The gap between statistics and claims
The statistician produces a table: estimated mean change in hydration at day 28 is +12.4% (95% CI: 9.1%–15.7%, p < 0.001). The marketing team wants to say: "Skin is visibly more hydrated in 4 weeks." These two sentences describe the same result. The question is whether the second sentence is an accurate, proportionate, and verifiable representation of the first.
This translation, from effect size to claim language, is where most cosmetics companies are exposed. The statistical analysis may be rigorous, but if the claim wording overstates the result, or if the connection between the two cannot be demonstrated to an inspector, the claim is non-compliant.
What regulators look for in the connection
Under EU 655/2013 and its Technical Document, inspectors look for three things when reviewing a claim dossier:
- Traceability: can you follow the line from raw data to final claim wording, through every analytical step?
- Proportionality: is the claim language commensurate with the effect size? "Clinically proven" requires a different standard than "users noticed an improvement."
- Specificity: is the claim limited to what was tested? If you tested on 30-45 year old women with normal skin, the claim should not be presented as universal.
Building a traceable results pipeline
A traceable pipeline means every number in your claim dossier has a documented origin. In R:
# File: R/03_claim_summary.R
# This script generates the claim evidence table directly from NCA/LMM results
# No copy-paste. Every number is computed.
library(dplyr); library(broom.mixed); library(gt)
# Load locked analysis results
lmm_results <- readRDS("results/lmm_hydration_locked.rds")
claim_table <- lmm_results %>%
tidy(effects = "fixed", conf.int = TRUE) %>%
filter(grepl("^visit", term)) %>%
transmute(
Timepoint = gsub("visit", "", term),
Change_pct = round(estimate, 1),
CI_lower = round(conf.low, 1),
CI_upper = round(conf.high, 1),
p_value = format.pval(p.value, digits = 2)
)
gt(claim_table) %>%
tab_header(title = "Hydration improvement vs baseline - Evidence table")
# This table is included verbatim in the claim dossier Word document
# via officer::body_add_flextable()When the claim dossier is assembled, every number comes from this script, not from a slide deck or email. The script is version-controlled. The analysis dataset is locked and checksummed.
Claim wording rules from EU 655/2013
The Technical Document is specific about what claim language implies. Key rules:
- "Clinically proven" / "dermatologically tested" implies a study conducted or supervised by a medical professional, with the full study available on request.
- Percentage claims must be accompanied by the methodology on request: "88% of users reported improvement" requires a defined measurement method, a defined improvement threshold, and a representative sample.
- Superlatives ("best," "most") require comparative data against market alternatives, not just baseline.
- Before/after visuals must show real results from the actual study, not cherry-picked subjects.
Translating effect sizes into honest claims
| Statistical result | Acceptable claim | Overclaim (non-compliant) |
|---|---|---|
| +12% hydration vs BL (p<0.001) | "Improves skin hydration" | "Dramatically boosts hydration" |
| HR = 0.6 for dryness recurrence | "Reduces dryness recurrence" | "Eliminates dryness" |
| 72% responders >1pt improvement | "Majority of users saw improvement" | "Almost all users see results" |
| d=0.3 (small effect size) | "Helps improve skin texture" | "Visible improvement in skin texture" |
The consumer test supplement
EU 655/2013 distinguishes between instrumental/expert assessment and consumer self-assessment. Many claims: particularly sensory and aesthetic claims, require or are strengthened by consumer test data: questionnaires administered at each time point, asking subjects to rate perceived improvement on defined scales.
Consumer test data is analyzed separately (typically mixed models on VAS or Likert scores) and supports claims that begin with "users felt..." or "skin feels..." rather than objective measurement claims.
Documentation that survives an inspection
The minimum claim dossier for an efficacy claim under EU 655/2013:
- Signed study protocol (pre-study)
- Raw data (locked, checksummed)
- Analysis code (version-controlled, with renv lockfile)
- Analysis outputs (tables, figures)
- Claim mapping document: a table linking each claim to the specific statistical result that supports it
- Claim wording rationale: why the chosen language is proportionate to the effect size
The last two items are the ones most often missing. The statistics may be solid. If you cannot produce a document that says "this claim is supported by this result from this analysis of this dataset," you are exposed.
Key takeaway
The data pipeline from trial to claim is not just an analytical pipeline: it is a documentation pipeline. Every decision point must leave a trace. The claim dossier should be assembled by the pipeline, not by a person copying from a presentation.