LL2-26 digest — Part E introduction and Ch26 Science for precautionary decision-making#

(EEA 2013, report pp. 621–642; PDF pp. 623–644. Philippe Grandjean: environmental epidemiologist, Environmental Health editor, LL2 editorial-team member. No panels.)

What it is. Part E has no introductory essay, only a title page and a contents list for Ch 26–28 (pp. 621–622). Ch 26 is a reflexive position essay by an insider about how environmental health science itself delays prevention. The author admits he is “part of the inertia” he criticises (p. 628 fn 3).

Core argument. Prevention has “too often been deferred due in part to the alleged absence of convincing scientific evidence” (p. 624). The chapter identifies four failings and proposes four remedies (pp. 625, 639):

Research neglects poorly known hazards. The author’s own bibliometrics cover 78 journals, 2000–2009: 119,636 articles and 760,056 chemical links, with the top 20 substances (mostly metals) taking 12% of links (pp. 626–627). The seven “early warnings” substances from LL1 still generated about 40% of all their articles since 1899 in that decade (Table 26.1, p. 626). By contrast, 13 data-poor chemicals on a 2006 US EPA priority list had only 352 links. Five had none, and the chapter reports no uptick by 2011, though without figures (pp. 627–628). For every article on triclosan or a major perfluorinated compound there were about 35 on lead (p. 628). The chapter tentatively (“may”, “perhaps”) attributes this to a Matthew effect (p. 629) driven by replication norms, cheap 1970s–80s instruments, publication-count metrics, funders’ risk aversion (“Uncomfortable surprises are unlikely”, p. 629), easy peer review and mentoring.
Methods are biased toward false reassurance. Nine default assumptions later proved wrong (Table 26.3, p. 630): “safe”/”natural” doses, healthy adult male reference subjects, averages, protective barriers, monotonic thresholds, dismissal of animal data. Lead and ALAD, and “natural” lead levels, illustrate the pattern (p. 631). “The greatest error” is inferring absence of harm from absence of proof (p. 631). Table 26.4 (p. 635) lists ten design features that push toward false negatives against three toward false positives.
Reporting and statistics mislead. The 5% threshold is “almost sacrosanct” (p. 632), and non-significant studies get labelled “negative” or “no risk” rather than “no information” (p. 635). The chapter proposes reporting confidence intervals and treating the upper limit as “a plausible worst case” (p. 633; Figure 26.1, p. 634), with worst-case and sensitivity analyses (p. 638). Monitoring data (Taylor et al., 2007) show that a “precipitous” decline, which that paper defines as 50% in 15 years, would go statistically undetected for 72–90% of whale stocks and 55% of polar bear stocks (p. 634). The chapter mis-glosses this as over half the polar bears having to disappear.
Research is not independent. Proprietary research stays secret, and neutral-sounding industry-funded bodies and friendly journals publish sponsor-favourable work (pp. 636–637). “Sound science” and “Good Epidemiological Practice” were used to discredit unwelcome findings (p. 632). Warners such as Needleman were harassed (p. 637), and hedged language is exploited by selective quotation (p. 637). BPA and perfluorinated compounds are the examples of long-claimed safety overturned by independent research, and the latter “cannot be recalled” (p. 636). (The chapter’s Eurostat figure, that industry funds over half of all EU R&D (p. 624), covers all R&D. The chapter uses it mainly to argue that public funding is large enough to expect research to reflect public priorities.)

Proposals. A “PATIO” research culture (Participatory, Accessible, Transparent, Inventive, Open-minded) to supplement CUDOS (which the chapter attributes to Merton) and Ziman’s PLACE (Table 26.5, p. 638). Uncertainty should be treated as an object of study (pp. 637–638). The research question becomes whether we are confident enough of possibly serious harm “to initiate transparent and democratic procedures” (p. 638). The four recommendations: stakeholder-informed topic choice; innovative rather than repetitive research; communication of magnitude; open and independent research (p. 639). Decisions stay provisional and evaluated (pp. 624, 639). The closing verdict: “science does not have a good track record for supporting decisions on improving environmental health” (p. 640).

Main mechanisms. Path-dependent research agendas; proof standards and error conventions favouring false negatives; asymmetric rigour turned into a weapon (manufactured doubt); funding shaping topics and conclusions; framing (“safe”, “natural”, “negative”); neglect of lay warners (p. 631); irreversibility; averages hiding subgroup harm (pp. 638–639); openness cutting both ways (p. 639).

Transferable insights (selected): - Research attention is path-dependent; flagging a priority doesn’t redirect it (pp. 625–629). Strong (descriptive); moderate (causal). - Absence of proof is read as proof of absence (pp. 631, 635). Strong. - Standard designs are tilted toward false negatives (p. 635). Moderate–strong. - Ask how large an effect a study could have missed, using upper limits and worst cases (pp. 633–635). Strong. - With small samples or rare outcomes, significance can arrive only after irreversible harm (p. 634). Strong. - Rigour and scepticism can be turned into weapons against unwelcome findings (pp. 631–633). Strong for tobacco; moderate in general. - Funding shapes what is studied and what is concluded (pp. 636–637). Strong for drug-trial sponsorship; moderate for environmental toxicology; suggestive for harms-versus-benefits neglect. - Persistent, dispersed agents make delay irreversible (p. 636). Strong as a principle; the chapter’s own support is one example. - Averages hide concentrated harm (pp. 638–639). Strong. - Governance needs evidence enough to trigger a democratic decision, not proof of causation (pp. 623, 638). Asserted (normative).

Caveats. - Single-author advocacy with no dissent, built on the author’s own study published in his own journal. - Counts are not content. The top 20 includes ethanol and essential trace metals (checked against the 2011 paper). The continued lead research he calls over-replication is also what revealed low-dose harm (pp. 624, 629, 631), which he partly concedes (p. 629). - The false-positive side is underweighted: 10:3 in Table 26.4; publication bias toward scares is conceded as “quite likely” but then set aside with topic counts that don’t address it (pp. 635–636); and it relies on Ch 2. - Predates the replication crisis, so “repetitive verification” reads differently now, though the chapter does grant that replication “will be useful” (p. 638) and that single studies are not firm evidence (p. 628). - The statistics contain their own errors: the p-value and CI definitions (pp. 632–633), and a misparaphrase of the whale data (p. 634, checked against the source). - Other errors: “Joe Forman” (Farman, p. 631), a garbled PFC name (p. 628), and CUDOS attributed to Merton alone. - Some labels are advocacy (EPRI grouped with tobacco-funded CIAR as an implied “front group”, though the industry funding itself is accurate; “trade magazines disguised as scientific journals”, pp. 636–637). - The 2010–2011 “no uptick” for the EPA list is stated without figures (pp. 627–628). - Table 26.3 shows only failed assumptions (hindsight). - The verdict of a “poor track record” (p. 640) sits uneasily with footnote 2 (p. 624), which places LL1’s main failures in political will, and with the chapter’s own opening, which grants that science “can provide powerful evidence for targeted prevention” (p. 624).

Hindsight leads (unverified). PFAS strongly vindicated (EFSA 2020 and US EPA 2022–24 used child vaccine-response data). BPA vindicated in the EU (EFSA 2023; EU ban 2024), still disputed elsewhere. Statistics reform broadly followed the chapter (ASA 2016), with a counter-push for stricter thresholds. Open access expanded. “Transparency” rules were used against epidemiology (US EPA 2021, vacated).