EFSA Scientific Committee 2017 — Update on the use of the benchmark dose approach in risk assessment

The EFSA Scientific Committee’s 41-page guidance document (adopted 17 November 2016, published EFSA Journal 2017;15(1):4658, question EFSA-Q-2014-00747) updates the Committee’s 2009 BMD guidance (EFSA 2009a) on the basis of seven years of in-house application experience and methodological developments. The opinion reconfirms that the benchmark dose (BMD) approach is scientifically more advanced than the no-observed-adverse-effect level (NOAEL) approach for deriving a Reference Point (RP), introduces model averaging as the preferred method for calculating the BMD confidence interval (with rejection-on-AIC as a suboptimal fallback while model-averaging tools mature), replaces the log-likelihood with the Akaike Information Criterion (AIC) for goodness-of-fit characterisation across mathematical models, narrows the default set of recommended dose-response models, and locks in the requirement to report the BMD confidence interval (BMDL as the potential RP, BMDU for the BMDU/BMDL uncertainty ratio) rather than the point BMD. The opinion does not call for general re-evaluation of past NOAEL- or BMD-based assessments, but updates downstream guidance for risk assessors performing dose-response analyses on quantal, continuous, and epidemiological data. Appendix A summarises differences between the BMDS (US EPA) and PROAST (RIVM) software; Appendix B is the BMD analysis reporting template.

The guidance applies to all chemicals in food regardless of category — pesticides, additives, contaminants — and explicitly extends to both experimental animal toxicity studies and observational epidemiological data, with the caveat that human-data BMD analysis will be the subject of a separate forthcoming Scientific Committee guidance. The PDF retrieved 2026-06-03 carries an “OBSOLETE” watermark on every page, indicating that this 2017 update has itself been superseded by a later EFSA Scientific Committee revision; this document nonetheless remains the historical RP-derivation methodology for EFSA’s pre-2022 contaminant opinions and is required reading for reconstructing the dose-response logic those opinions are built on.

Key numbers

Default benchmark response (BMR) values recommended for RP derivation (Section 2.5.2, Conclusions):

Illustrative upper bounds on the effect size at the NOAEL (Section 2.3.3, Table 1; selection of 10 substances previously evaluated by JMPR or EFSA):

Two-sided 90% confidence interval (one-sided 95%) for the effect size at the NOAEL is the Scientific Committee’s chosen reporting convention (Table 1 footnotes a-c).

Worked Example 1 — NTP subchronic rat body-weight study (Section 2.3.4, Table 2; doses 0, 8, 25, 76, 215, 419 mg/kg b.w., n=10 per group):

NOAEL = 76 mg/kg (highest dose without statistically significant body-weight depression); upper 95% confidence bound on the effect size at the NOAEL = 4.7% body-weight decrease. Applying BMR = 5% to a fitted exponential model y = a·exp(bxᵈ) (PROAST v. 61.5) yields BMDL₀₅ = 170 mg/kg with associated effect-size upper bound of 1.3% (vs the 4.7% bound at the NOAEL). The BMDL₀₅ (170 mg/kg) is therefore higher than the NOAEL (76 mg/kg) in this example, while the effect-size guarantee at the BMDL₀₅ is tighter than at the NOAEL.

Worked Example 2 — gastric-impaction quantal data (Section 2.3.4, Figure 3; pairwise NOAEL = 450 mg/kg with upper 95% CI of ≈ 47% extra risk; BMR = 10% extra risk; log-logistic dose-response model fitted via PROAST and BMDS): BMDL₁₀ = 171 mg/kg, i.e. 2.6-fold lower than the NOAEL of 450 mg/kg.

Worked Example 3 — human epidemiological dose-response (Section 2.3.4, Figure 4; eye-hand coordination scores 0/1 in workers, individual exposure CRD; one-stage model y = a + (1−a)(1−exp(−x/b)) fitted via PROAST v. 61.4; BMR = 10% extra risk): BMD₁₀ = 173 (units = log₁₀ CRD), BMDL₁₀ = 92.

Model-averaging weights example (Section 2.5.9 Example 2, quantal thyroid epithelial cell vacuolation incidence in a 2-year female-rat study with three dose groups, PROAST v. 62.3, BMR = 0.10 extra risk):

Model-averaged BMDL = 1.5 mg/kg. Surrogate-method confidence interval using log-probit and log-logistic (the two models within two AIC units of the lowest) = (1.8, 5.1) mg/kg; the model-averaged BMDL is slightly lower because the other five models carry residual weight.

Methods (brief)

Guidance document, not an experimental study. Three statistical approaches are explicitly compared and constrained:

• Goodness-of-fit selection. AIC (Akaike 1974; Burnham & Anderson 2004) replaces the log-likelihood for ranking dose-response models. Two AIC units is the locked threshold for “indistinguishable fit” — models whose AIC is within two units of the lowest AIC are considered to fit the data equally well.
• Model averaging. Preferred method for the BMD confidence interval. The MADr-BMD program described in Wheeler and Bailer (2008) is the implementing reference cited for worked examples; Wheeler and Bailer (2007, 2009) and Burnham and Anderson (2004) are cited as the broader methodology basis. When model-averaging software is unavailable, the surrogate method is to take the lowest BMDL and highest BMDU from models within two AIC units of the lowest AIC (suboptimal fallback explicitly flagged).
• Default recommended model set. Reduced relative to the 2009 guidance, with nested-family handling (full Exponential and Hill four-parameter models for continuous data following Slob and Setzer 2014) re-specified.

Two software platforms are recommended and Appendix A documents their differences: BMDS (US EPA, www.epa.gov/bmds, version-controlled) and PROAST (RIVM, www.rivm.nl/proast, version-controlled). The opinion notes that BMDS and PROAST “will lead to the same answer (possibly with minor numerical differences)” when the same models, BMRs, and assumptions are applied; differences are in default distributional assumptions (normal in BMDS vs log-normal in PROAST for continuous data), output format, modelling options, and the availability of the three-parameter Hill model (PROAST only).

Confidence-interval convention: two-sided 90% (equivalent to one-sided 95%) for the BMD confidence interval, applied throughout the worked examples.

Uncertainty factors: the opinion concludes that the default uncertainty factors currently applied to the NOAEL (typically 100 for inter- and intra-species variation) are equally applicable to the BMDL. No additional uncertainty factor is required for BMDL-derived health-based guidance values, because the BMDL on average coincides with the NOAEL when default BMR values are used. For LOAEL situations where pairwise comparison shows a statistically significant effect at the lowest tested dose, the opinion notes that the BMD approach can derive a BMDL at the desired BMR without applying a separate LOAEL-to-NOAEL uncertainty factor, unless extrapolation outside the observed dose-range forces a higher BMR.

Per-cent change vs extra-risk distinction: for continuous data, the BMR is expressed as a per-cent change in mean response relative to the fitted background response (Section 2.5.2). For quantal data, the BMR is expressed as “extra risk” — the additional risk above background divided by (100 minus the background response in %) (Section 2.5.2 and footnote 5 on page 9). In Table 1, the upper bounds on effect size at the NOAEL were calculated by the likelihood profile method for quantal endpoints (footnote b) and by computing the 90% CI on log-scale and back-transforming for continuous endpoints (footnote c); these are reporting conventions for the table, not the BMR specification itself.

Reporting template: Appendix B is a standardised template for documenting a BMD analysis, covering (A) Data description, (B) Selection of the BMR, (C) Software used, (D) Specification of deviations from default assumptions, (E) Results (including the table of fitted models with parameters, log-likelihood, AIC, BMDL, and BMDU per Table B.3 and a requirement to highlight models complying with AIC ≤ AIC_min + 2), (F) Plots of fitted models, and (G) Conclusions. The same A–G structure is specified in Section 2.5.9 with minor wording variants. The template is strongly recommended for all dose-response analyses submitted to EFSA for peer review.

Implications

Certification: contributes the methodological floor under any EFSA-derived heavy-metal Reference Point used in HBGV-vs-exposure margin-of-exposure (MOE) reasoning that postdates 2017 and pre-dates the next BMD-guidance update. The opinion does not name specific contaminants or thresholds; its contribution to certification work is upstream of any metal-specific decision.

Courses: BMD-vs-NOAEL is a central methodological concept for the “How heavy-metal risk assessment works” educational module. The three worked examples (continuous body-weight, quantal gastric impaction, epidemiological eye-hand coordination) illustrate how the same conceptual machine handles three very different data types. Figure 1 (key concepts) is the canonical teaching diagram.

App: no contamination occurrence data, no consumer-facing thresholds, no per-product values. The guidance is invisible to the app surface.

Microbiome: not applicable.

Verification notes

• Document carries an “OBSOLETE” watermark on every page of the retrieved PDF, indicating it has been superseded by a subsequent EFSA Scientific Committee BMD guidance update. The exact successor citation was not verified inside this paper; downstream pages should cite the current EFSA BMD guidance for prospective dose-response work and cite this 2017 document only as the historical methodology backbone for pre-2022 EFSA contaminant opinions. superseded_by: left null pending verified citation of the successor; can be filled when the successor source page is ingested.
• Metals, ingredients, products, and matrices arrays are intentionally empty. This is a pure methodology guidance document; it measures no contamination, identifies no food matrix, and proposes no metal-specific or product-specific limits. Routing audit will report empty products: and ingredients: as advisory-level (non-blocking) rather than malformed.
• License CC-BY-ND corrected on the basis of the licence-statement page (page 2 of the PDF): “Creative Commons Attribution-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited and no modifications or adaptations are made.” Not the looser CC-BY used by many later EFSA opinions.
• Sample size null because the document is methodological. The sample_population field documents the illustrative substances used as worked examples (drawn from earlier JMPR/EFSA evaluations) so that downstream synthesis is not misled into treating Table 1 as primary occurrence data.
• Author list reflects the full Scientific Committee membership plus working-group members and EFSA staff per the suggested-citation block on page 2 of the PDF.
• Audit subagent (2026-06-03, fresh-context general-purpose Agent) returned REVISE. Applied corrections: (1) Appendix B section labels in Methods corrected from invented (C) “recommended dose-response models” / (D) “fitting results per model” / (E) “BMD confidence interval” / (F) “sensitivity / model-averaging output” to the source’s actual (C) Software used / (D) Specification of deviations from default assumptions / (E) Results / (F) Plots of fitted models per pages 40–41; (2) model-averaging example relocation from “Appendix’s worked thyroid epithelial cell vacuolation analysis” to its actual location in Section 2.5.9 Example 2 (pages 31–32); (3) MADr-BMD implementing reference narrowed from Wheeler & Bailer 2007/2008/2009 to Wheeler & Bailer 2008 specifically, with 2007/2009 and Burnham & Anderson 2004 kept as the broader methodology basis; (4) BMDS-vs-PROAST equivalence claim restored to the source’s exact “possibly with minor numerical differences” qualifier (page 9), with the distributional-assumption and Hill-model caveats added; (5) continuous-data BMR section restructured so that the per-cent-change-in-mean-response BMR specification (Section 2.5.2) is distinguished from the Table 1 effect-size CI computed on log-scale (footnote c); (6) opening-paragraph enumeration of downstream EFSA CONTAM heavy-metal opinions (Pb 2010 BMDL₀₁ values, Cd 2009 BMDL₅, iAs 2009 cancer BMDL₀₁ range, Cr-VI 2014 NTP-derived BMDLs, Ni 2015/2020 BMDL₁₀ values, JECFA Al) removed because those values are not stated in this paper; the body now restricts itself to what the 2017 guidance actually says, and the methodological-backbone relationship is left to downstream metal-page synthesis. (7) “The wiki should cite this opinion whenever a downstream metal page describes a BMDL or BMDU as the basis for a TWI, TDI, or MOE threshold.” removed from the Implications section as policy-prescription drift; the relationship is implicit and does not need stating in this source’s page. All seven corrections verified independently against the PDF before applying. Numerical fidelity (Tables 1, 2, the seven-model AIC-weight table, the three worked-example BMDL values) verified as ✅ clean by the subagent and re-verified independently here.

Page history

The five most recent substantive edits to this page. The full version history lives in git; when DOI minting comes online (see schema docs), each entry below will also link to a version-pinned DataCite DOI.