LL2-26 — Part E introduction (Implications for science and governance) and Ch26 Science for precautionary decision-making#

(Late lessons from early warnings: science, precaution, innovation, EEA Report No 1/2013; report pp. 621–642; PDF pp. 623–644)

Reading record. I read the whole text extract in order, from the Part E title page (p. 621) to the last page of references (p. 642). Every table, Figure 26.1, the chapter summary box and the footnoted pages were checked visually against the PDF (report pp. 621, 623, 626, 627, 630–632, 634, 635, 637, 638). The extract garbles Table 26.4 (which features map to which direction of error) and the row layout of Table 26.3; both are corrected below from the page images. To establish the author’s standpoint I also consulted the report’s contents and Acknowledgements (pp. 3–5) and Annex 1, Author biographies (p. 690). Two citations were checked at source: the abstract and full text of Grandjean et al. (2011), via Europe PMC and PMC3229577, and the abstract of Taylor et al. (2007), via Crossref. Later developments listed in the bias-check section come from general knowledge and could not be verified by web search in this pass (the search budget ran out), so treat them as leads to verify.

Authors and standpoint#

Part E “introduction”. Part E has no introductory essay. It opens with a title page carrying a stock photograph of laboratory glassware (p. 621) and a contents page (p. 622) listing the three chapters of Part E: - Ch 26, Science for precautionary decision-making (Grandjean), pp. 623–642. - Ch 27, More or less precaution? (David Gee), pp. 643–669. Its subsections: the power of corporations to oppose action; the precautionary principle’s key elements and misunderstandings; complex biological and ecological systems; “conflicts between the high strength of evidence needed for scientific causality and the lower strength of evidence needed for timely public policy”; the pros and cons of actions and inactions; political and financial short-termism; public participation in hazard and options analysis. - Ch 28, In conclusion, pp. 670–682, including “2001–2013: what new insights emerge?” and “Governance of innovation and innovation in governance”. - A memorial, “In memory of Masazumi Harada and Poul Harremoës” (p. 683).

The report’s contents (pp. 3–4) suggest that Parts A–D each have a few introductory pages before their first chapter, whereas Part E goes straight from contents to Ch 26. The contents page does show the arc of Part E: first the science and how it is produced (Ch 26), then the case for and against precaution and its governance (Ch 27), then overall conclusions (Ch 28). These notes cover only pp. 621–642. Ch 27 and Ch 28 are outside scope.

Author. Philippe Grandjean is the sole author. Annex 1 (p. 690) gives his position in 2013: - Chair of Environmental Medicine, Institute of Public Health, University of Southern Denmark (Odense). - Adjunct Professor of Environmental Health, Harvard School of Public Health. - Editor-in-Chief of the journal Environmental Health, and a member of several other editorial boards. - Research “mainly within the environmental epidemiology of developmental toxicity due to environmental chemicals”. - His work on prenatal methylmercury neurotoxicity “helped inspire international efforts to control mercury pollution”. - He has also published on research ethics, genetic susceptibility, exposure limits and “the impact of the precautionary principle on prevention and research”.

He was also a member of the LL2 editorial team (Acknowledgements, p. 5), so this chapter is written by an insider to the report, not by an outside reviewer. Footnote 1 (p. 623) thanks Mette Eriksen for the SciFinder searches, and John Bailar, Carl Cranor, David Gee and David Kriebel for comments. Gee originated the Late Lessons project and wrote Ch 27. Cranor wrote Ch 24.

Genre and stance. This is a cross-cutting methodological and position essay, not a case study. There are no panels or commentaries, and no response from industry, regulators or dissenting scientists. The stance is openly normative and pro-precaution: “I shall focus primarily on the weaknesses of current applied research in the environmental field” (p. 624). Unusually, it is also reflexive: - Footnote 3 (p. 628). The author “has published numerous articles on lead, mercury, and other top‑20 chemicals … thus being part of the inertia”. He adds that the lead and mercury chapters show “real or alleged uncertainties were often used to argue against hazard abatement, thereby requiring more research”. This is a partial defence of his own position. - Footnote 6 (p. 637). He flags how often he himself writes “may” and “perhaps”, explaining that he prefers understatement.

Self-citation and the evidence base. The chapter’s empirical core is the author’s own bibliometric study (Grandjean et al., 2011). It was published in Environmental Health, the journal he then edited; the paper declares no competing interests (PMC full text). The chapter cites his own work six times: Grandjean 2008a, 2008b, 2004, the 2008 Faroes statement, 2011 and 2012. He is also one of about forty co-signatories of the Axelson et al. (2003) letter it cites to criticise Regulatory Toxicology and Pharmacology (reference list, p. 640). None of this is improper for a position essay, but a reader should know that much of the “independent” support is the author’s own.

Section-by-section notes#

Part E front matter (pp. 621–622)#

Title page and contents only (see above). There is no editorial framing text.

Chapter summary box (p. 623)#

Five paragraphs set out the whole argument: - Academic goals “may differ from those of regulatory agencies”. Feasibility, merit and institutional agendas “may lead to inflexibility and inertia”. - A large proportion of academic research “seems to focus on a small number of well studied environmental chemicals, such as metals”, so research should “to a greater extent consider poorly known problems”. - “Misinterpretation may occur when results published in scientific journals are expressed in hedged language.” A non-significant study “is often said to be negative” and is then misread as “evidence that a hazard is absent”. This comes from traditions that “demand meticulous and repeated examination before a hypothesis can be said to be substantiated”. - Research should focus on “the possible magnitude of potential hazards”, because uncertainties “can blur a real association … thereby resulting in an underestimated risk”. - Research should be “complementary … rather than being repetitive for verification purposes”, and “openly available”.

The reframed question asks whether the evidence is serious enough “to initiate transparent and democratic procedures to decide on appropriate intervention”. On this view, science’s job is not to decide on action but to decide whether a democratic decision process should be triggered. The point recurs on p. 638 and is easy to miss.

26.1 Science and the Precautionary Principle (pp. 624–625)#

• Science’s double role. Science can provide “powerful evidence for targeted prevention”, but it can also be “insufficient … misinterpreted or ignored, so that appropriate intervention is deferred or abandoned” (p. 624). Basic research is not dismissed, but the focus is on “the weaknesses of current applied research” (p. 624).
• Reply to Goldstein and Carruth (2004), who feared the PP makes further research redundant. Decisions “should be considered tentative and amenable to change”, and research, “including intervention studies to determine if the action had the intended effect”, remains needed (p. 624).
• Core diagnosis. “prevention has too often been deferred due in part to the alleged absence of convincing scientific evidence”. The error is recognised only once decisive evidence arrives (p. 624).
• Exposure-limit ratchet. “nearly all exposure limits for hazardous agents have decreased” as harm was found at lower levels, so “when scientific evidence is incomplete, environmental standards are more lenient” (p. 624). This is asserted without data.
• Footnote 2 (p. 624) quotes the LL1 preface: absence of political will “seems to be an even more important factor in these histories than is the availability of trusted information”. The report’s own earlier verdict puts politics ahead of information. This chapter addresses the information side.
• Validity and relevance. Methodologically “‘poor’” research can still be highly relevant, though “a study of limited validity is most likely also to have little impact”. Researchers should avoid becoming “sceptical ivory‑tower nit‑pickers”, and should also avoid “inappropriate (or apparent) advocacy … inspired, though perhaps not justified, by the research” (p. 624). Both failure modes are named.
• Institutional economics. Universities “are enterprises” that must satisfy funders and “avoid going into debt”. Industry funds “more than half” of EU R&D and public sources “slightly more than one‑third” (Eurostat, 2011). Horizon 2020 aimed at “smart investment” for societal challenges. With so much public money, “one would anticipate” research would reflect regulators’ priorities (p. 624).
• Evidence threshold (p. 625). Under the PP, “scientific proof or a very high degree of certainty are not required”, and “Incomplete, but reliable evidence can be sufficient”. The word “reliable” keeps a quality floor. Some uncertainties are accepted as “inevitable or impossible to remove in the time available for preventing plausible harm”. Where evidence is extensive, ordinary risk assessment applies. But “traditional risk assessment is sometimes anti‑precautionary when it demands convincing evidence”, and it inspires “continued elaboration of fairly well documented hazards”. Science does not automatically lead to prudent decisions, and how incomplete data are interpreted “can obfuscate the discussion on the urgency” of protection (p. 625). Asked whether science can better support policy, the author answers: “I think that the answer is yes”, though it “may not be easy to achieve” (p. 625).
• Four questions (p. 625): coverage of societal needs; exploration of emerging hazards as an early-warning system; appropriate reporting; reliability and independence “of vested interests”. These map one-to-one onto the recommendations on p. 639.

26.2 Current research focus is on well-known hazards (pp. 625–627)#

Method. Bibliometrics across 78 journals in toxicology, environmental science and public health. Web of Science is used for title searches back to 1899; SciFinder links articles to chemicals by CAS number for recent years (p. 625). The source is Grandjean et al. (2011), whose abstract I checked: 78 journals, 119,636 articles, 760,056 CAS links, top‑20 at 12%, each top‑20 substance in 2,000–10,000 articles. All match the chapter.

Table 26.1 (p. 626), Web of Science title counts for the LL1 “early warnings” substances.

*MTBE use expanded rather than being restricted (table note). The table spells it “MBTE”.

• Far from fading, the decade 2000–2009 produced about 40% of all articles since 1899 on these substances (the seven substances from LL1’s 14 case studies). The chapter notes the exceptions itself: sulfur dioxide and DES “clearly faded” (about a quarter of titles in 2000–2009), while MTBE “became more popular” (p. 625). SciFinder found 8,267 articles (asbestos not included), “over 10 scientific articles per substance per month” (p. 626).
• All chemicals (p. 626). 119,636 articles carried 760,056 CAS links (about 6 per article), and coverage was “extremely uneven”. The top 100 chemicals each had 600–10,000 articles (180,822 links in total). The top 20 had 2,000–10,000 links each, 12% of all links. Assuming they also account for 12% of articles (the chapter’s stated assumption), that is about 14,264 papers (119 a month): “five or six papers every work day, without holiday breaks”.
• Composition (pp. 626–627). “All of the top‑10 substances are metals” (arsenic counted as a semimetal). “Also well covered” are PAHs, solvents and the PCBs. On average, 51% of all articles on the top 20 were from the latest decade. Arsenic rose to 74%, aluminium fell to 31%. “the chemicals that were popular during the previous century remained a focus”. Checked against the 2011 paper (PMC3229577): the top 10 are copper, lead, zinc, cadmium, iron, nickel, chromium, arsenic, mercury and manganese. Several are essential trace elements, and ethanol is ranked 17. So not every counted article is hazard research. (The paper’s 1,1’-biphenyl entry has the same count, 3,897, as its “Polychlorinated biphenyls” entry. It is the PCB proxy, not a general-chemistry artefact.)

Table 26.2 (p. 627), SciFinder links 2000–2009 for LL2 chemicals.

Some chemicals have “become persistent and highly prominent in the scientific literature”, and topic choice “greatly benefits the well‑known chemicals” (p. 627).

26.3 Ignoring new potential environmental hazards (pp. 627–628)#

• Testing gaps. NRC (1984) found 78% of the most commonly produced industrial chemicals “without even minimal test data for toxicity”. US EPA (1998) found little improvement, and ECHA was still complaining of data gaps in 2011 (Gilbert, 2011) (p. 627).
• EPA’s 2006 high-priority list (cited as US EPA, 2009, Risk-Based Prioritization). Thirteen high-production chemicals were listed as lacking hazard and exposure data. They had 352 links in total for 2000–2009, “only three per month for the entire group”. Five had none. The chapter excuses the gap up to 2006 “and perhaps 2007”, but by 2010–2011 “the priority listing had not inspired any increased number of publications” (pp. 627–628). The chapter gives no figures for 2010–2011. The 2011 paper reports only 2010: 34 links for the whole group, which “did not support any overall increasing trend”. This is the chapter’s clearest measured lag: no research response four to five years after a regulator flagged the need.
• Emerging hazards (p. 628). Triclosan had 259 articles. The most prevalent perfluorinated compound, garbled as “perfluorinated octanoic sulfate” (which conflates PFOA and PFOS), had 271. That is about two dozen a year each, against roughly 35 lead articles (and about 20 on mercury) for every one on these. These figures are not in the 2011 paper (checked), so they can’t be checked against a published source. They were presumably produced by the same SciFinder method.
• Robustness (p. 628). The counts don’t distinguish short reports from reviews, and they miss non-academic work. But “one would have to imagine huge numbers of reports outside the mainstream journals to make up for the differences”. The conclusion: prominence in journals “does not match the societal needs or those of regulatory agencies”.

26.4 Inertia and its reasons (pp. 628–629)#

The mechanisms are argued and hedged, not measured. - Replication norm “stretched to the extreme” when well-known chemicals generate “almost 1 000 publications per year” (p. 628). - Researchers may not know regulators’ needs. Funding limits free topic choice, “especially … young researchers of low academic rank”. Mentoring reproduces “the same type of expertise and narrow focus” (p. 628). - Infrastructure. Atomic absorption spectrometers and gas chromatographs, common since the 1970s–80s, produce a publishable result “within a week or so”. New substances need “expensive equipment and arduous development of new methods” (p. 628). A narrow focus is “counter‑productive in regard to scientific discovery and innovation” (p. 628). The chapter also grants that “a single study should not be relied upon as firm evidence”: its complaint is replication “stretched to the extreme” (p. 628), not replication as such. - Careers (p. 629). Publication count is “a crucial metric for academic prestige and for obtaining a tenured position”. “small incremental manuscripts” and “So‑called vanity publications” result, and budgets tied to counts favour “quantity over quality”. - Short-cuts (p. 629). Questionnaires in place of measurements give “non‑informative” results that “may be interpreted as evidence against the hazard causing any risk at all”. The chapter calls this a “Type III error”, which is a looser use than Schwartz and Carpenter’s “right answer to the wrong question”. - Funders (p. 629). Known problems bring track records, feasible protocols and easy reviewer recruitment. “Uncomfortable surprises are unlikely. Hence, it may be safer and more convenient for grant managers to concentrate on the known hazards.” - Journals (p. 629). Peer review “is rarely a problem with a manuscript on lead exposure”. Reviewers like being cited, and prestige attaches to certain chemicals. The result is “circular reasoning” and “a self‑prophetic bias”. This is Merton’s (1968) Matthew effect. “The opposite strategy would appear more attractive from the point of view of innovation.” - Concession (p. 629). Research on lead and mercury has produced “novel scientific breakthroughs” that were valuable “via analogy, to many other substances”.

26.5 Research methodologies and assumptions (pp. 629–631)#

• Traditional question (p. 630). “Have we reliably documented through meticulous study and replication that this substance is mechanistically and causally linked to an adverse biological change?” The chapter grants the standard its virtue: such standards “protect science from making mistakes”, and well-defined studies “more likely will lead to firm or indisputable conclusions”. But controlled single-factor designs give only “incremental” knowledge and neglect “multiple or complex exposure scenarios and the significance of individual vulnerability”. The result is “reductionism”.
• Table 26.3 (p. 630), “Erroneous assumptions made in initial evaluations … and the subsequent scientific recognition of the true complexity”. It is compiled from “the chapters on human health hazards” in LL2 and LL1. Nine initial assumptions, each followed by the later lesson (mostly hedged with “may” in the source): 1. “Safe” or natural doses are tolerable → delayed, cumulative or re-mobilised doses and toxic metabolites may cause harm at exposures thought safe. 2. No harm in adult male workers means no public risk → children and the elderly are more vulnerable. 3. Acute effects reflect chronic ones → the dose-responses differ. 4. Biological effects need not be adverse → early changes predict later harm. 5. Monotonic dose-response with thresholds → “low dose” effects exist. 6. Short-term, single-pathway exposure assessment is valid → imprecision leads to underestimated toxicity. 7. The placenta and blood-brain barrier protect → both can be bypassed. 8. Averages show the potential for harm → sensitive subgroups are hidden in them. 9. Animal and wildlife data are irrelevant → animal data “have reliably predicted most known carcinogens”.
• Relying on these assumptions “led to proliferation of environmental hazards due to the substantial delay in their recognition” (p. 630).
• Examples (p. 631). In general, biological changes were dismissed as adaptation or “hormesis”. The chapter concedes that “With proper justification, the assumption may be true”, but such changes “should not be disregarded just because they are prevalent (or unwelcome)”. Lead inhibition of the red-cell enzyme ALAD was long believed to have no health implications (LL2 Ch 3). The chapter concedes this “may be true in a strict sense” because the enzyme “has no important function”. The point is that “serious adverse effects do occur at lead exposures that were previously regarded as too low to be harmful”. Habitual lead levels were called “natural”, but lead isotope and mummy studies showed they were “far above what could be considered natural”. Using inbred adult male rats hid variability that mattered once endocrine disruption emerged (LL2 Chs 10, 13; LL1 on PCBs, DES, TBT). Developmental windows are “only a recent discovery”. Birth cohorts are “extremely costly”, and “multi‑generation animal assays are often resisted due to economic burdens on industry”.
• “The greatest error” (p. 631). “If adverse effects were not proven to exist, the erroneous conclusion was drawn that adverse effects must be absent.” Ignored imprecision “will most often result in underestimation”. Such uncertainties “are not likely to create spurious associations, unless confounding factors are present”. That last claim is too strong (see bias check).

26.6 Vulnerability of research to criticism (pp. 631–632)#

• Asymmetric scepticism. Seeking the “‘full’ truth” makes science “vulnerable to purported weaknesses”. Some scientists apply rigour mainly “when judging the work of colleagues”, earning a “halo” from admiring peers (p. 631).
• Early warnings, being “innovative and necessarily tentative”, suffer most. Valid hypotheses often come from “clinicians, factory inspectors, workers, anglers, bee keepers and community members”, confirmed “much later” (p. 631).
• Footnote 4 (p. 631), ozone. “Joe Forman” [sic: Joe Farman, British Antarctic Survey] doubted his ground-based readings because satellite data disagreed. He reportedly returned “three times” before, “under pressure from his funding sources”, he reported. The name is wrong, and this compressed second-hand account needs checking against the LL1 halocarbons chapter and the primary history. It is also unclear which way the funding pressure pushed.
• Hill (1965), quoted on p. 632: incomplete science “does not confer upon us the freedom to ignore the knowledge we already have, or to postpone the action that it appears to demand”. Some researchers “may mistake the validity of their own conclusions for meticulousness in identifying presumed violations of the causal criteria”.
• “Sound science”, as praised by “special interest groups”, supports “conclusions that are considered attractive” and “is sometimes actually less” reliable. The example is the “harsh critique” of “so‑called black‑box epidemiology” (Taubes, 1995). “Some of that exaggerated critique is echoed in the chapters of the present volume” is ambiguous: it could mean the case chapters document such critique, or it could be a veiled criticism of some LL2 content (p. 632).
• Expert committees “stress the preponderance of uncertainties”, and methodological focus “may be coupled with blindness to environmental degradation and social injustice”. Vested interests exploit this to manufacture doubt (Michaels; Oreskes and Conway) (p. 632).
• “Good Epidemiological Practice” guidelines were “at first embraced by researchers as a useful tool”, but “It turned out that the initiative originated with industry groups in order to disqualify unwelcome ‘junk science’” (Ong and Glantz, 2001; LL2 Ch 7 on tobacco). Rigour became “an unrealistic requirement for repetitive, controlled studies that could furnish virtual statistical certainty” (p. 632).
• Footnote 5 (p. 632), the two-fold rule. Only a doubling of risk would be “believable” (power lines and childhood leukaemia; passive smoking and heart disease). Yet 1.9 and 2.1 are not meaningfully different, and a doubling of a common disease has large impact. Not in the chapter: doubling of risk is also a toxic-tort legal heuristic, which helps explain why the rule persisted.

26.7 Statistics and confidence limits (pp. 632–635)#

• The 5% limit. Fisher’s threshold, first used in agricultural designs, is “almost sacrosanct”. p = 4.9% and p = 5.1% are treated differently (Holman et al., 2001) despite “no meaningful difference” (p. 632).
• The null is often implausible. “Would we ever be tempted to conclude that lead is not toxic, just because a small study has resulted in a p value that is greater than 5 %? Of course not.” Frequentist analysis “considers the data in isolation” (Goodman, 2008). “early warnings are often initially not statistically significant, such as the first IARC study of passive smoking, but nevertheless turned out to be robust” (p. 633). Studies that are individually non-significant may reach significance when pooled by meta-analysis (p. 633).
• Bonferroni correction for multiple comparisons can also be “used erroneously to disregard an unwelcome study” (Perneger, 1998) (p. 633).
• Bayesian updating is presented fairly. It uses all data, but has “serious mathematical complications”, priors “may be impossible to obtain”, and it is criticised as “subjective and overly laborious”. Its use is “gaining support” (p. 633).
• Confidence intervals (Thompson, 1987; some journals reject p values, Lang et al., 1998). “In a precautionary setting, the upper limit would often represent a plausible worst case scenario that would serve as a useful basis when considering intervention” (p. 633).
• Technical note. The chapter defines the CI as the range “within which 95 % of averages would fall if a large number of similar studies were conducted” (p. 633), a standard misinterpretation. The frequentist property is that 95% of intervals so constructed contain the true value. The p-value description (“The probability that their results are significant”, p. 632) is also imprecise. The practical argument survives.
• Figure 26.1 (p. 634). Two studies share a point estimate. The imprecise one crosses zero (not significant) but “cannot exclude a large effect”. The precise one is significant but “can exclude the presence of a large effect”. “Both of these perspectives are relevant, for both studies.” Upper-limit thinking “would inspire larger studies”. The frequentist reading gives the small study no further attention; the precautionary reading calls for follow-up. “a narrow focus on p values should be avoided” (Stang et al., 2010) (p. 634).
• Rare outcomes and small populations (p. 634). Taylor and Gerrodette (1993): “a population might go extinct before a significant decline could be detected”. Taylor et al. (2007): undetected “precipitous declines” run at “72 % and 90 % for various whale species and 55 % for polar bears”.
• Verification (Crossref abstract, checked). A precipitous decline is a 50% fall over 15 years. It would go undetected for 72% of large-whale, 90% of beaked-whale and 78% of dolphin/porpoise stocks, and 55% of polar bear and sea otter stocks. The chapter’s gloss, “more than half of the world’s polar bears … would have to disappear”, is a misreading. The percentages are shares of stocks where a halving would be missed, not fractions of animals lost. The correct reading supports the point just as well. The chapter’s first sentence (“the percentage of precipitous declines that would not be detected”) is broadly accurate. The same abstract gives 5% for pinnipeds on land and 100% for pinnipeds on ice, which the chapter does not report. They show that detection failure depends on monitoring effort, which fits the chapter’s call to expand underpowered protocols.
• Monitoring and power (pp. 634–635). Patchy monitoring means “we are probably overlooking adverse effects, even those that are serious”. An underpowered protocol should be “disbanded” or expanded (p. 635).

26.8 Bias in research (pp. 635–636)#

• Language. Inconclusive studies are called “negative” and treated as “‘no risk’, rather than ‘no information’”, which amounts to “inherent biases toward the null hypothesis” (p. 635).
• Table 26.4 (p. 635), revised from Gee (2009) and Grandjean et al. (2004). Checked visually, since the extract loses the grouping.
• False negative (10): inadequate power; lost cases and inadequate follow-up; exposure misclassification; imprecise outcomes; adjusting for confounders measured more precisely than the exposure; failing to adjust for opposite-direction confounders; disregarding vulnerable subgroups; the 5% Type I level; the 20% Type II level; “pressure to avoid false alarm”.
• False positive (3): incomplete adjustment for same-direction confounders; post hoc hypotheses; publication bias toward positive findings.

The conventional 5% false-positive limit and 20% false-negative tolerance (80% power) together tilt designs toward missing real effects. (Layout re-checked in the PDF this pass: the “False negative” label is centred on the first ten rows and “False positive” on the last three.) The text says “Two entries in the table refer to the possible existence of publication bias” (p. 635). Only one is labelled as such. The other is presumably “Pressure to avoid false alarm”, on the false-negative side. - Key question (p. 635). “How large an effect can the study have overlooked?”, including subgroups and long-term effects. Methodology matters (“greater attention to methodology would be beneficial”), but so does extracting all the information in the data. - Publication bias (pp. 635–636). The chapter concedes it is “quite likely” that some journals, “and more often the mass media”, prefer scares (Ioannidis, 2008), though the bias may run the other way (Oreskes, 2004). It then argues that its topic data “suggest the opposite”, because journals mostly cover familiar chemicals “where new scares are rare”. This answers a different question. Topic-level attention says nothing about whether findings within a topic are reported selectively in the alarming direction. The argument therefore sidesteps rather than rebuts the Ioannidis concern. It adds: “Chapter 2 on false positives shows that erroneous alarms are fairly rare” (p. 636). - Summary (p. 636). Balance “the context of justification … with the context of application”. “While polishing the same stone over and over again, we should not ignore all the other shingles and rocks where some scientific gems may yet be hiding.”

26.9 The changing research paradigm (pp. 636–638)#

• CUDOS (p. 636). Communalism, Universalism, Disinterestedness, Originality, Scepticism, attributed to Merton. Academic “preoccupation with publication, credentials and funding … can lead to social apathy, thereby providing fertile ground for dependence on narrow interests that may include corporate money”. Attribution note: Merton’s 1942 norms were communism, universalism, disinterestedness and organised scepticism. The acronym, and “Originality”, are generally credited to Ziman. “Merton, 1973” (Table 26.5) is missing from the references.
• Vested interests (p. 636) “of whatever origin” reduce reliability. Case chapters show industries that “withheld evidence, lambasted whistle‑blowers and promoted research that supported the conclusions desired”. “complete elimination of financial ties may be the best way to secure trust‑worthy research” (Krimsky, 2003). The chapter concedes that academic conflicts exist and that academic agendas differ from regulators’.
• Contract research (p. 636) may be secret or unpublished, so it “will not inspire further studies”. BPA (952 publications) “was said to be safe at the very low exposures that consumers were likely to receive”, until, “after several decades of expanding use, independent research eventually uncovered evidence of health risks” (Myers et al., 2009). With perfluorinated compounds, “a major US producer for decades claimed that little would escape into the environment, and that essentially no toxicity occurred”. “Only recently” (Grandjean et al., 2012, the author’s own JAMA vaccine-antibody study) did it emerge that “current exposures may be far from safe”. These chemicals “cannot be recalled”.
• PLACE (Ziman): Proprietary, Local, Authoritarian, Commissioned, Expert (pp. 636–637). The chapter adds that “The same characteristics may apply to contract research carried out with public funding, but the initiator may not always be apparent” (p. 636). CIAR, EPRI and the Chlorine Council “may sound like charitable donors, rather than industry front groups. But they are in fact organisations funded by corporations with vested interests in the research outcome”, so readers may wrongly think the research “reflects CUDOS values” (p. 636). Industry funding is accurate for all three. Treating them all as “front groups” is more contestable: the label is well documented for the tobacco-funded CIAR, less obviously apt for a utility-funded research institute such as EPRI.
• Funding shapes topics (pp. 636–637). Little research on pesticide risks or alternatives (Krimsky). Booster biocides (Ch 12): Diuron 389 links, dichlofluanid 39, “some of them much less than the organotin compounds (see tributyltin in Table 26.1) that have been phased out”. Gaucho® (Ch 16): 133 links, “the pesticide that endangered bee populations”. “much less attention is paid to adverse effects of new technology than to its advantages, although this has recently changed in regard to mobile telephony” (asserted without citation; LL2 Ch 21 covers mobile phones, but the chapter does not cite it here). Off-patent drugs are under-researched (Washburn, 2005).
• Intimidation (p. 637). Needleman “was angrily persecuted and harassed with unfounded accusations of dishonesty” (Needleman, 2000). The chapter notes that “Disagreement usually focuses on the uncertainties and the scientific inference, not the choice of study topic”, with angry accusations of bias (citing Gori, 1996, an RTP paper). But research with abatement implications “usually receives more wrath”, and “Perhaps this is another key as to why researchers favour well‑known hazards”. This links back to the inertia argument in 26.4.
• Captured outlets (p. 637). “trade magazines disguised as scientific journals”: Indoor and Built Environment (Tong et al., 2005) and Regulatory Toxicology and Pharmacology (Axelson et al., 2003, a letter the author co-signed). Such outlets “argue for ‘no risk’ when the evidence is uncertain”. Sponsored drug studies favour sponsors (Jorgensen et al., 2006), and “the same seems to happen in toxicology” (Myers et al., 2009). “public trust is abused by deceit.”
• Hedging (p. 637). “incessant use of words, such as ‘maybe’, ‘perhaps’” protects researchers. Lay readers hear that results “do not prove anything”, and interested readers use “selective quotation”. Footnote 6 notes the author’s own hedging.
• PATIO (pp. 637–638). Contract science often “pos[es] like independent, basic research”. “If all research today earned CUDOS, no matter its funding, there would be little to worry about.” PLACE “needs to be supplemented” by a complementary paradigm. For PP-based decisions the chapter proposes Participatory (involving stakeholders), Accessible, Transparent, Inventive, Open-minded research. It concedes that the attributes “may perhaps not be compared horizontally in Table 26.5” (p. 637).

Table 26.5 (p. 638), “Main properties of research in three different settings”:

“‘Industrial’ science is driven by private or other special interest and may violate some of the CUDOS norms in its pursuit of knowledge within fields of commercial or other defined interest, where public interest may be ignored.” Sources: (a) Merton, 1973; (b) Ziman, 2000. PATIO has no cited source and appears to be the author’s own coinage. - Uncertainty as the object of study.* Treat uncertainty “as a normal condition that needs to be explored and addressed, rather than minimised”, and include the public in deciding how it should affect decisions (p. 637). The PP “does not inspire repetitive verification”. But “If available, replication will be useful”, and “a hypothesis may well be plausible even in the (temporary) absence of supportive evidence” (p. 638). Research should “document the extent of uncertainty” and narrow it toward risk assessments “that will no longer need to be precautionary”. “research into uncertainty [is] a very urgent need” (p. 638).

26.10 Precautionary science (pp. 638–640)#

• PP-based question (p. 638). The PP “does not specifically demand testing of a null hypothesis”; “information is required whether a hazard could potentially be serious”. The rephrasing is attributed to Neutra (2002), the California EMF evaluation. It asks whether exposure “leads to doses of a magnitude that can result in adverse effects that are serious enough to initiate transparent and democratic procedures to decide on appropriate intervention?”
• Measurement error. Standard methods assume exposure is measured precisely, which leads to “underestimation of a hazard”. “While uncertainties may be erroneously thought to cause exaggeration of alleged risks, most often the opposite is true.” “One or more worst‑case scenarios deserve as careful scrutiny as the null hypothesis” (p. 638).
• Weakness is not absence. “the mere occurrence of some scientific weakness does not prove the absence of a risk”. Methodological inconsistencies “have been used to derail conclusions otherwise adopted by the scientific community”, and “acceptance of the null hypothesis has sometimes been interpreted as proof of safety” (p. 638).
• Distribution. “a population‑wide shift in the distribution may represent substantial harm”. Harm to populations at risk “can be diluted by the results of non‑vulnerable groups” (pp. 638–639).
• What could the study detect? “what could possibly be known, given the type of evidence available?” Crude studies detect “only the most serious risks” (p. 639).
• Provisionality. “all conclusions must be accepted as being provisional and temporary.” Ioannidis shows that many conclusions later prove wrong, but “this does not mean that environmental hazards are exaggerated”. With most chemicals poorly documented, neglect implies “a very large number of false negative conclusions” (p. 639). The described study sounds like Ioannidis’s 2005 JAMA paper, not the cited 2008 one.
• Access. Paywalls and 6–12 month embargoes still limit access. Open access is growing: the EC’s 12-month recommendation, Wellcome (spelled “Welcome”), Dutch repositories. “So in regard to the Participatory aspect of the PATIO paradigm, access to information is improving” (p. 639). Here participation is discussed only as reading access. Elsewhere the chapter defines Participatory as involving stakeholders (p. 637), and recommendation 1 involves them in topic choice. How that involvement would work is not spelled out.
• Data sharing. Even preliminary data can feed later meta-analyses and follow-up studies, if the data are available (p. 639). Trade secrets and “suppression of information and withholding of evidence” block access (Kurland, 2003). “Some public funding agencies now demand” data-sharing strategies for major projects, but “Hostile analyses have occurred, thus making researchers wary with whom they share their raw data” (Pearce and Smith, 2011) (p. 639). The tension is noted but not resolved.
• Four recommendations (p. 639): 1. “The choice of research topic should involve stakeholders and consider the societal needs for information on poorly known hazards”; 2. “The research should be innovative and complementary with the aim of extending current knowledge, rather than repetitive for verification purposes”; 3. “The findings should be communicated in such a way as to facilitate judgements concerning the possible magnitude of suspected environmental hazards”; 4. “The research should be openly available and independent of vested interests.”
• Close (p. 640). Science “does not provide a prescription for the right decisions”. Under precaution, evidence “does not have to meet the most rigorous demands of science”, and “world views, political and other preferences, technical and economic feasibility, and alternative options are crucial”. The final verdict: “science does not have a good track record for supporting decisions on improving environmental health.”

References (pp. 640–642)#

About 50 references, spanning statistical methodology, sociology of science, the literature on manufactured doubt and sponsorship bias, regulatory reports, and six of the author’s own papers. Minor errors: “EC, 2001” in the list against “EC, 2011” in the text; Merton 1973 missing; Neutra et al.’s co-authors garbled.

Case timeline#

This is not a case study, so there is no single hazard timeline. Below are the dated evidence and events the chapter uses. Where a date is not given in the chapter, that is stated.

Lags the chapter identifies: - Research does not respond to a regulator’s priority list for at least four to five years (2006 to 2011) (pp. 627–628). - Recognising harm at lower doses comes long after the original limits were set, as with ALAD and lead (pp. 624, 631). - Industry claims about BPA and perfluorinated compounds lasted “decades” before independent evidence emerged (p. 636). - Monitoring can miss a “precipitous” decline (p. 634). Per Taylor et al. (2007), not the chapter, that means a halving within 15 years.

The authors’ own lessons and conclusions#

Lessons derived from evidence the chapter presents (with the strength of that evidence): 1. Academic environmental research is heavily concentrated on a small number of long-studied substances, and data-poor, regulator-flagged chemicals get almost no attention (pp. 625–628). This rests on original bibliometric data. 2. Initial assumptions in hazard assessment were repeatedly wrong, and in the same direction, delaying recognition of harm (Table 26.3, p. 630). This is compiled from the case studies with hindsight. 3. Conventional study design and statistical conventions are biased toward false negatives (Table 26.4, p. 635). This rests on methodological reasoning plus citations. 4. Early warnings are often statistically non-significant at first, yet prove robust (IARC passive smoking, p. 633). One example is given. 5. Vested interests have used methodological critique, captured outlets, front organisations and intimidation to manufacture doubt (pp. 632, 636–637). This rests on cited secondary literature and case chapters. 6. Monitoring can fail to detect catastrophic declines (whales, polar bears, p. 634). This rests on cited primary studies.

Explanatory claims (mechanisms proposed, largely untested in the chapter): 7. The inertia comes from replication norms, career metrics, infrastructure, funder risk aversion, peer review and prestige, and mentoring (pp. 628–629). The chapter offers these as hypotheses: “may”, “perhaps”. 8. Research with pollution-abatement implications draws more hostility, which pushes researchers toward safe topics (p. 637). Offered as “Perhaps”.

Recommendations and advocacy: 9. Reframe the research question around possible magnitude and the threshold for initiating democratic decision procedures (pp. 623, 638). 10. Report confidence intervals and upper limits, run worst-case and sensitivity analyses, use power analysis, and ask “how large an effect can the study have overlooked?” (pp. 633–635, 638). 11. Treat uncertainty as an object of research, not a nuisance to be minimised (pp. 637–638). 12. Adopt a PATIO research culture alongside CUDOS and PLACE (pp. 637–638). 13. The four recommendations: stakeholder topic choice; innovative rather than repetitive research; communication of magnitude; open and independent research (p. 639). 14. Eliminating financial ties may be better than disclosing them (p. 636, citing Krimsky). 15. Precautionary decisions should be provisional and evaluated by intervention studies (pp. 624, 639).

Overarching verdict (p. 640): “science does not have a good track record for supporting decisions on improving environmental health”. This is an assertion drawn from a set of cases selected because action came late (see bias check).

Mechanisms and dynamics#

The section notes above carry the detail. This section pulls out the causal structure.

1. Research agendas lock in (path dependence). Several feedback loops reinforce each other: cheap, validated 1970s–80s instruments (p. 628); publication-count careers and budgets (p. 629); funders who prefer predictable projects (p. 629); easy peer review and reviewer self-citation (p. 629); prestige attached to well-studied substances (p. 629); expertise reproduced through mentoring (p. 628); extra hostility toward regulatory-relevant findings (p. 637). Together they produce a Matthew effect (p. 629). An explicit regulatory priority list did not move output (pp. 627–628). The notes’ inference, not a claim the chapter states or tests: governance signals may not reach agenda-setting unless incentives change. The cost falls on discovery as well as protection (p. 628). The chapter offers these loops as hypotheses (“may”, “perhaps”), not measured effects.

2. Standards of proof produce false reassurance. The traditional question puts the burden on demonstrating harm, at a high standard (p. 630). A “no effect” null, a 5% false-positive limit against a 20% false-negative tolerance, exposure imprecision and averaging then tilt results toward the null (Table 26.4, pp. 635, 638). Lack of proof becomes proof of safety, “the greatest error” (p. 631). Risk assessment that demands convincing evidence becomes “anti‑precautionary” and keeps refining known hazards (p. 625). The bias is structural: well-meant rigour produces it without any bad faith.

3. Rigour is turned into a weapon. Interested parties exploit the same virtues: “sound science” and “junk science”, “Good Epidemiological Practice”, strict readings of causal criteria, the two-fold rule, Bonferroni corrections, expert committees stressing uncertainty, and selective quotation of hedges (pp. 631–633, 637). The tell is asymmetric scepticism, and scientific culture rewards the critic’s “halo” (p. 631).

4. Interests shape both topics and conclusions. Industry funds more than half of all EU R&D activity (p. 624, Eurostat). That figure covers all R&D, much of it industry’s own in-house work, and the chapter uses it mainly to argue that the public share is large enough that academic research should reflect public priorities. It is not direct evidence about who funds environmental health research. Proprietary research stays secret (p. 636). Harms get less study than benefits, and alternatives and off-patent options get little (pp. 636–637). Sponsored studies favour sponsors (p. 637). Neutral-sounding funders and friendly journals blur where research comes from (pp. 636–637). Warners are intimidated (p. 637). The chapter’s remedy is to eliminate financial ties rather than disclose them (p. 636). It also names a non-financial misalignment: academic agendas differ from regulators’ needs (p. 636).

5. Mental models and language. The blind spots include a reductionist study design (p. 630), the healthy adult male (or inbred male rat) as reference subject, protective barriers, monotonic thresholds, dismissal of animal data, and biomarker changes deemed “not adverse” (Table 26.3; p. 631). Words do the work: “safe” and “natural” doses, “negative” studies, “no risk” for “no information” (pp. 631, 635), “sound” and “junk” science (p. 632), hedges heard as “nothing proven” (p. 637), benign-sounding names (p. 636). Experts’ self-image as guardians of rigour can hide “blindness to environmental degradation and social injustice” (p. 632).

6. Whose knowledge counts. Early warnings often come from clinicians, inspectors, workers, anglers, beekeepers and communities (p. 631), so hyper-scepticism hits the earliest detectors hardest. PATIO’s “Participatory” element (stakeholder involvement, p. 637; recommendation 1, p. 639) answers this. But the only concrete progress the chapter reports under that heading is open access (p. 639), and it does not say how lay observers’ warnings would be taken up.

7. Time, irreversibility and distribution. Developmental and multigenerational effects need long, costly studies that industry resists (p. 631). Dispersed persistent chemicals “cannot be recalled” (p. 636). A population can halve before monitoring can detect it (p. 634). Averages hide harm to vulnerable groups, and small population-wide shifts add up to large harm (pp. 638–639). The burden of scientific uncertainty falls on the exposed, especially children and the fetus.

8. Institutions and governance. Universities behave as enterprises (p. 624). Grant managers minimise their own risk and journals chase prestige (p. 629). Publishers restrict access (p. 639). No single actor needs to act in bad faith for the system to neglect emerging hazards. Science’s proper role is to judge whether evidence warrants triggering “transparent and democratic procedures” (pp. 623, 638). The decisions then turn on values, feasibility and alternatives (p. 640) and should stay provisional (p. 624). Footnote 2 (p. 624) concedes that political will often matters more than information. Openness cuts both ways: it enables meta-analysis and it enables “hostile analyses” (p. 639).

Transferable insights (technology-neutral)#

1. Research attention is path-dependent and self-reinforcing. It tends to stay on what is already well studied, even when regulators explicitly flag other priorities. Evidence: pp. 625–628; Tables 26.1–26.2; EPA list of 13 with 352 links, 5 absent, no change by 2011. Strength: strong for the descriptive pattern in 2000–2009. It rests on original bibliometric data, and I verified the headline numbers against the 2011 paper. It is limited to 78 academic journals and to article counts, not content. The 2010–2011 “no uptick” claim is reported without figures (the 2011 paper gives only a 2010 count).

2. Signalling a priority is not enough to redirect knowledge production. Incentives (metrics, funding, infrastructure, reviewer familiarity) have to change. Evidence: pp. 627–629. Strength: moderate. The non-response to the EPA list is measured, but the causal mechanisms are argued and hedged (“may”), not tested.

3. Absence of evidence gets converted into evidence of absence, through vocabulary (“negative”, “safe”, “natural”) as well as statistics. Evidence: pp. 623, 631, 635, 638; ALAD and “natural” lead, p. 631. Strength: strong as logic, and the lead examples are well documented. How often it happened across cases is asserted from the case studies, not counted.

4. Conventional study designs and error-rate conventions are tilted toward false negatives. The asymmetry is built into standard practice, not only into bad faith. Evidence: Table 26.4, p. 635; pp. 631, 638. Strength: moderate to strong. Non-differential exposure error biasing toward the null and the 5% versus 20% convention are standard epidemiology. But the table is a list, not a quantification, and some biases that produce false positives (selective reporting, flexible analysis) are under-represented.

5. Ask how large an effect a study could have missed. Use upper confidence limits and worst-case or sensitivity analyses as decision inputs, alongside tests of “no effect”. Evidence: pp. 633–635, 638; Figure 26.1. Strength: strong as methodological advice. This became mainstream in statistical reform after 2013, though the chapter defines the CI imprecisely (p. 633).

6. When outcomes are rare or populations small, statistical significance may be reachable only after irreversible harm. Detection power must be considered before “no significant change” is read as reassurance. Evidence: p. 634 (Taylor & Gerrodette 1993; Taylor et al. 2007); p. 635 (power analysis). Strength: strong in principle and in the verified source data, but the chapter misparaphrases the whale and polar bear figures.

7. Rigour can be turned into a weapon. Demanding ever-higher proof, strict causal criteria, arbitrary thresholds (such as a two-fold risk) or multiple-comparison corrections can be used to dismiss unwelcome findings. Scepticism applied asymmetrically is a sign of this. Evidence: pp. 631–633; fn 5. Strength: strong for the documented tobacco case (Ong & Glantz). Moderate as a generalisation, since the chapter gives few non-tobacco examples in its own text.

8. Early warnings often come from practitioners and lay observers, and are tentative and underpowered, so hyper-scepticism suppresses them disproportionately. Evidence: p. 631; IARC passive smoking, p. 633; the Farman footnote, p. 631. Strength: moderate. The chapter asserts this from the case studies of both volumes but does not count cases, and the ozone footnote is garbled.

9. Early assessments tend to rest on convenient default assumptions: tolerable “safe” doses, a healthy adult reference subject, averages, protective barriers, monotonic thresholds, irrelevance of other-species evidence. These fail in the same direction. Evidence: Table 26.3, p. 630; p. 631. Strength: moderate to strong. Each item is illustrated in case chapters, but the table is compiled with hindsight and lists only assumptions that failed.

10. What is “normal” is not necessarily “natural” or safe. Baselines can themselves reflect accumulated exposure. Evidence: p. 631 (lead isotopes, mummified tissues). Strength: strong for lead. Moderate as a general heuristic.

11. Who funds research shapes both what is studied and what is concluded. Proprietary research may be secret, and harms get less study than benefits. Evidence: pp. 636–637; Diuron/dichlofluanid/Gaucho counts, p. 636; Jorgensen et al., Myers et al., p. 637. (The Eurostat R&D figure on p. 624 covers all R&D and is not evidence for this point.) Strength: strong for pharmaceutical sponsorship bias (Jorgensen). Moderate for environmental toxicology. Suggestive for the benefits-versus-harms claim, which is asserted.

12. Neutral-sounding intermediaries and friendly outlets can make interested research look like independent science. Evidence: pp. 636–637. Strength: moderate. Well documented for tobacco (CIAR, Tong et al.), but the chapter lumps together very different bodies and uses advocacy language.

13. People who produce findings with regulatory consequences draw disproportionate hostility, which pushes the research community toward safe topics. Evidence: p. 637 (Needleman). Strength: moderate for the harassment itself (documented case). Suggestive for the effect on topic choice (“Perhaps”).

14. Scientific hedging protects researchers but reads to the public as “nothing proven”, and can be quoted selectively by interested parties. Evidence: pp. 623, 637. Strength: suggestive. It is plausible and the author illustrates it himself (fn 6), but no evidence on reception is given.

15. Delay is costliest when agents are persistent and dispersed, because exposure cannot be recalled once it has spread. Evidence: p. 636 (perfluorinated compounds). Strength: strong as a principle. The chapter’s own support is a single sentence on one example, and later evidence on PFAS strongly supports it (see hindsight leads).

16. Early standards set on incomplete evidence tend to be too lenient and to ratchet down as evidence accumulates. Evidence: p. 624; ALAD and lead, p. 631. Strength: moderate. Widely observed, but “nearly all exposure limits” is asserted without data.

17. Averages hide concentrated harm, and small shifts across a whole population can add up to large harm. Evidence: pp. 638–639; Table 26.3 #2 and #8. Strength: strong as standard population-health reasoning. The chapter asserts it without a worked example (lead would fit, but it is not used here).

18. The evidentiary question for governance differs from the one for scientific proof. It is whether possible harm is serious enough to trigger a transparent, democratic decision process, and decisions taken on that basis should be provisional and evaluated. Evidence: pp. 623, 624, 638–640. Strength: asserted. This is a normative proposal (drawing on Neutra et al. 2002). It is coherent, but the chapter offers no evidence that this framing produces better outcomes.

19. Transparency and data sharing cut both ways. They enable accountability, meta-analysis and follow-up, and they also enable hostile reanalysis. Evidence: p. 639 (Pearce and Smith, 2011); p. 636. Strength: suggestive in the chapter itself, which gives one line and one citation. Later use of “transparency” requirements to exclude evidence is not in the chapter. It is a hindsight lead that may strengthen the point (see below).

20. The system that produces knowledge, not only individual researchers, creates blind spots. Researchers, funders, journals, expert committees and publishers each act sensibly on their own terms and together neglect emerging hazards. Evidence: pp. 624, 628–629, 632, 639. Strength: moderate. The synthesis is persuasive, but mostly argued from plausibility.

Limitations, contestation and bias check#

Genre and balance. - This is a single-author position essay by a member of the report’s editorial team, with no panel or dissenting commentary (pp. 3–5, 623). - The empirical core is the author’s own bibliometric study, published in his own journal (Annex 1, p. 690). - The chapter criticises “inappropriate (or apparent) advocacy” (p. 624) and is itself partly advocacy. That isn’t hypocrisy, but it shapes the tone. The chapter does mark its reasoning as hedged, and says so (fn 6, p. 637).

What the bibliometrics can and cannot show. - Counts of CAS links in 78 academic journals do not measure depth, novelty or whether a paper studies hazards at all. The 2011 top 20 includes ethanol and essential trace metals (copper, zinc, iron, manganese), whose literature spans nutrition and geochemistry as well as toxicity. They also miss regulatory, industry and grey literature. The chapter concedes most of this (p. 628), and the 2011 paper states that “the depth of our study is based solely on the mere number of articles located”. - The chapter’s own examples cut against its “over-replication” charge. Continued work on lead is how harm at lower exposures, and the lowering of limits, came to be recognised (pp. 624, 631). The chapter concedes that research on well-known substances produced breakthroughs (p. 629) but does not reconcile this with “polishing the same stone” (p. 636). A fair reading: heavy study of known hazards may be both partly wasteful and the source of the low-dose findings that precautionary advocates rely on. - The perfluorinated compound count is attached to a garbled chemical name (p. 628). The triclosan and PFC numbers are not in the 2011 paper, so they cannot be checked against a published source.

Replication and the false-positive side. - The call to favour “complementary” over “repetitive” research (pp. 623, 639) was written just before the replication crisis became widely recognised (e.g., Begley & Ellis 2012; Open Science Collaboration 2015). The chapter’s point concerns substances studied thousands of times, and it does qualify it: “a single study should not be relied upon as firm evidence” (p. 628), and “If available, replication will be useful” (p. 638). Even so, the general phrasing of recommendation 2 could be read as undervaluing replication, which later evidence showed to be too scarce in many fields. - The false-positive side is underweighted: - Table 26.4 lists 10 false-negative drivers against 3 false-positive ones. Flexible analysis is only partly covered by “post hoc hypothesis”, and selective outcome reporting and small-study effects are omitted. - The claim that uncertainties “are not likely to create spurious associations, unless confounding factors are present” (p. 631) is too strong: differential misclassification and selection bias can also create associations. - The chapter concedes that journals and media “quite likely” favour scares, then sets the concern aside with topic-count data. Those data do not speak to the direction of findings within topics (pp. 635–636). - The chapter relies on Ch 2 for “erroneous alarms are fairly rare” (p. 636). Ch 2’s conclusion depends on how it defined and selected false positives; cross-check with the LL2-02 notes. - Ioannidis is cited but his point is contained rather than engaged with (p. 639).

Statistical accuracy. The critique of p-values is sound and was mainstream among epidemiologists by 2013 (Goodman, Greenland, Stang, Lang all cited), but the chapter’s own definitions are imprecise: - the p-value as “the probability that their results are significant” (p. 632); - the CI as the range containing “95 % of averages” (p. 633); - the misparaphrase of Taylor et al. (p. 634).

None of these invalidates the argument, but they matter in a chapter about misinterpreting statistics. On Bayesian methods the chapter is balanced, and it lists their drawbacks (p. 633).

Hindsight and case selection. - Table 26.3 lists only assumptions that later failed. Default assumptions that held up (real thresholds, for instance, or animal data that over-predicted human risk) are not shown. - “Science does not have a good track record” (p. 640) generalises from cases selected for late action. In several of those cases, lead and ozone among them, science eventually drove action. The chapter’s own opening grants that the case studies show science “can provide powerful evidence for targeted prevention” (p. 624). - Footnote 2 (p. 624) itself says political will mattered more than information in LL1’s cases, which would place the main failure outside science.

Vested interests. The analysis concentrates on corporate interests. It gestures at “vested interests of whatever origin” and at academic conflicts (p. 636), but does not examine incentives that could inflate hazard findings: advocacy, novelty-seeking in high-profile journals, media attention. That is the direction a critic of precaution would press. Some labels are advocacy rather than analysis: - EPRI grouped with CIAR as an implied “industry front group” (p. 636). The chapter’s factual claim, that all three bodies are industry-funded, is accurate; the “front group” framing is the contestable part; - RTP and Indoor and Built Environment as “trade magazines disguised as scientific journals” (p. 637), where the chapter’s supporting citation for RTP is a letter the author co-signed.

The substantive concern about sponsor influence is well founded. The characterisations are contestable.

Internal tensions. - Openness is advocated (recommendation 4) while the chapter acknowledges “hostile analyses” (p. 639). - “Participatory” is defined as stakeholder involvement (p. 637), but the only progress reported under it is open access (p. 639). The mechanism of participation is left undeveloped. - Research agendas “should” follow regulators’ and stakeholders’ needs (p. 639), yet the chapter values originality and innovation. Tying academic agendas to regulatory lists could create its own narrowing and loss of independence, which the chapter does not discuss. - PATIO is aspirational. No example of PATIO research, and no evidence that it improves decisions, is offered.

Factual and citation errors (none central to the argument, but worth knowing): - “Joe Forman” should be Joe Farman, and the footnote’s account needs checking (p. 631 fn 4). - “perfluorinated octanoic sulfate” (p. 628). - CUDOS attributed to Merton alone (p. 636). - “Merton, 1973” missing from the references (p. 638). - The Ioannidis finding on p. 639 probably comes from the 2005 JAMA paper, not the cited 2008 one. - “EC, 2001/2011” mismatch; “Welcome Trust”; “MBTE”; Ioannidis 2008 volume printed as “(195)” for 19(5). - Schwartz and Carpenter’s “Type III error” is used loosely (p. 629). - Taylor et al. is misparaphrased (p. 634).

Hindsight leads (post-2013). From general knowledge, not verified in this pass; treat as leads for the hindsight strand. - Perfluorinated compounds. This is probably the chapter’s most strongly vindicated claim. EFSA (2020) set a tolerable weekly intake for four PFAS with reduced vaccine antibody response in children as the critical effect. US EPA’s 2022 interim health advisories for PFOA and PFOS rested on immune effects in children drawing on the Faroese cohort, and the 2024 drinking-water rule set limits of 4 ppt. The main US producer announced its exit from PFAS manufacture (announced 2022, by end-2025), and large water-utility settlements followed in 2023. Check the current status of the US rule after the 2025 reconsiderations. - BPA. EFSA (2023) cut the tolerable daily intake by about 20,000-fold, and the EU banned BPA in food-contact materials (Regulation (EU) 2024/3190). But the German BfR dissented from EFSA’s value, and US FDA (after CLARITY-BPA, 2018) maintained that current exposures are safe. The contest the chapter describes continued, and divergent regulatory assessments persisted. - Triclosan. US FDA removed it from consumer antiseptic washes (2016 rule). - Gaucho/imidacloprid. EU restrictions from 2013 and an outdoor-use ban in 2018. - Statistics reform. The ASA statement on p-values (2016); Greenland et al. (2016) on misinterpretations, which lists the CI misreading the chapter makes; the 2019 calls to retire “statistical significance”. But there was also a counter-proposal to tighten the threshold to p < 0.005 (Benjamin et al. 2018), aimed at false positives. Reform went in both directions. - Open access. Plan S (2018) and the US OSTP public-access memo (2022) moved much further than the chapter anticipated. - Data-transparency demands used as a weapon. US EPA’s “Strengthening Transparency in Regulatory Science” rule (finalised January 2021, vacated by a federal court soon after) was widely criticised as a way to exclude epidemiological studies built on confidential health data. This is a close later instance of the chapter’s “hostile analyses” and “sound science” concerns. Check also later US “gold standard science” policy framing. - Sponsorship bias. A 2017 Cochrane review (Lundh et al.) confirmed that industry-sponsored drug and device studies more often report favourable results and conclusions. - Exposure limits ratchet. US CDC lowered its blood-lead reference value (to 3.5 µg/dL in 2021). EFSA identified no threshold for lead’s developmental neurotoxicity (2010). - Chemical data gaps. Check ECHA’s later evaluation reports on REACH dossier compliance, the US TSCA reform (2016), and whether bibliometric concentration persisted (for example, any update of Grandjean et al. 2011). Also check the EU Partnership for the Assessment of Risks from Chemicals (PARC, from 2022) as a partial institutional answer to recommendation 1.

Notable quotes#

1. “prevention has too often been deferred due in part to the alleged absence of convincing scientific evidence” (p. 624)
2. “traditional risk assessment is sometimes anti‑precautionary when it demands convincing evidence and thus ignores emerging insight and incomplete documentation” (p. 625)
3. “the long‑term prominence of substances commonly covered in articles in environmental journals does not match the societal needs or those of regulatory agencies” (p. 628)
4. “Uncomfortable surprises are unlikely. Hence, it may be safer and more convenient for grant managers to concentrate on the known hazards.” (p. 629)
5. “If adverse effects were not proven to exist, the erroneous conclusion was drawn that adverse effects must be absent. Perhaps this is the underlying assumption, which represents the greatest error.” (p. 631)
6. “In a precautionary setting, the upper limit would often represent a plausible worst case scenario that would serve as a useful basis when considering intervention.” (p. 633)
7. “such studies have sometimes been thought to represent ‘no risk’, rather than ‘no information’” (p. 635)
8. “While polishing the same stone over and over again, we should not ignore all the other shingles and rocks where some scientific gems may yet be hiding.” (p. 636)
9. “Are we sufficiently confident that this exposure to a potential hazard leads to doses of a magnitude that can result in adverse effects that are serious enough to initiate transparent and democratic procedures to decide on appropriate intervention?” (p. 638)
10. “science does not have a good track record for supporting decisions on improving environmental health” (p. 640)

Open questions#

1. Did the concentration of academic research on well-known substances persist after 2011? Did REACH, TSCA reform or programmes such as PARC shift attention toward data-poor substances? Is there an update of Grandjean et al. (2011)?
2. Which perfluorinated compound (PFOA or PFOS) was counted on p. 628, and how were the triclosan and PFC counts produced?
3. What did “Some of that exaggerated critique is echoed in the chapters of the present volume” (p. 632) refer to: documentation in the case chapters, or criticism of particular LL2 chapters or panels?
4. How does the LL1 halocarbons chapter tell the Farman story that footnote 4 compresses? Did funding pressure push toward publication or toward caution?
5. How can recommendation 2 (complementary, not repetitive, research) be squared with evidence from the replication crisis that replication is too scarce in many fields? Where is the line between wasteful re-verification and necessary replication?
6. Is there evidence that the reframed PP question, or confidence-limit reporting, changes regulatory outcomes or timing? Any natural experiments?
7. How should the chapter’s call for open data be squared with its worry about “hostile analyses” and later uses of transparency demands to exclude evidence? What governance designs protect both openness and the integrity of evidence?
8. Tying academic agendas to regulators’ and stakeholders’ needs could itself narrow science or compromise its independence. What safeguards would PATIO need?
9. How does Ch 2’s finding that false alarms are rare hold up, given its definitions and case selection? The chapter’s dismissal of false-positive concern leans on it (p. 636).
10. Is the verdict of a “poor track record” (p. 640) consistent with footnote 2 (p. 624), which places the main failures in political will rather than information? Which failures belong to science and which to governance?

Audit log#

Independent audit, 2026-09-26. I re-read the whole extract (pp. 621–642) against the notes. I re-checked the Table 26.4 layout, the contents (Ch 21, 24, 27 authors), the Acknowledgements (editorial team) and Annex 1 in the PDF. I also checked Grandjean et al. (2011) (PMC3229577) and the Taylor et al. (2007) abstract (Crossref). Table figures, the four recommendations and the main quotes were confirmed verbatim.

1. Summary box: restored the hedge “seems to focus” and the missing point that misinterpretation comes from hedged language.
2. 26.1: added “proof or a very high degree of certainty are not required”, the accepted “inevitable” uncertainties, and the author’s “I think that the answer is yes” (p. 625).
3. Table 26.1 bullet: added the chapter’s own exceptions (SO2 and DES “clearly faded”, MTBE rose), “asbestos not included” for the 8,267 SciFinder count, and corrected the page to p. 626.
4. Top-20 bullet: flagged that 14,264 papers rests on the chapter’s stated assumption that 12% of links equals 12% of articles.
5. Composition bullet: corrected an error. The notes had called 1,1’-biphenyl a “general chemistry” artefact, but the 2011 paper gives it the same count (3,897) as its PCB entry, so it is the PCB proxy. Replaced it with the verified point about essential trace metals and ethanol in the top 20.
6. EPA priority list: added the citation (US EPA, 2009), the chapter’s “perhaps 2007” allowance, and the fact that the 2010–2011 “no uptick” claim has no figures in the chapter (the 2011 paper gives only 34 links for 2010).
7. Triclosan/PFC counts: softened “method undocumented” to “cannot be checked against a published source”.
8. 26.4: corrected the page for “counter-productive” (p. 628) and added the chapter’s concession that single studies are not firm evidence.
9. 26.5: added the chapter’s acknowledgement of the traditional paradigm’s virtues, that Table 26.3 is drawn from the human-health chapters, and the “may” hedging of the table’s lessons.
10. ALAD example: corrected a misrepresentation. Hormesis and adaptation were the general argument, not specific to ALAD. The chapter concedes that ALAD inhibition itself “may be true in a strict sense” to be harmless, and its point is harm at exposures once thought too low.
11. 26.6: attributed “sound science” to “special interest groups”; replaced “Taubes’s attack” with “harsh critique”; added that GEP was “at first embraced by researchers”; changed “Some critics” to “Some researchers may” to match the source.
12. 26.7: added the meta-analysis point (p. 633). Added to the Taylor et al. verification the 5% (land pinnipeds) and 100% (ice pinnipeds) figures the chapter omits, and noted that the chapter’s first sentence is broadly accurate.
13. Table 26.4: confirmed the 10/3 grouping from the PDF layout. Added the text’s “Two entries … publication bias”, with “Pressure to avoid false alarm” as the presumed second entry.
14. Publication bias: softened “non sequitur” to “answers a different question” and added the chapter’s concession that journals and media “quite likely” favour scares.
15. PLACE: added that PLACE traits “may apply to contract research carried out with public funding”, and the chapter’s substantive claim that CIAR, EPRI and the Chlorine Council are industry-funded (accurate). Rebalanced the critique so only the “front group” label is contestable.
16. Funding bullet: added the booster biocide versus organotin comparison. Flagged “(LL2 Ch 21)” as a cross-reference added by the notes, not a citation in the chapter.
17. Intimidation: added the Needleman (2000) citation, the chapter’s concession that disagreement usually targets inference rather than topic, and Gori (1996).
18. PATIO: added “supplemented” (complementary, not replacement), the stakeholder definition of Participatory, and the chapter’s caveat that the attributes “may perhaps not be compared horizontally”.
19. Table 26.5: restored the printed spelling “Skepticism”, the full table note, and the source line (Merton 1973; Ziman 2000).
20. Uncertainty bullet: added “does not inspire repetitive verification” and “If available, replication will be useful” (p. 638).
21. 26.10: added that the PP “does not specifically demand testing of a null hypothesis”. Changed “from Neutra et al.” to “attributed to Neutra (2002)”.
22. Access and participation: softened “equates participation with reading access”. The chapter defines Participatory as stakeholder involvement, and only its reported progress is open access.
23. Data sharing: added the preliminary-data and meta-analysis point, the Kurland citation, and “Some public funding agencies” in place of “increasingly”.
24. Timeline: fixed the Merton row (1942 is not in the chapter; “Merton, 1973” is missing from the references). Marked the GEP “1990s” date as outside knowledge. Added the 2001 LL1 preface row (fn 2). Added the missing-figures caveat to the 2010–2011 row.
25. Lags: attributed the “halving in 15 years” definition to Taylor et al., not the chapter.
26. Mechanism 1: marked “governance signals do not reach agenda-setting unless incentives change” as the notes’ inference, and noted that the loops are hypotheses in the chapter.
27. Mechanism 4 and Insight 11: corrected a misuse of evidence. The Eurostat “more than half of EU R&D” figure covers all R&D and is used by the chapter to argue that public funding is substantial. It is not evidence that environmental research is industry-funded.
28. Mechanism 6: reworded the PATIO “Participatory” point in line with item 22.
29. Insights 1, 3, 15 and 17: adjusted the strength notes (the 2010–11 figures are missing; frequency of the pattern is asserted not counted; the chapter’s support for 15 is a single example; 17 is asserted without a worked example in the chapter).
30. Insight 19: removed “‘transparency’ demands used to exclude evidence” from the insight statement, because it is not in the chapter, and kept it only as a hindsight lead.
31. Limitations: replaced the biphenyl example; softened the triclosan/PFC method point; added the chapter’s own replication qualifications; noted that “post hoc hypothesis” partly covers flexible analysis; softened the publication-bias critique; added the chapter’s p. 624 concession to the track-record critique; rebalanced the EPRI point; reworded the participation tension; added the Ioannidis “(195)” reference typo.
32. Digest: hedged the Matthew-effect attribution. Corrected the whale/polar bear line (50% is Taylor et al.’s definition, and the figures are shares of stocks). Moved the Eurostat figure out of the “not independent” evidence with a qualifier. Changed “set alongside Merton’s CUDOS” to “supplement CUDOS (attributed to Merton)”. Split the funding insight’s strength rating. Qualified the persistence rating. Added the caveats on replication concessions, the trace metals and ethanol in the counts, EPRI, the missing 2010–11 figures and the p. 624 concession. Softened the publication-bias wording.