Source
Douglas (2000), LIT-139 (Philosophy of Science 67(4), 559–579), read in full in NOTE-145. Page references are the journal’s.
What was actually shown
Inductive risk is “the chance that one will be wrong in accepting (or rejecting) a scientific hypothesis” (p. 561, after Hempel 1965). Douglas extends it to accepting methodologies, data and interpretations (p. 565).
Significance level (§4, pp. 565–569). For a fixed test, the significance level trades false positives against false negatives. In dioxin bioassays a false positive leads to overregulation and a false negative to underregulation. Dose groups are about ten-fold apart, so a false negative under the threshold model sets the “safe” dose “at least one order of magnitude less protective than it should be” (p. 568). Reducing both errors needs more power, and one two-year study of 200 rats costs about $3 million (p. 569).
Characterising data (§5, pp. 569–572; Table 1). The same female Sprague-Dawley liver slides from the 1978 Kociba study were scored in 1978, 1980 and 1990. Total tumours at 100 ng/kg/day: 34/50, 33/47, 18/50. At 10 ng/kg/day: 20/50, 27/50, 9/50. Controls: 9/86, 16/86, 2/86. The 1990 panel voted by majority. Borderline calls trade the two errors, and the 1990 re-evaluation led Maine to relax its regulation. Blinding was impossible, because liver toxicity marked the high-dose slides.
Interpreting results (§6, pp. 573–577). With 50 animals per group a 1% excess cannot be detected, so an apparent threshold at 1 ng/kg/day may be “a limit of detection, not a limit in the activity of dioxins”. But every tumour-bearing liver at 100 ng, and 8 of 9 at 10 ng, showed acute toxicity, which makes a threshold plausible if the tumours come from cytotoxic regeneration. Countries assuming a threshold set acceptable levels about 100 times higher than no-threshold countries such as the US (p. 577).
What could have come out otherwise. The panels could have agreed, and the data could have favoured one dose-response reading. Neither happened.
What this does not say
- That every internal choice carries inductive risk that matters. Douglas says “not all scientific decisions in the internal stages of science will require the consultation of non-epistemic values” (§7, p. 577). Uncertainty may be negligible, or error may have no outside consequences.
- That the conventional significance levels are wrong. She sets the conventions aside: “I will not argue that such conventions have not worked reasonably well” (p. 566). The argument is about what a deliberate choice of level would involve.
- That values tell which reading is true. On her account they bear on which error to risk, not on whether evidence confirms a hypothesis (p. 565).
- That the pattern holds outside toxicology. All three cases come from one domain where the policy consequences are unusually direct.
- That rodent results transfer to humans. That extrapolation is set aside (p. 567).
Connections
- The normative claim built on this, that weighing the consequences of error is then required and that values may enter only indirectly, is THEORY-442.
- The majority-vote panel is a case of adjudicating ambiguous labels. The reading draws the analogy to ML evaluation itself; the paper does not (NOTE-145, Bearing).