What EU 655/2013 actually requires
EU Regulation 655/2013 (and its associated Technical Document, last updated 2021) establishes the legal framework for cosmetic claims in the European Union. The regulation does not specify statistical methods. It specifies principles, and the statistical implications of those principles are where most companies get into trouble.
The regulation applies to all cosmetic claims: marketing materials, packaging, point-of-sale, digital advertising. It applies to claims made in any EU member state. Non-compliance is enforced at the national level, typically by market surveillance authorities.
The six criteria and their statistical implications
The regulation's six common criteria each have statistical consequences:
- Criterion 1 (Legal compliance): No direct statistical implication, but the claim wording constrains what you can test.
- Criterion 2 (Truthfulness): The claim must be backed by evidence. For efficacy claims, this means a study with an appropriate design and adequate statistical power.
- Criterion 3 (Evidential support): Evidence must be adequate and verifiable. Study protocols, raw data, and analysis code must be retained.
- Criterion 4 (Honesty): No exaggeration. If your study shows 20% improvement, you cannot claim "dramatically improved." The statistics constrain the claim wording.
- Criterion 5 (Fairness): Comparative claims require head-to-head data. "Better than X" requires a controlled study comparing directly to X.
- Criterion 6 (Informed decision-making): Claims must not create confusion. Statistical significance alone is not sufficient: the effect must be meaningful in absolute terms.
Designing a valid efficacy study
Most cosmetic efficacy studies use a within-subject design: the same panel evaluates the product before and after use over a defined period. This design maximizes statistical power because each subject serves as their own control.
The minimum requirements for a claim-supporting study under EU 655/2013 Technical Document guidance:
- Pre-defined primary endpoint and statistical test (written protocol, signed before data collection)
- Sample size justification based on expected effect size and desired power (typically 80% at α = 0.05)
- Defined inclusion/exclusion criteria for panelists
- Standardized measurement conditions (same evaluator, same time of day, same lighting)
- Intent-to-treat analysis as primary (no exclusion of non-responders post-hoc)
# Sample size calculation for a cosmetic efficacy study
# Within-subject design, continuous endpoint (e.g., hydration score)
library(pwr)
# Assumptions: expected effect d = 0.5 (moderate), power = 80%, alpha = 0.05
pwr.t.test(
d = 0.5,
sig.level = 0.05,
power = 0.80,
type = "one.sample" # within-subject: one-sample t-test on change scores
)
# Result: n = 33.4 → recruit 35-40 to account for dropoutChoosing the right statistical test for your claim
The test depends on the endpoint and the claim:
| Claim type | Endpoint | Primary test |
|---|---|---|
| Product improves X vs baseline | Continuous | Paired t-test or LMM |
| % subjects show improvement | Binary responder | Exact binomial (vs 50%) |
| Product better than comparator | Continuous | Two-sample t-test or LMM with treatment factor |
| Expert-graded ordinal score | Ordinal (e.g., IGA) | CLMM or Wilcoxon signed-rank |
| Sensory perception claim | Continuous (VAS) | Paired t-test or LMM |
Mixed models vs ANOVA for panel data
Most cosmetics studies have repeated measures on the same panelists across multiple time points. ANOVA treats panelists as fixed effects and cannot handle missing data without imputation. A linear mixed model (LMM) handles both correctly.
library(lme4)
library(emmeans)
# Hydration score measured at BL, D14, D28, D56
# Random intercept per subject handles between-subject variability
model <- lmer(
hydration ~ visit * treatment + (1 | subject_id),
data = panel_data,
REML = TRUE
)
# Estimated marginal means at each visit
emmeans(model, ~ visit | treatment) %>%
contrast("trt.vs.ctrl", ref = "Baseline") %>%
summary(infer = TRUE)The emmeans output gives you the estimated change from baseline with 95% confidence intervals, exactly what you need to support a claim and document the effect size honestly (Criterion 4).
The baseline correction problem
A common mistake: testing the absolute endpoint value at follow-up rather than the change from baseline. This conflates treatment effect with baseline imbalance. For a within-subject design, always test the change score or include baseline as a covariate.
# WRONG: tests final value, not change
t.test(panel_data$hydration_D28)
# CORRECT: tests change from baseline
panel_data$delta_D28 <- panel_data$hydration_D28 - panel_data$hydration_BL
t.test(panel_data$delta_D28, mu = 0, alternative = "greater")Common mistakes and how to avoid them
- Post-hoc endpoint selection: running multiple tests and reporting only the significant one. Pre-specify your primary endpoint in the protocol.
- Inadequate sample size: 10-subject pilots presented as efficacy evidence. Power analysis is required.
- Responder rate without definition: "89% of users saw improvement" without specifying what threshold defines improvement.
- No correction for multiple comparisons: testing at 3 time points with 5 endpoints and reporting all significant results without adjustment.
Documenting your evidence dossier
The evidence dossier that supports a claim must be retained and available to authorities on request. It should contain: the signed study protocol, raw data, analysis code (with version-controlled R environment), analysis outputs, and an interpretation document linking statistical results to specific claim language.
This documentation is where many cosmetics companies fail. The study may be valid, but if the dossier cannot be assembled and presented to an inspector within days, the claim is effectively unsupported.
Key takeaway
EU 655/2013 compliance is as much about documentation and study design as it is about statistical significance. A well-designed 35-person within-subject study with a pre-specified primary endpoint, a mixed model analysis, and a complete evidence dossier is more defensible than a 200-person study analyzed post-hoc and poorly documented.