NOTE-145: Douglas — Inductive risk and values in science

Contribution

Douglas revives the 1948–65 inductive-risk argument (Churchman 1948, Rudner 1953, Hempel 1965 for; Jeffrey 1956, Levi 1962, McMullin 1983 against) and extends it in two ways. First, Hempel located inductive risk only at hypothesis acceptance, and she places it at every internal stage: choice of methodology, gathering and characterisation of data, and interpretation. Second, she answers the Jeffrey–McMullin reply (leave the utilities to users of science) by pointing to the authority of scientists and to risks taken at stages that never reach public decision-makers. Her conclusion is that the “externality” model of values is overthrown: “non-epistemic values are required for good reasoning” (p. 565) whenever inductive risk carries foreseeable non-epistemic consequences. For such science, “value-free science is inadequate science” (p. 559).

Key insight

Values do not tell you what is true. They tell you which errors you should be more willing to risk. Any choice made under uncertainty has two ways of being wrong, and when the consequences of those errors fall outside science, which outside consequences matter more is a value question that cannot be avoided. It is also answered implicitly by whatever threshold, classification rule or background assumption is adopted. Because this happens at internal choice points (α, borderline slides, background assumptions), a value-free science would require scientists to be exempt from moral responsibility for foreseeable consequences of their choices.

Assumptions

Premises and authorities rather than formal conditions:

• Following Hempel (1965): no evidence establishes a hypothesis with certainty, so acceptance carries inductive risk, and rules of acceptance are “special instances of decision rules” that need valuations of the four outcomes: true or false acceptance, true or false rejection (pp. 561–562).
• Also following Hempel: value judgements have no role in whether evidence is confirmatory of h given background assumptions. They enter in choosing which evidence and which assumptions to accept (p. 565).
• Scientists have the same moral responsibilities as everyone else, including for the foreseeable consequences of error. This is explicitly “beyond the scope of this paper” (p. 563) and is assumed rather than argued.
• Scientific claims are taken as authoritative in policy, and many inductive risks are taken before the acceptance stage, where public decision-makers cannot see them (p. 563).
• The epistemic/non-epistemic distinction is porous (Rooney 1992; Longino 1996; Machamer & Douglas 1999). It is used only to indicate which goals a value primarily serves (p. 560).
• For the case studies: dioxin (2,3,7,8-TCDD) is assumed to be a tumour promoter and not a mutagen (p. 574). The extrapolation of rodent results to humans is set aside (p. 567).

Key results

These are arguments and case findings; there are no theorems.

• Structure of values (§3, pp. 563–565). Longino’s three uncontroversial roles (problem selection, use of results, ethical limits on methods) are direct and external. Inductive risk gives an indirect role, via consequences of error, that is internal. At each internal stage “one can make a wrong (i.e., epistemically incorrect) choice, with consequences following from that choice”. Values alone never determine the choice: the magnitude of inductive risk (how likely one is to be wrong) is weighed together with the valuation of the consequences (p. 565).
• Significance level (§4, pp. 565–569). For a fixed test, α trades false positives against false negatives. In dioxin bioassays, false positives lead to overregulation (costs to industry) and false negatives to underregulation (costs to public health). In the threshold model a false negative sets the “safe” dose too high, and because dose groups are spaced about 10× apart, it is “at least one order of magnitude less protective than it should be” (p. 568). Reducing both errors means more power, but typical groups are 50–100 animals, and a single two-year study with 200 rats (50 per dose × 3 doses + 50 controls) costs about $3 million (Graham & Rhomberg 1996) (p. 569). The studies use 95% significance, i.e. a false-positive rate ≤5% (fn. 11).
• Evidence characterisation (§5, pp. 569–572; Table 1, adapted from EPA 1994). The same female Sprague-Dawley liver slides were scored differently across the 1978, 1980 and 1990 evaluations. Total tumours: control 9/86, 16/86, 2/86; 1 ng/kg/day 3/50, 8/50, 1/50; 10 ng 20/50, 27/50, 9/50; 100 ng 34/50, 33/47, 18/50. Borderline calls in either direction trade false positives against false negatives, with regulatory consequences (the 1990 re-evaluation led Maine to relax its regulation). Blinding helps prevent the direct use of values and spreads a pathologist’s tendency evenly across groups. Blinding was impossible here, because acute liver toxicity visible at 10 and 100 ng/kg/day identified the high-dose slides, where most borderline cases were.
• Interpretation (§6, pp. 573–577). “Is there a carcinogenic threshold?” admits two plausible readings of the same Kociba data. No threshold: n = 50 cannot detect a 1% excess, so an apparent threshold at 1 ng/kg/day may be “a limit of detection, not a limit in the activity of dioxins”. Threshold: every tumour-bearing liver at 100 ng, and 8 of 9 at 10 ng, showed acute toxicity, while the lower-dose and control cancers did not. If the tumours are a by-product of cytotoxic regeneration, a threshold is plausible. Countries assuming a threshold set acceptable dioxin levels about 100× higher than no-threshold countries such as the US (Finkel 1988; Greenlee et al. 1991) (p. 577). Choosing the background assumption means choosing which error to risk.
• Limits of the argument (§7, pp. 577–578). Values are not needed when uncertainty is negligible (the meteorite analogy) or when errors have no outside consequences (atom-beam coherence). There is a grey area where consequences are hard to foresee. Recognising inductive risk should increase the motivation to reduce error.

Claims

id	claim	strength	support
C1	Inductive risk is present not only at hypothesis acceptance but at methodology choice, data characterisation and interpretation	strong	argument §3 plus three worked dioxin cases §§4–6
C2	Where the non-epistemic consequences of error are foreseeable, weighing them, and hence non-epistemic values, is a required part of sound scientific reasoning	moderate	argument §3/§7; rests on the premise of scientists’ moral responsibility, which is explicitly not argued (p. 563)
C3	The legitimate role of non-epistemic values in internal science is only indirect (through consequences of error); a direct role (“I want X, therefore X”) remains illegitimate	moderate	conceptual distinction §3, §5 (blinding), §7; stated rather than fully developed here
C4	For a fixed test, lowering α trades false negatives for false positives; reducing both requires more power, which is costly (~$3M per 200-rat, 2-year study)	strong	standard statistics plus cited cost figure (Graham & Rhomberg 1996)
C5	Expert characterisation of identical evidence varies materially, with regulatory consequences	strong	Table 1 (EPA 1994): three evaluations of the same Kociba slides; the 1990 panel voted by majority
C6	Both threshold and no-threshold readings of the Kociba data are plausible, and the choice between them reflects background assumptions and has ~100× regulatory consequences	moderate	§6 analysis; regulatory figure cited from Finkel 1988 and Greenlee et al. 1991
C7	The McMullin/Jeffrey division of labour (scientists ignore utilities; users apply them) fails because scientists’ claims are authoritative and many risks are taken invisibly before acceptance	moderate	informal argument §2
C8	Research without foreseeable outside consequences (“rapidly becoming a minority”) need not consider non-epistemic values	weak	example (atom beams) and an assertion about funding trends, §7
C9	”Value-free” science is inadequate science for areas with clear non-epistemic impacts	moderate	abstract/§1 headline; follows from C1–C2, so it inherits C2’s unargued moral premise

Method

This is conceptual argument grounded in Hempel’s decision-theoretic account of acceptance, applied to three case studies drawn from the author’s 1998 dissertation on dioxin science. The data are secondary: EPA 1994, Kociba et al. 1978, Goodman & Sauer 1992, Brown 1991 and regulatory comparisons.

Concepts

• Inductive risk — “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).
• Epistemic / non-epistemic values — non-epistemic means “social, ethical, political” (p. 559). The distinction is porous and marks which goals a value mainly serves.
• Direct role of values — values weigh the direct consequences of an action, as in the ethics of methods or the choice of technology. This is legitimate in the external stages.
• Indirect role of values — values weigh the consequences of possible error, alongside an estimate of how likely error is. This is legitimate, and required, in the internal stages when consequences are foreseeable.
• Externality model — the view (Longino 1990, pp. 85–86) that values constrain science only from outside, leaving internal reasoning value-free. This is Douglas’s target.
• Internal stages — choice of methodology, gathering and characterisation of data, and interpretation of data.
• Threshold vs linear-extrapolation (no-threshold) model — the two dose-response assumptions whose choice is at stake in §6.

Connections

Douglas builds directly on Hempel 1965 and on Churchman’s (1956) recognition of the many decision points in science. She positions herself against Jeffrey 1956, Levi 1962 and McMullin 1983. She is sympathetic to feminist critiques of the epistemic/non-epistemic distinction (Rooney, Longino), but her argument is normative where Longino’s is descriptive (fn. 8). Cranor 1993 is the regulatory-law ally on the costs of false negatives.

In the record: Reiss (LIT-118) independently builds the same indirect role of values into his theory of evidence: higher welfare stakes should lower the bar for rejecting alternatives, and further probing is a cost-benefit decision. Leonelli (LIT-189) cites Douglas among those who deny a strict internal/external boundary. Worrall (LIT-175) is relevant to §4, since both treat significance testing as a convention whose epistemic credentials must be argued, not assumed.

Epistemology tag. The work is about the norms of acceptance and justification: when evidence suffices to accept a hypothesis, a methodology or a characterisation of data. Its position is that sufficiency is not a purely evidential matter under uncertainty. The evidential relation itself stays value-free, following Hempel, but the threshold of acceptance properly depends on the costs of error. This is the “inductive risk” that the tag’s blurb names. The tag is justified.

Bearing on the record

This is a THEORY-grade argument, not an ML practice, and the anthology holds no document on it. Its implications for ML evaluation are real but are this reader’s extrapolation, not the paper’s content. Choice of decision threshold, of significance level in comparing systems, of how ambiguous labels are adjudicated (compare the majority-vote pathologist panel), and of which reading of an ambiguous capability or safety result to adopt all carry inductive risk, and when a deployment or policy consequence is foreseeable, those choices are value-laden. The paper’s endorsement of blinding as a guard against the direct role of values has a clear analogue in blind annotation and held-out evaluation. If the anthology writes a THEORY on reporting thresholds for safety evaluations, this paper is the source for “state the error costs that justify the threshold”. The nearest held anthology practices on statistical reporting (ANTH-SOTA-307, error bars over seeds; ANTH-SOTA-350, many splits and seeds) concern reducing uncertainty, which Douglas agrees is always desirable (§7). They do not address the value question she raises.

Limitations

• The central normative premise, that scientists are morally responsible for the foreseeable consequences of error, is explicitly left unargued (p. 563). The “required” in the main thesis inherits this.
• All three cases come from one domain (dioxin toxicology and regulation), where the policy consequences are unusually direct. The claim that consequence-free research is “rapidly becoming a minority” is asserted without evidence.
• The direct/indirect distinction is drawn sharply in the conclusion but only sketched. Whether the indirect role can in practice be kept from sliding into the direct one, for instance when a borderline-slide rule is chosen because of its known regulatory effect, is not addressed.
• The argument does not engage the rejoinder that scientists could report uncertainty (probabilities or likelihoods) rather than accept or reject. The seed rightly flags later responses (e.g. Betz), and they are not anticipated here.
• What the argument means for the “objectivity” of science is explicitly deferred to “future work” (p. 578). The argument against the Jeffrey/McMullin division of labour is a single paragraph (p. 563).

Open questions

• Can the indirect/direct line be operationalised, i.e. is there a test that distinguishes a value-weighted threshold choice from motivated reasoning? Douglas’s later book develops this. This paper does not.
• Can uncertainty reporting in place of acceptance (Jeffrey-style) remove the need for values at the characterisation and interpretation stages, where there is no explicit α?
• Who should supply the valuations: individual scientists, panels, or the public? The paper argues only that scientists cannot avoid them.
• Is the grey area of “useful but unforeseeable” consequences large or small in practice? An empirical survey of research areas would say.

Corrections to the seeded skim

• The skim says §4 is about “the conventional choice of a statistical-significance level”. Douglas explicitly sets the conventional choice aside. Levels are usually fixed “by the tradition of an area of research or the choice of a computer statistical package”, and “I will not argue that such conventions have not worked reasonably well” (p. 566). Her argument concerns what a deliberate and explicit choice of level would require. The critique of the conventions is attributed to Cranor 1993 (fn. 9).
• The skim describes §5 as “the TCDD bioassay (Brown 1991 re-reading)”. The case is the 1978 Kociba et al. (Dow Chemical) two-year study, and there were three evaluations of the same female rat-liver slides: Kociba 1978; Squire 1980 for the EPA; and 1990 by seven pathologists assembled by PATHCO at the paper industry’s request, reported in Goodman & Sauer 1992. The 1990 panel “resorted to majority voting” (≥4 of 7), which Douglas says is “not a consensus as normally understood” (pp. 569–571). Brown 1991 is cited for the relaxed regulation in Maine and for the toxicity-mediated argument, not as “the” re-reading.
• The seed summary says “every internal scientific choice carries inductive risk” and values are “necessarily required … wherever those consequences can be foreseen”. The second clause is right. The first overstates the paper: 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 (atom-beam coherence research), and a grey area is left case by case.
• On what the skim asks a deeper reading to check (how direct and indirect roles are separated): in this paper the direct role of non-epistemic values is legitimate only in the “external” decisions (problem selection, use of knowledge, ethical limits on methods; §3, after Longino 1990). In the internal stages it is illegitimate: blinding is endorsed precisely to prevent “the seeing of something because one wants to see it” (p. 572), and “I want X to be true, therefore X is true” remains a bad argument (p. 578). The fuller direct/indirect theory is in Douglas’s later work, not here.
• The paper also concedes that the epistemic/non-epistemic distinction “cannot do the boundary work many philosophers would like it to” and uses it only to mark which goals a value serves (§1, p. 560). The skim does not record this.
• epistemology tag: justified. Inductive risk is named in the tag’s own blurb, and the paper takes a position on the norms of hypothesis acceptance and on evidence characterisation.
• primary topic: philosophy-of-science is right. Consider adding natural-sciences, since the three cases are toxicology and risk regulation and are read closely.