Data & Methodology

Primary data sources

We draw on a curated set of authoritative resources. These are the foundation of our ingredient entries and dataset fields:

• Regulatory dossiers & opinions: SCCS (EU), ECHA, FDA summaries, Health Canada, and national regulator publications.
• Peer-reviewed literature: PubMed/MEDLINE-indexed clinical studies, systematic reviews, and meta-analyses.
• Toxicology databases: OECD, ECHA, PubChem (incorporating TOXNET/HSDB archives), and comparable repositories.
• Manufacturer technical data: supplier safety data sheets (SDS), formulation guides, used cautiously and labelled as industry-sourced.
• Adverse event surveillance: national reporting systems and pharmacovigilance-type datasets for signal detection (used as signals, not proof of causation).

For every cited study we capture persistent identifiers (DOI, PMID) and regulator document URLs to enable retrieval.

How we collect data

Data collection is a mix of automated retrieval of publicly available metadata (from compliant APIs and open-access sources) and manual curation.

Automated collection

We run scheduled queries against literature databases (e.g., PubMed Entrez / E-utilities) to retrieve citation metadata (not full-text content). Automated scripts retrieve metadata and DOIs; these results are stored in a staging table for manual verification.

Manual curation

Researchers or contributing research staff manually verify INCI names, match synonyms, check CAS numbers, and extract concentration ranges where available. Label photos submitted voluntarily by users are reviewed and archived for research purposes with minimal metadata (submission date and batch code when present); any personal identifiers are removed prior to storage.

Data provenance

Every record includes provenance metadata: source_type, source_id (DOI/PMID/URL), date_retrieved, and extractor initials. This provenance is preserved in reproducibility packs.

Categorisation & taxonomy

Ingredients and products are organised using a stable taxonomy designed for clarity and searchability. Main categorical fields include:

• Ingredient class: surfactant, preservative, fragrance, emollient, chelating agent, colorant, etc.
• Use-case: rinse-off, leave-on, antibacterial, baby-safe, etc.
• Evidence flag: Evidence Overview | Context Dependent | Under Review.
• Exposure context: typical concentration ranges, frequency, vulnerable populations.

This taxonomy is versioned; changes are logged in our public changelog to preserve interpretability of historical records.

Dataset schema (example)

Below is a simplified example of the ingredient dataset fields we publish in reproducibility packs.

{
"inchi_name": "Sodium Laureth Sulfate",
"cas_number": "9004-82-4",
"synonyms": ["SLES","Laureth sulfate"],
"evidence_flag": "Use with Caution",
"evidence_tiers": [
"Tier A: SCCS opinion",
"Tier B: Patch test study"
],
"typical_concentration_range": "0.5–5%",
"use_case": ["rinse-off"],
"last_reviewed": "2025-10-20",
"primary_sources": [
"https://doi.org/10.1000/xyz",
"https://ec.europa.eu/health/scientific_committees/consumer_safety"
],
"provenance":{
"retrieved":"2025-10-05",
"extractor_role":"Research staff"
}
}

Fields are documented in full in our dataset readme included in reproducibility requests.

How we interpret and score evidence

Interpretation follows the Ingredient Framework: evidence tiers, weighting for clinical relevance, and exposure-adjusted scoring. Key rules include:

• Regulatory opinions and human randomized/controlled trials carry the greatest weight.
• Patch-test and cohort dermatitis data are interpreted in context of concentration and formulation.
• Animal and in vitro mechanistic data inform hypotheses but do not by themselves determine consumer-facing flags.

Our scoring matrix and thresholds are published on the Ingredient Framework page and applied consistently using documented, reproducible worksheets.

Quality control and validation

We employ several QA measures:

• Dual extraction: two research staff members independently extract key data points for critical entries to minimise interpretation errors.
• Source verification: persistent identifiers are checked against registries to prevent citation drift.
• Sanity checks: automated scripts flag improbable concentration ranges or missing CAS/INCI mappings.
• External review may be considered where feasible for complex or high-impact topics.

Analytical methods & statistics

Most of our work is qualitative evidence synthesis. When quantitative synthesis is required (e.g., meta-summary of reported patch-test positivity rates), we apply standard statistical approaches:

• Descriptive statistics (means, medians, ranges) for concentration data.
• Meta-analysis methods where multiple comparable studies exist, using random-effects models when heterogeneity is expected.
• Sensitivity analyses to assess how assumptions about concentration or exposure affect flags.

All code used for statistical analysis is archived in reproducibility packs and documented with versions of packages used (R/Python environments).

Limitations & responsible use

No dataset is perfect. Common limitations:

• Many manufacturers do not publish exact ingredient concentrations, we use published ranges and conservative, clearly stated assumptions.
• Publication bias in clinical literature may skew apparent risk profiles for well-studied ingredients.
• Adverse event reports are useful signals but cannot by themselves establish causation.

Our datasets and summaries are designed for consumer guidance and reproducible research, not clinical decision-making. For clinical concerns consult a licensed healthcare professional.

Data sharing, licensing & how to request datasets

We provide reproducibility packs and dataset extracts to researchers and public-interest projects on request. Typical contents:

• Dataset CSV/JSON with documented schema.
• Search strings and database query logs (dates and exact terms).
• Extraction worksheets and provenance notes (redacted for personal data).
• Analytical code and environment specification for reproducible analyses.

To request a pack, use the contact page and include affiliation, purpose, and the dataset or page(s) you need. Small academic requests are often fulfilled free of charge, subject to capacity; commercial requests are handled under a data license described at /legal/data-license.

Case examples (how data informs decisions)

Example 1: A regulator publishes a new SCCS opinion lowering the acceptable concentration for preservative Y. We record the opinion, update the ingredient entry, recalculate flags for typical product concentrations, and publish a changelog explaining the net effect for consumers.

Example 2: Multiple small patch-test studies report differing irritation rates for surfactant Z. We meta-synthesise comparable results, run sensitivity analyses excluding outliers, and present conservative consumer guidance highlighting uncertainty.