis
Report tail probabilities,
missing-data patterns,
despite
Implausibly
narrow
Truly uniform response,
if real
inappropriate
Rounding, copy-
forward, selective
discontinuation, sensitivity
Examine raw data, digit
patterns, variance, and
distribution
reporting, data
source records
A Tail and Quality Plan should be included in relevant trial protocols. It should specify how
extreme values will be verified, adjudicated, interpreted, and reported, and when observed
patterns require independent monitoring or regulatory notification.
Illustrative Cases
ACCORD: Benefit, Harm, and Retrospective Explanation
The Action to Control Cardiovascular Risk in Diabetes trial compared intensive with standard
glycemic management in adults with type 2 diabetes at high cardiovascular risk. Intensive
treatment lowered glycated hemoglobin more aggressively but was stopped early because all-
cause mortality was higher in the intensive group: 257 deaths compared with 203 deaths in the
standard group. The primary report estimated a hazard ratio of 1.22, with a 95% confidence
interval of 1.01 to 1.46.
The clinical lesson is not that glycemic control is unimportant. It is that improvement in a central
biomarker cannot stand as a complete representation of clinical value. Mortality, severe
hypoglycemia, treatment complexity, patient vulnerability, and the possibility of differential
benefit and harm also require attention. Long-term follow-up found fewer nonfatal myocardial
infarctions with intensive treatment but continued evidence of increased mortality, demonstrating
competing clinical effects not captured by glycated hemoglobin alone.Subgroup analyses did not provide a simple explanation for the mortality signal. Most baseline
subgroups showed no convincing differential effect. Exploratory interactions involving self-
reported neuropathy, high baseline glycated hemoglobin, and aspirin use were reported, but the
investigators explicitly cautioned that among 38 examined variables, approximately 1.9 findings
with a p-value below 0.05 would be expected by chance alone. These results were therefore
suggestive rather than confirmatory.
This is a direct illustration of the multiplicity problem: a serious tail signal may demand
explanation, but searching many subgroups can produce plausible-looking explanations that are
not reproducible.
A later post-hoc gradient-forest analysis used a different analytic dataset and modeling approach.
It reported an estimated absolute mortality reduction of 2.3% in a low-risk subgroup defined by
hemoglobin glycation index below 0.44, body mass index below 30 kg/m², and age below 61
years; this subgroup represented 8.6% of participants. It also identified a higher-risk subgroup
with hemoglobin glycation index of at least 0.44, representing 24.7% of participants, with an
estimated absolute mortality increase of 3.7% and a number needed to harm of 27.
These model-based findings are important but not definitive. They differ from the original trial’s
primary analysis because they use a different analytic approach and follow-up structure. They
were generated post hoc from one trial and internally validated, but they require external
validation before they can guide individualized treatment. Their value is educational:
multivariable methods may reveal patterns that simple subgroup comparisons miss, yet they do
not remove the need for replication, causal caution, and prospective confirmation.
Oncology: Exceptional Responders
In oncology, a small number of patients may experience deep or unusually durable responses
even when the overall response rate is low. These “exceptional responders” may provide clues
about tumor genomics, host immunity, drug exposure, disease biology, and mechanisms of
sensitivity or resistance.
The U.S. National Cancer Institute’s Exceptional Responders Initiative examined the feasibility
of identifying such patients retrospectively and characterizing molecular features associated with
their responses. Of 520 cases initially proposed to the initiative, 476 were accepted for further
review, and 222 met the criteria for exceptional response. Molecular analysis of available
pretreatment tumor material was feasible in most eligible cases, illustrating the potential of
systematic investigation of exceptional outcomes.
The favorable tail should not be treated as proof that treatment works in a presumed molecular
subgroup. Exceptional response may reflect disease biology, treatment timing, co-interventions,
prognosis, or chance. Its value lies in structured hypothesis generation.
A tail-sensitive approach converts unusual responses into a prospective research agenda:
standardized outcome definitions, tissue collection, genomic and immune profiling, assessment
of drug exposure, appropriate comparator information, and subsequent biomarker-stratifiedstudies.
Exceptional outcomes are not anecdotal decorations; they are starting points for
disciplined discovery.
Rofecoxib: The Temporal Tail
Rofecoxib illustrates why tail-sensitive evidence must include time, cumulative exposure, and
long-term benefit–risk assessment. The Adenomatous Polyp Prevention on Vioxx trial evaluated
rofecoxib for prevention of recurrent colorectal adenomas. The study was terminated early after
an excess of adjudicated thrombotic cardiovascular events emerged in participants receiving
rofecoxib.
In the final analysis, 46 participants in the rofecoxib group experienced a confirmed thrombotic
event during 3,059 patient-years of follow-up, compared with 26 participants in the placebo
group during 3,327 patient-years. The event rates were 1.50 and 0.78 per 100 patient-years,
respectively, corresponding to a relative risk of 1.92, with a 95% confidence interval of 1.19 to
3.11. The risk increase was not evident during the first 18 months of treatment; it emerged only
after longer exposure.
This is not simply a patient-level subgroup problem. It is a temporal tail: a serious adverse effect
that became visible only after cumulative exposure and sufficient duration of observation.
The case does not imply that every serious safety problem can be predicted before approval or
that one trial can eliminate all uncertainty. It does demonstrate how the Tail-Sensitive Evidence
Framework could improve planning. For an intervention with plausible cardiovascular risk, the
protocol should include prospective analysis of time patterns, cumulative incidence, exposure
duration, and late-emerging events; independent adjudication; periodic safety review; and an
explicit plan for continued surveillance beyond the trial.
The rofecoxib case also demonstrates the regulatory importance of time-dependent harm. The
product was voluntarily withdrawn in 2004 after interim safety findings, and subsequent FDA
analyses concluded that selective COX-2 inhibitors were associated with increased risk of
serious cardiovascular events compared with placebo. Regulatory actions emphasized
cardiovascular and gastrointestinal safety information in prescription NSAID labeling.
The absolute risk difference was modest in numerical terms—approximately 0.72 events per 100
patient-years—but the events were serious, potentially irreversible, and more clearly associated
with longer exposure. Frequency, severity, reversibility, exposure duration, and availability of
safer alternatives all matter in benefit–risk judgment.
The Price of Looking More Closely
Tail-sensitive analysis increases clinical visibility, but it also expands inferential opportunity.
Every additional threshold, quantile, subgroup, biomarker, interaction, and machine-learning
search increases the risk of false-positive findings.Prespecification is necessary but not sufficient.
A protocol can prespecify many weakly justified
analyses and still create a substantial false-discovery problem.
Confirmatory inference should therefore be limited to a small number of clinically decisive tail
outcomes and biologically credible effect modifiers. These analyses should be placed within a
formal multiplicity strategy, such as hierarchical testing, gatekeeping, allocation of type-I error,
or a prespecified Bayesian decision rule.
All other analyses should be classified as supportive, sensitivity-based, or exploratory.
Sparse-event and small-subgroup estimates require particular restraint. Reports should emphasize
absolute risks, risk differences, effect estimates, uncertainty intervals, clinical consequences, and
reproducibility rather than p-values alone. When many predictors are examined, regularization,
shrinkage, partial pooling, internal validation, external validation, and replication are essential.
A Bayesian tail decision rule, for example, may specify in advance that a safety review is
triggered when the posterior probability that treatment increases a clinically important severe-
harm risk by more than a predefined margin exceeds a stated threshold. Such a rule does not
eliminate uncertainty or bias; it makes the decision criterion explicit and permits incorporation of
justified prior information.
Tail-sensitive research must not create new false certainties while attempting to reveal hidden
ones.
Operationalizing Tail-Sensitive Analysis
Tail-sensitive analysis must be planned before recruitment and before database lock. It should be
included in the protocol and Statistical Analysis Plan as a dedicated module.
Define Clinically Meaningful Tails
The protocol should identify favorable and unfavorable outcomes that could alter a clinical
decision.
Favorable outcomes may include:
• Complete remission
• Major functional recovery
• Sustained disease control
• Durable survival
• Major improvement in a validated patient-reported outcome
• Large reduction in hospitalization, disability, or care dependence
Unfavorable outcomes may include:
• Death
• Irreversible organ injury• Severe or life-threatening toxicity
• Serious infection
• Major bleeding
• Severe hypoglycemia
• Syncope, falls, delirium, or acute kidney injury
• Treatment discontinuation due to intolerance
• Major deterioration in symptoms, function, or quality of life
These thresholds should be clinically justified and selected before unblinding.
Select Measures Appropriate to the Outcome
Outcome
type
Continuous
outcomes
Potential tail-sensitive measures
Group-specific distributions, cumulative distribution curves, clinically defined
threshold probabilities, selected quantiles, quantile treatment effects
Ordinal
outcomes
Binary
outcomes
Time-to-
event
Category-specific probabilities, shift analyses, severe deterioration and major-
improvement probabilities, justified ordinal models
Absolute event rates, risk differences, relative risks, responder and deterioration
rates, number needed to treat or harm when interpretable
Survival or cumulative-incidence curves, event probabilities at clinically relevant
time points, early and late event patterns, competing-risk analyses
outcomes
Recurrent
outcomes
Safety
outcomes
Longitudinal
outcomes
Total event burden, recurrent-event rates, time to first and subsequent events,
participant-level trajectories
Serious-event incidence, severity, timing, reversibility, exposure relation,
treatment discontinuation, events of special interest
Individual trajectories, sustained response and deterioration, relapse and recovery
patterns, time-varying exposure and risk
No universal rule should require the same percentiles in every trial. In large datasets, lower and
upper quantiles may be informative. In modest samples, extreme percentiles may be unstable.
For rare harms, direct reporting of event counts, denominators, exact uncertainty intervals,
clinically defined thresholds, and time patterns is often more informative.
Minimum Tail-Sensitive Reporting Standards
The following standards are intended as a proportionate minimum. They should be adapted to
disease severity, endpoint type, treatment duration, sample size, expected event rate, and the
clinical consequences of benefit and harm.
Require
ment
Clinical-
tail
All confirmatory trials When clinically relevant and feasible
Prespecify clinically meaningful
deterioration and major safety
Prespecify profound benefit, remission, or
exceptional durability
definition
Central
estimate
Absolute
effects
outcomes
Report primary effect estimate with
uncertainty interval
Report absolute event rates and risk
differences for major benefits and
Include adjusted and unadjusted estimates
where appropriate
Report number needed to treat or harm when
robust and interpretable
harmsDistributi
onal
Provide group-specific distributional
information for primary continuous
Provide cumulative distribution curves,
selected quantiles, or quantile treatment
informati
Threshol
d
or ordinal outcomes
Report prespecified rates of
meaningful benefit and clinically
effects
Examine alternative thresholds as clearly
labeled sensitivity analyses
outcomes
Safety important deterioration
Report serious adverse events,
deaths, discontinuations, and events
Report timing, severity, reversibility, dose
relation, and exposure-adjusted rates
Heteroge
neity
of special interest by group
Report prespecified subgroup effects
with interaction estimates and
Use risk-based, biomarker-based, or causal-
machine-learning methods as exploratory
Data
quality
uncertainty intervals
Verify clinically important extreme
observations and report major
approaches requiring validation
Report site-level review where extreme-event
clustering or data irregularities occur
Uncertain
ty
missing-data patterns
State which tail-related risks or
subgroup effects could not be
Specify the evidence-generation plan for
unresolved questions
estimated reliably
This template does not replace CONSORT, CONSORT Harms, ICH guidance, or regulatory
requirements. It organizes relevant elements around a practical question: what does treatment do
at the margins of patient experience?
Making Tail-Sensitive Trials Affordable
Tail-sensitive research should not mean collecting every possible data element from every
participant. That approach can increase cost, burden, missingness, and analytic noise. The
objective is targeted, decision-relevant measurement.
Several strategies can improve feasibility.
First, use a staged evidence program. Early studies should clarify dose, exposure,
pharmacokinetics, pharmacodynamics, biomarkers, foreseeable harms, and major sources of
variability. Pivotal trials should answer a focused causal question and evaluate a limited number
of clinically decisive tail outcomes. Postmarketing surveillance, registries, pragmatic trials, and
real-world evidence can address rare harms, long-term effects, and complex populations.
Second, use practical, prognostic, or predictive enrichment only when supported by credible
biological and clinical rationale. Enrichment can reduce irrelevant variability, increase event
rates, or focus study on patients more likely to respond, but it should not become a way to avoid
studying populations likely to receive treatment.
Third, consider adaptive enrichment designs when the target population is uncertain but
biologically plausible. Prospectively defined interim rules can permit broad initial learning while
allowing later concentration of enrollment in groups with credible evidence of benefit or harm.
Such designs require strict planning, error control, and operational integrity.
Fourth, use Bayesian and hierarchical models where appropriate. Partial pooling across related
subgroups can stabilize sparse estimates and reduce exaggerated effects, provided that modeling
assumptions are transparent and clinically defensible. Bayesian approaches can incorporate
relevant prior evidence, but they do not substitute for sufficient data or external validation.Fifth,
link trials to high-quality routine clinical data, registries, claims data, and mortality
databases for long-term follow-up and rare-event surveillance. Such linkage can reduce the cost
of extended observation, but data must be fit for purpose and causal analyses require appropriate
comparators and design.
Sixth, use AI-supported tools to reduce operational waste: eligibility prescreening, recruitment
planning, automated data extraction, detection of missingness and inconsistency, safety-signal
prioritization, and site-level anomaly detection.
AI for Tail Detection, Not Automated Clinical Judgment
AI is relevant to tail-sensitive research because tails often reflect complex combinations of age,
frailty, organ function, genotype, biomarker status, drug exposure, concomitant medication,
disease stage, and time.
Potential uses include:
• Identifying candidate participants from electronic health records
• Detecting inconsistent, missing, or implausible data
• Flagging clusters of adverse events or laboratory abnormalities
• Prioritizing safety signals for expert review
• Estimating conditional average treatment effects using causal forests, meta-learners,
Bayesian additive regression trees, and related methods
• Identifying hypotheses for prospective testing and external validation
• Supporting postmarketing surveillance and long-term follow-up
Causal forests estimate heterogeneous treatment effects by aggregating many causal trees,
whereas meta-learners offer related approaches for modeling conditional average treatment
effects in complex data.
But these methods have strict limits. They estimate risk-stratified or conditional average
treatment effects under stated assumptions. They do not observe an individual patient’s
unobserved counterfactual outcome. They cannot by themselves prove an individual treatment
effect, correct poor-quality data, eliminate unmeasured confounding, or replace randomized
evidence. The methodological challenges of individualized treatment-effect prediction include
validation, calibration, transportability, and causal assumptions.
AI can detect patterns in the tails. It cannot decide, by itself, what those patterns mean.
Making Tail-Sensitive Evidence More Visible in Regulatory
Decisions
Tail-sensitive evidence does not require an entirely new regulatory system. Regulators already
assess benefit, harm, uncertainty, and risk management. The practical objective is to make
clinically decisive distributional information, rare harms, and susceptibility patterns more visible
within existing benefit–risk assessment, product labeling, risk-management plans, and
postmarketing requirements.
FDA guidance recognizes that benefit–risk assessments may support premarket and postmarket
decisions, including approval, labeling, warnings, postmarketing requirements, postmarketing
commitments, and risk-management measures. FDA describes postmarketing requirements as
tools to assess known serious risks, investigate signals of serious risk, or identify unexpected
serious risks when available information suggests potential concern.
Regulatory
stage
Premarket review Role of tail-sensitive evidence
Identify severe toxicity, clinically important deterioration, exceptional
benefit, potentially vulnerable groups, and uncertainty around uncommon
Product labeling events
Inform contraindications, warnings, precautions, dose adjustment,
monitoring recommendations, and population-specific statements
Conditional or
accelerated
Define unresolved tail-related questions requiring confirmation, enhanced
surveillance, or additional studies
approval
Postmarketing
surveillance
Detect rare, delayed, or subgroup-concentrated harm; reassess benefit–risk;
revise labeling; require risk minimization or further research
European risk-management planning follows a similar logic. The EMA Guideline on Good
Pharmacovigilance Practices, Module V, describes the risk-management plan as a structured tool
for identifying and characterizing important risks, documenting missing information, planning
pharmacovigilance activities, and specifying risk-minimization measures over the medicinal
product life cycle.
Tail-sensitive information should not automatically prevent approval. A rare serious harm has
different implications for preventive treatment in healthy people than for therapy in advanced
cancer without effective alternatives. Decisions depend on disease severity, unmet need,
magnitude and certainty of benefit, reversibility and preventability of harm, available
alternatives, and whether vulnerable patients can be identified and protected.
Conclusion
Clinical trials should be interpreted as decision tools, not simply as demonstrations of average
efficacy. Average effects remain indispensable, but the clinical value of treatment depends on
how benefits and harms are distributed across people, time, and levels of vulnerability.
Human beings differ across interacting biological, clinical, psychological, social, and time-
dependent spectra. Disease and response to treatment vary accordingly. Severe harm, profound
nonresponse, clinically important deterioration, exceptional benefit, vulnerability patterns, data-
quality problems, and unexpected threats to benefit–risk balance may all be concentrated in the
tails of outcome distributions.The tails should not be treated as statistical debris. They are where
clinical reality often becomes
visible, and where the adequacy of trial design, measurement, monitoring, and interpretation can
be tested.
A Tail-Sensitive Evidence Framework does not require every trial to answer every question. It
requires investigators to identify clinically decisive tails prospectively; collect the information
needed to interpret them; apply an analytic hierarchy that limits false discovery; investigate
extreme outcomes as possible biological, safety, and quality signals; report findings and
uncertainty transparently; and continue learning through postmarketing surveillance, registries,
pragmatic trials, real-world evidence, and carefully validated computational methods.
The purpose of tail-sensitive clinical research is not to displace the mean. It is to place the mean
within the fuller reality required for clinical decision-making: who benefits, who does not, who is
harmed, when risk emerges, what uncertainty remains, and how that uncertainty will be resolved.
References
International Council for Harmonisation of Technical Requirements
for Pharmaceuticals for Human Use. ICH E8(R1): General considerations for clinical studies.
Geneva, Switzerland: ICH; 2021.
International Council for Harmonisation of Technical Requirements for Pharmaceuticals
for Human Use. ICH E9: Statistical principles for clinical trials. Geneva, Switzerland:
ICH; 1998.
International Council for Harmonisation of Technical Requirements for Pharmaceuticals
for Human Use. ICH E9(R1): Addendum on estimands and sensitivity analysis in clinical
trials to the guideline on statistical principles for clinical trials. Geneva, Switzerland:
ICH; 2020.
Hopewell S, Chan AW, Collins GS, et al. CONSORT 2025 statement: updated guideline
for reporting randomised trials. BMJ. 2025;389:e081123.
Junqueira DRG, Zorzela L, Golder S, et al. CONSORT Harms 2022 statement,
explanation, and elaboration: updated guideline for the reporting of harms in randomised
trials. BMJ. 2023;381:e073725.
Action to Control Cardiovascular Risk in Diabetes Study Group. Effects of intensive
glucose lowering in type 2 diabetes. N Engl J Med. 2008;358:2545-2559.
Action to Control Cardiovascular Risk in Diabetes Study Group. Long-term effects of
intensive glucose lowering on cardiovascular outcomes. N Engl J Med.
2011;364:818-828.
Calles-Escandón J, Lovato LC, Simons-Morton DG, et al. Effect of intensive compared
with standard glycemia treatment strategies on mortality by baseline subgroup
characteristics: the ACCORD trial. Diabetes Care.2010;33:721-727.
Basu S, Raghavan S, Wexler DJ, Berkowitz SA. Characteristics associated with
decreased or increased mortality risk from glycemic therapy among patients with type 2
diabetes and high cardiovascular risk: machine learning analysis of the ACCORD trial.
Diabetes Care. 2018;41:604-612. doi:10.2337/dc17-2252.
Conley BA, et al. The Exceptional Responders Initiative: feasibility of a National Cancer
Institute pilot study. J Natl Cancer Inst. 2021;113:27-37. doi:10.1093/jnci/djaa061.
Bresalier RS, Sandler RS, Quan H, et al; Adenomatous Polyp Prevention on Vioxx Trial
Investigators. Cardiovascular events associated with rofecoxib in a colorectal adenoma
chemoprevention trial. N Engl J Med.2005;352:1092-1102. doi:10.1056/NEJMoa050493.
Graham DJ. Rofecoxib voluntarily withdrawn from market. CMAJ. 2004;171:1027-1028.
U.S. Food and Drug Administration. Decisional summary: COX-2 selective and
nonselective non-steroidal anti-inflammatory drugs. Silver Spring, MD: FDA; 2005.
U.S. Food and Drug Administration. Enrichment strategies for clinical trials to support
approval of human drugs and biological products: guidance for industry. Silver Spring,
MD: FDA; 2019.
Sherman RE, Anderson SA, Dal Pan GJ, et al. Real-world evidence—what is it and what
can it tell us? N Engl J Med. 2016;375:2293-2297.
U.S. Food and Drug Administration. Considerations for the use of real-world data and
real-world evidence to support regulatory decision-making for drug and biological
products: guidance for industry. Silver Spring, MD: FDA; 2023.
U.S. Food and Drug Administration. Real-world evidence. FDA Science and Research
Special Topics. Accessed August 2026.
Wager S, Athey S. Estimation and inference of heterogeneous treatment effects using
random forests. J Am Stat Assoc. 2018;113:1228-1242.
Künzel SR, Sekhon JS, Bickel PJ, Yu B. Metalearners for estimating heterogeneous
treatment effects using machine learning. Proc Natl Acad Sci U S A.
2019;116:4156-4165.
Kent DM, Hayward RA. Limitations of applying summary results of clinical trials to
individual patients: the need for risk stratification. JAMA. 2007;298:1209-1212.
Kravitz RL, Duan N, Braslow J. Evidence-based medicine, heterogeneity of treatment
effects, and the trouble with averages. Milbank Q. 2004;82:661-687.
U.S. Food and Drug Administration. Benefit-risk assessment for new drug and biological
products: guidance for industry. Silver Spring, MD: FDA; 2021.
U.S. Food and Drug Administration. Postmarketing requirements and commitments:
introduction. Silver Spring, MD: FDA. Accessed August 2026.
European Medicines Agency. Guideline on good pharmacovigilance practices (GVP):
Module V—Risk management systems (Rev 2). EMA/838713/2011 Rev 2. Amsterdam,
Netherlands: EMA; 2017.
Comments
Post a Comment