Biosimilar Statistical Assessment Requirements

Why "No Significant Difference" Is Not Enough

Suppose a comparative efficacy study produces:

That does not establish equivalence.

A non-significant difference may simply reflect inadequate precision.

For example, an estimate could be:

with a very wide confidence interval:

A conclusion of equivalence depends on whether the entire confidence interval lies within the prespecified equivalence margin.

Choosing the Equivalence Margin

The equivalence margin is one of the most important statistical assumptions in a comparative efficacy study.

A defensible margin should be based on clinical considerations and the established efficacy of the reference product.

It should not simply be selected to make the sample size convenient.

The margin should be justified before unblinding the comparative analysis.

Sample-Size Planning

For a simple two-arm equivalence comparison of a continuous endpoint, a conceptual sample-size expression is:

where:

• \(\sigma\) = standard deviation
• \(M\) = equivalence margin
• \(\alpha\) = type I error component
• \(1-\beta\) = desired power

This is only a simplified planning equation.

Actual biosimilar study calculations may need to account for:

• Ratio-based endpoints
• Log transformation
• Unequal allocation
• Dropout
• Repeated measurements
• Covariate adjustment
• Stratification
• Binary response endpoints
• Non-normal endpoints
• Multiple primary endpoints

Example Sample-Size Calculation

Suppose an equivalence study uses:

• Equivalence margin = 10 units
• Standard deviation = 20 units
• Power = 90%
• Two-sided alpha = 5%

The simplified calculation is:

which produces approximately:

before accounting for dropout and other design considerations.

Multiplicity Across Endpoints

Biosimilar studies can involve multiple endpoints.

For example:

Not every endpoint necessarily belongs to one conventional multiplicity family.

The statistical analysis plan should clearly distinguish:

• Primary similarity criteria
• Supportive endpoints
• Descriptive analyses
• Safety summaries
• Exploratory analyses

Analytical Similarity and Multiplicity

Analytical similarity presents a particularly interesting multiplicity problem.

A development program may compare dozens or hundreds of attributes.

It is usually inappropriate to treat every individual comparison as though it were a standalone confirmatory hypothesis test.

Instead, the statistical framework should reflect the role of each quality attribute and the overall comparability strategy.

EMA's statistical reflection paper specifically addresses statistical methodology for comparative quality attributes, including the comparison objective, sampling strategy, variability, acceptance ranges, and inferential methods.

Distribution-Based Comparisons

For continuous analytical attributes, useful summaries may include:

• Means
• Standard deviations
• Variances
• Quantiles
• Reference-product ranges
• Probability distributions
• Confidence intervals
• Prediction intervals

The choice depends on the nature of the attribute and the scientific question.

Difference in Means

A simple comparison can be represented as:

with confidence interval:

This approach may be useful for some normally distributed quality attributes, but it should not be applied automatically to every attribute.

Ratio of Means

For multiplicative attributes:

the logarithmic scale may provide a more appropriate analysis:

Reference Variability

Suppose the reference product has substantial lot-to-lot variability.

The statistical assessment should distinguish:

• True product variability
• Sampling variability
• Analytical assay variability
• Batch effects
• Measurement error

Failure to separate these sources can produce either an overly permissive or overly restrictive similarity assessment.

Batch Selection

The batches selected for analytical comparison should be scientifically representative.

Potential considerations include:

• Manufacturing history
• Commercial relevance
• Lot age
• Storage conditions
• Manufacturing sites
• Reference-product variability

Statistical analysis cannot rescue a poorly designed sampling strategy.

Immunogenicity

Immunogenicity is an important component of biosimilar development.

Relevant outcomes can include:

• Anti-drug antibody incidence
• Neutralizing antibody incidence
• Time to antibody development
• Antibody persistence
• Antibody titers
• Effects on PK
• Effects on efficacy
• Effects on safety

The analysis may include categorical comparisons, time-to-event methods, longitudinal analyses, and exposure-response evaluations.

Binary Immunogenicity Endpoints

Suppose ADA incidence is:

The crude risk difference is:

or −1.1 percentage points.

The interpretation should consider the confidence interval and the prespecified analysis rather than the point estimate alone.

Time-to-Immunogenicity

If timing of ADA development is clinically important, survival-analysis methods can be considered.

For example:

• Kaplan-Meier curves
• Cox proportional hazards models
• Restricted mean survival time

The choice should be driven by the estimand and scientific question.

Missing Data

Missing data are especially important in clinical comparative studies.

Potential causes include:

• Study withdrawal
• Adverse events
• Loss to follow-up
• Assay failure
• Protocol deviations
• Sample handling problems

Missingness should not automatically be handled using simple last-observation carried forward.

Instead, the statistical analysis plan should define appropriate estimands, analysis populations, and sensitivity analyses.

Estimands in Biosimilar Studies

The estimand framework provides a structured way to specify exactly what treatment comparison is being estimated.

Important dimensions include:

• Population
• Treatment condition
• Variable
• Intercurrent events
• Summary measure

For example, a clinical efficacy estimand might specify whether treatment discontinuation because of adverse events is handled using a treatment-policy, hypothetical, composite, or other strategy.

Analysis Populations

A biosimilar clinical study may define several analysis populations.

For equivalence studies, interpretation of both full-analysis and per-protocol analyses can be particularly important.

Concordance Between Analysis Populations

Suppose the primary analysis demonstrates equivalence in the per-protocol set but not in the full analysis set.

This discrepancy requires investigation.

Conversely, agreement between appropriately defined populations strengthens confidence in the robustness of the conclusion.

Sensitivity Analyses

Sensitivity analyses can assess whether the similarity conclusion depends on specific assumptions.

Examples include:

• Alternative missing-data assumptions
• Alternative covariance structures
• Alternative baseline adjustments
• Alternative analysis populations
• Exclusion of influential observations
• Alternative handling of protocol deviations

Protocol Deviations

Protocol deviations can be particularly important in equivalence studies.

A deviation that affects treatment exposure, assessment timing, or endpoint measurement can influence the ability of the study to distinguish products.

The analysis plan should prospectively define which deviations are considered important and how they affect analysis populations.

Outliers

An outlier should not be removed simply because it makes a similarity criterion fail.

The analyst should determine whether the observation represents:

• A genuine biological value
• Measurement error
• Sample contamination
• Data-entry error
• Protocol deviation
• Assay failure

Any exclusion should be supported by predefined criteria and documented.

Reference Product Drift

Reference products can evolve over time because of manufacturing changes.

This creates a statistical and scientific question about which reference batches should represent the relevant reference-product distribution.

The solution is not simply to average every historical batch available.

The selection should reflect the scientific and regulatory context of the development program.

FDA Versus EMA: High-Level Comparison

This table is intentionally high-level. Regulatory requirements should always be assessed against the current product-specific guidance and agency advice.

FDA: Current Statistical Direction

FDA's September 2025 final guidance on comparative analytical assessment provides the current finalized framework for analytical similarity assessment of therapeutic protein biosimilars.

FDA also maintains a clinical pharmacology guidance describing the clinical pharmacology data used to support biosimilarity determinations.

The FDA's October 2025 draft guidance on comparative efficacy studies signals a more selective approach to when comparative efficacy studies may be needed. Because it remains draft, sponsors should not treat its recommendations as binding requirements.

EMA: Current Statistical Direction

EMA's overarching biosimilar guideline remains an important foundation for the European framework. The current effective overarching guideline is Revision 1, while a new revision was opened for consultation in July 2026.

More importantly for the current clinical-development strategy, EMA adopted the Reflection Paper on a Tailored Clinical Approach in Biosimilar Development in March 2026.

The reflection paper recognizes that dedicated comparative efficacy studies may not be required for well-characterized biological products when analytical and PK evidence is sufficiently informative, together with appropriate safety and other relevant evidence.

Why the Analytical Assessment Can Reduce Clinical Uncertainty

Consider two hypothetical development programs.

The statistical strategy should therefore be driven by residual uncertainty, not by a fixed checklist applied identically to every molecule.

Totality of Evidence

A biosimilarity conclusion can be represented conceptually as:

This is not a literal arithmetic formula.

It represents the conceptual integration of different evidence streams.

Residual Uncertainty

A useful way to think about the development program is:

The purpose of each study is to reduce a specific type of uncertainty.

Statistical Power in Biosimilar Development

Power should be considered carefully because a biosimilar study is often designed to establish similarity rather than superiority.

For an equivalence study:

Conceptually, power is the probability that the confidence interval will fall entirely inside the prespecified similarity region when the products are sufficiently similar.

Precision Is Critical

Suppose two studies have the same estimated treatment ratio:

Study A:

Study B:

Both point estimates are identical.

But Study A provides substantially more precise evidence.

This illustrates why sample size and variance control are central to biosimilar study design.

Confidence Intervals Are More Informative Than P-Values

A p-value can answer whether a null hypothesis is rejected under a particular test.

A confidence interval additionally shows:

• The estimated magnitude of the difference
• The direction of the difference
• The uncertainty surrounding the estimate
• Whether the complete interval fits within the similarity margin

For biosimilarity, these features are particularly important.

Common Statistical Mistake: Testing for Equality

Suppose an analyst performs:

and obtains:

It is incorrect to conclude:

"The products are statistically equivalent."

The appropriate interpretation is that the analysis did not provide evidence against the null hypothesis of equality.

That is not the same as demonstrating that any possible difference is sufficiently small.

Common Statistical Mistake: Using an Arbitrary Margin

Suppose the observed difference is only 2%.

An analyst might choose an equivalence margin of 20% and conclude equivalence.

But if a 20% difference would be clinically meaningful, the analysis is not scientifically persuasive.

Common Statistical Mistake: Ignoring Assay Variability

Suppose an analytical assay has a coefficient of variation of 12%.

A measured difference of 5% may be indistinguishable from assay variability.

A measured difference of 30% may be much more difficult to explain as random measurement variation.

The statistical model should therefore account for the properties of the measurement process.

Common Statistical Mistake: Overinterpreting One Quality Attribute

A biosimilar may show a modest difference in one quality attribute while remaining highly similar across many other characteristics.

That single difference should be evaluated in context.

Questions include:

• Is the attribute clinically relevant?
• Is the difference reproducible?
• Does it affect function?
• Is it within reference-product variability?
• Does it affect PK?
• Does it affect immunogenicity?
• Does it alter the mechanism of action?

R Implementation: PK Equivalence

A simple log-transformed PK analysis in R can be implemented using a linear model.

fit <- lm(
  log(AUC) ~ TRT + SEQUENCE + PERIOD,
  data = pk
)

summary(fit)

coef_ci <- confint(
  fit,
  "TRTTest",
  level = 0.90
)

gmr <- exp(coef(fit)["TRTTest"])

ci_gmr <- exp(coef_ci)

gmr
ci_gmr

The exact model depends on the study design.

For a crossover study, subject effects and the appropriate period and sequence structure should be incorporated according to the prespecified analysis.

R Implementation: Equivalence Confidence Interval

For a continuous endpoint, the estimated difference and confidence interval can be compared with the equivalence limits.

estimate <- -0.8
lower <- -5.2
upper <- 3.6

margin <- 10

equivalent <-
  lower > -margin &&
  upper < margin

equivalent

The result is TRUE only when the entire confidence interval lies within the prespecified equivalence region.

R Implementation: Ratio-Based Similarity

estimate <- 1.03
lower <- 0.98
upper <- 1.08

lower_margin <- 0.80
upper_margin <- 1.25

similar <-
  lower > lower_margin &&
  upper < upper_margin

similar

Again, the limits shown here are illustrative. Actual biosimilar acceptance criteria must be scientifically justified and prespecified.

R Implementation: Analytical Quality Attributes

Suppose a dataset contains:

product
batch
attribute
value

A simple descriptive comparison might begin with:

library(dplyr)

summary_stats <- data %>%
  group_by(product, attribute) %>%
  summarise(
    n = n(),
    mean = mean(value, na.rm = TRUE),
    sd = sd(value, na.rm = TRUE),
    median = median(value, na.rm = TRUE),
    .groups = "drop"
  )

This is only a descriptive starting point.

The regulatory similarity assessment may require more sophisticated analyses depending on the attribute and agreed analytical strategy.

R Visualization of Analytical Similarity

library(ggplot2)

ggplot(
  data,
  aes(
    x = product,
    y = value
  )
) +
  geom_boxplot() +
  facet_wrap(~attribute, scales = "free_y") +
  labs(
    x = NULL,
    y = "Measured Attribute"
  ) +
  theme_classic()

Distributional graphics can help identify:

• Differences in central tendency
• Differences in spread
• Outliers
• Batch effects
• Potential shifts in the reference distribution

Forest Plot for Similarity

A forest plot can summarize confidence intervals across multiple endpoints.

library(ggplot2)

ggplot(
  results,
  aes(
    x = estimate,
    y = endpoint
  )
) +
  geom_errorbarh(
    aes(
      xmin = lower,
      xmax = upper
    ),
    height = 0.15
  ) +
  geom_point() +
  geom_vline(
    xintercept = 1,
    linetype = "dashed"
  ) +
  theme_classic()

For ratio-based endpoints, the reference value is:

For difference-based endpoints, the reference value is:

Recommended Statistical Workflow

Statistical Analysis Plan

The SAP should clearly define:

• Analysis populations
• Primary and secondary endpoints
• Baseline definitions
• Transformation rules
• Statistical models
• Covariates
• Random effects where applicable
• Confidence interval level
• Acceptance criteria
• Missing-data handling
• Protocol-deviation rules
• Outlier handling
• Sensitivity analyses
• Multiplicity strategy
• Software and validation procedures

Pre-Specification Is Especially Important

A similarity margin selected after reviewing the results is not credible as a confirmatory criterion.

Likewise, selectively changing the analysis population after seeing which population produces the desired conclusion undermines the interpretation.

The critical statistical decisions should therefore be established before unblinding or database lock, as applicable to the analysis.

Quality-Control Checklist

Example Statistical Assessment

This table is illustrative only.

In an actual submission, each criterion must be scientifically justified and aligned with the applicable regulatory strategy.

What Happens When a Criterion Fails?

Suppose a PK endpoint produces:

with:

If the upper similarity limit is 1.25, the confidence interval extends just beyond the limit.

The appropriate response is not simply to discard the study.

Questions should include:

• Was the study appropriately designed?
• Was the endpoint measured reliably?
• Was there unexpected variability?
• Were important covariates imbalanced?
• Were there assay problems?
• Were there anti-drug antibodies?
• Was the observed difference reproducible?
• What do the analytical and functional data show?

EMA's clinical pharmacology Q&A specifically discusses investigation of root causes when PK confidence intervals fail the prespecified acceptance limits and emphasizes interpretation of the complete evidence rather than simply ignoring an unsuccessful study.

Why Failed Studies Cannot Simply Be Hidden

Suppose two PK studies are performed.

The existence of Study 1 does not automatically make Study 2 irrelevant.

The discrepancy should be investigated and incorporated into the overall scientific assessment.

This is an important principle for regulatory statistics:

Interchangeability Is Different From Biosimilarity

In the United States, a biosimilar designation and an interchangeable designation are related but distinct regulatory concepts.

A product can satisfy biosimilarity requirements without automatically satisfying the additional requirements for interchangeability.

Therefore, statistical planning for an interchangeability program should not simply copy the statistical strategy used for an ordinary biosimilarity demonstration.

FDA's guidance framework separately addresses considerations for demonstrating interchangeability.

Switching and Multiple-Period Designs

Where interchangeability is being investigated, a study may involve switching between reference and biosimilar products.

Such designs introduce additional statistical considerations, including:

• Sequence effects
• Period effects
• Carryover considerations
• Within-subject variability
• Repeated exposure
• Immunogenicity
• Safety following switching

The analysis must be designed around the specific regulatory question rather than simply around generic crossover-study methods.

EMA's Tailored Clinical Approach

The 2026 EMA reflection paper is particularly important for statistical programmers because it reinforces the idea that the clinical program should be proportionate to residual uncertainty.

For sufficiently well-characterized biological substances, analytical and functional similarity may provide highly sensitive evidence of similarity, while PK and appropriate safety information can address remaining clinical uncertainty.

The implication is that statistical planning should begin with the question:

rather than:

EMA's adopted reflection paper explicitly states that comparative efficacy studies may not be required for well-characterized biological substances when analytical and PK evidence is sufficiently informative, together with appropriate safety evidence.

Regulatory Strategy and Statistical Strategy

The statistical plan should therefore be integrated with the scientific development strategy.

Common Mistakes

1. Treating biosimilarity as ordinary superiority testing. The statistical objective is different.
2. Assuming 80–125% applies to every biosimilar comparison. Acceptance criteria are endpoint- and scientifically dependent.
3. Using "p > 0.05" as proof of similarity. Non-significance is not evidence of equivalence.
4. Choosing the margin after reviewing results. Similarity margins must be scientifically justified and prospectively defined.
5. Ignoring reference-product variability. The reference distribution is central to analytical similarity.
6. Ignoring assay variability. Observed differences must be interpreted in light of measurement precision.
7. Automatically requiring a large efficacy trial. Current FDA and EMA approaches increasingly emphasize scientifically tailored clinical development.
8. Automatically eliminating all clinical efficacy data. A tailored approach does not mean that clinical evidence is universally unnecessary.
9. Ignoring failed or discordant studies. All meaningful evidence should be integrated into the totality-of-evidence assessment.
10. Confusing biosimilarity with interchangeability. The regulatory questions are related but distinct.
11. Using unvalidated analytical datasets. The statistical result is only as reliable as the underlying measurements.
12. Overinterpreting individual quality attributes. An attribute should be interpreted in the context of function and clinical relevance.

What Should a Statistical Report Contain?

A statistical report supporting biosimilarity should allow a reviewer to understand the complete analytical pathway.

At minimum, consider including:

• Study objectives
• Statistical estimands
• Analysis populations
• Endpoint definitions
• Data transformations
• Statistical models
• Variance assumptions
• Similarity margins
• Confidence intervals
• Missing-data handling
• Protocol-deviation handling
• Sensitivity analyses
• Individual-study results
• Integrated interpretation

Suggested Analytical Similarity Table

The exact presentation should follow the statistical methodology agreed for the development program.

Statistical Interpretation of Quality Attributes

An attribute can be statistically different without necessarily being clinically meaningful.

Conversely, a statistically non-significant difference may still require scientific attention if the estimate is large or the study is imprecise.

Therefore, the reviewer should consider:

and:

The complete evidence package is what supports the regulatory conclusion.

Validation of Statistical Programming

Biosimilar analyses may be particularly sensitive to programming errors because small changes in confidence intervals, transformations, or reference populations can affect conclusions.

Programming QC should include:

• Independent programming
• Double-programming of key endpoints
• Dataset-level reconciliation
• Table-to-source verification
• Confidence-interval validation
• Margin validation
• Graphical review
• Extreme-value review
• Traceability to source data

Example of a Statistical QC Check in R

results %>%
  mutate(
    criterion_met =
      lower_ci > lower_margin &
      upper_ci < upper_margin
  ) %>%
  select(
    endpoint,
    estimate,
    lower_ci,
    upper_ci,
    lower_margin,
    upper_margin,
    criterion_met
  )

The output should be independently checked against the prespecified statistical criteria.

Reporting the Overall Conclusion

A biosimilarity conclusion should not be reduced to a single p-value.

A stronger summary describes:

• The analytical evidence
• The functional evidence
• The PK evidence
• The PD evidence where applicable
• The immunogenicity evidence
• The clinical evidence where applicable
• Any residual uncertainties
• The rationale for the overall conclusion

Example Overall Interpretation

This wording is intentionally generic and should not be copied into a regulatory submission without adapting it to the actual product, data, and regulatory strategy.

A Practical Decision Framework

The Most Important Statistical Concepts

Regulatory References

FDA. Development of Therapeutic Protein Biosimilars: Comparative Analytical Assessment and Other Quality-Related Considerations. Final Guidance, September 2025.
FDA. Clinical Pharmacology Data to Support a Demonstration of Biosimilarity to a Reference Product. Final Guidance, December 2016.
FDA. Scientific Considerations in Demonstrating Biosimilarity to a Reference Product: Updated Recommendations for Assessing the Need for Comparative Efficacy Studies. Draft Guidance, October 2025.
EMA. Guideline on Similar Biological Medicinal Products. CHMP/437/04 Rev.1, effective April 2015; revision under consultation in 2026.
EMA. Similar Biological Medicinal Products Containing Biotechnology-Derived Proteins as Active Substance: Non-Clinical and Clinical Issues. EMEA/CHMP/BMWP/42832/2005 Rev.1.
EMA. Reflection Paper on a Tailored Clinical Approach in Biosimilar Development. EMA/CHMP/BMWP/60916/2025, adopted March 2026.
EMA. Reflection Paper on Statistical Methodology for the Comparative Assessment of Quality Attributes in Drug Development. EMA/CHMP/138502/2017.

Key Takeaways