Source
Reiss (2015), LIT-118 (Philosophy of Science 82(3)), read in full in NOTE-197 from the February 2015 Durham CHESS working paper. Page references are the working paper’s.
What was actually shown
Support and warrant (§2, pp. 5–7). Evidence is two things. Support is a fact relevant to a hypothesis h. Warrant is a reason to infer h, and comes in degrees.
Support (§3, pp. 9–15). Deductive entailment by h is neither necessary nor sufficient for support: a “biased coin” hypothesis entails no particular data set, and a cause acting by two cancelling routes entails no correlation (Fig. 1). Support is a pattern in the data “we are entitled to expect to obtain under the supposition that h is true” (p. 12). Direct support bears on h. Indirect support is a pattern incompatible with an alternative that could account for h’s support.
Relevance and elimination (§§4.2–4.3). An alternative is relevant only with an empirical reason, generic (selection bias for observational data) or case-specific. Eliminating one is a judgement, not a proof, and values enter it: when stakes are high, “the standards for rejecting an alternative should be lower” (p. 23).
Stopping (§4.4, pp. 24–27). A default entitlement to other scientists’ claims, and four guidelines: large effects, specific manner and timing, small inferential distance, and the costs of further inquiry against its benefits.
Grades (§5, Table 1, p. 29). Proof, when all relevant alternatives are eliminated; strong warrant, all salient and some others; moderate, most including some salient; weak, some. “Proof” is empirical and fallible.
Against experimentalism (§6, pp. 31–33). Treating randomised experiments as basic cannot explain why non-experimental data count at all. The elimination account explains both why trials succeed, removing “all alternative causal hypotheses – in one fell swoop”, and how they fail, through measurement, coding and pooling errors that randomisation does not prevent.
What could have come out otherwise. Nothing in the paper could have. The smoking case is an illustration: whether the framework delivers “is, alas, a matter for future research” (§7, p. 35).
What this does not say
- That it holds outside its scope. It is restricted to type-level causal hypotheses in domains where randomised experiments “can be and are frequently employed”, and excludes most of physics and chemistry, in vitro work and the historical sciences (§1; §2).
- How to settle what is relevant. What one is “entitled to expect”, which alternatives to consider and when to stop are handed to context and background knowledge, with no account of how disputes over context are settled. The four open questions are relocated, not answered.
- That the grades are principled. Exactly three grades short of proof is “arbitrary” by the author’s concession, and grading needs judgements of “most” and “some” that the theory does not fix.
- That there is “no gold standard whatsoever”. That follows from the framework and is asserted, not argued independently (NOTE-197, C8).
- That probability is the wrong model of credence. Reiss says so, but in a footnote, by citation of Norton (fn. 3).
- How lowered thresholds are kept from motivated reasoning. Values may lower the bar for rejecting an alternative, and the paper does not say how that is kept honest.
Connections
- Worrall (LIT-175) reaches an overlapping conclusion from inside the randomised-trial debate: telling evidence tells against plausible rivals, and randomisation has no special standing beyond that (NOTE-197, Connections). His account is THEORY-491.
- Cartwright’s account of singular causal claims (THEORY-451) is the population-to-setting step this account leaves aside: a well-run trial secures at most its study population’s average effect, and warrant for another setting needs evidence on the causal process there.
- The value-laden rejection of alternatives is the indirect role of values that Douglas argues for (LIT-139; THEORY-442). Reiss does not cite her. Whether “values lower the bar” licenses the direct role she forbids is open.