315 | Branden Fitelson on the Logic and Use of Probability

Episode Summary

Executive Summary: Sean Carroll and philosopher Brandon Feitelson explore how science uses probability without ever proving anything deductively. The discussion argues for a pluralistic Bayesian view: probabilities are theory-dependent, confirmation is distinct from posterior belief, and strong evidence is best understood through likelihood ratios plus priors. They also apply this to classic paradoxes and experimental design.

Main Topics: Science versus proof (Priority: 5/5): Science does not provide mathematical proof; it updates models using evidence that can be overturned by new data or better theories. What probability is (Priority: 5/5): Feitelson distinguishes objective probabilities built into scientific theories from epistemic probabilities used to assess arguments and evidence. Induction and skeptical challenges (Priority: 5/5): The conversation reviews Hume’s problem of induction and argues that scientific success is best explained by real evidential support, not deductive certainty. Confirmation, likelihoods, and priors (Priority: 5/5): A key theme is the difference between posterior probability and confirmation: evidence can strongly confirm a hypothesis even when the hypothesis remains unlikely due to low prior probability. Two-dimensional argument strength (Priority: 4/5): Feitelson argues that argument strength has two dimensions: conditional probability and relevance/confirmation, rather than a single summary number. Paradoxes and experiments (Priority: 4/5): Cases like the Raven paradox, conjunction fallacy, and Wason selection task are used to show how confirmation theory clarifies puzzling reasoning patterns. Pluralist Bayesianism in practice (Priority: 4/5): The talk concludes that different scientific contexts require different probability models, and experimental design should aim to maximize confirmational power.

Key Arguments: Science yields theories and models, not proofs; evidence can support one theory more than another without guaranteeing truth. Probabilities in science are often objective features of theories (like mass), but epistemic probabilities are needed to compare theories or evaluate arguments neutrally. Frequency theory is inadequate as a general account of probability because it cannot handle finite cases, odd numbers of trials, or irrational probabilities. Confirmation is distinct from posterior probability: a result can strongly confirm a hypothesis while the hypothesis remains low-probability because of a low prior. The Bayes factor / likelihood ratio measures confirmational force and is objective in the sense that it depends on test characteristics rather than a subject’s prior beliefs. Argument strength should be treated as two-dimensional: posterior probability plus confirmation/relevance. Popper was partly right that refutation has special force, but wrong that only refutation matters; negative evidence can be graded without collapsing all inference into falsification. The Raven paradox is resolved by quantitative confirmation: non-ravens/non-black things may confirm a universal claim, but vastly less than observing ravens. The Wason selection task mirrors the Raven paradox; choosing the counterexample card is more informative than checking a confirming instance. No single universal probability distribution can assess every argument; instead, each case requires constructing an appropriate model and probability assignment.

Data Points: Odd tosses in finite frequency view: 5 tosses, 3 heads and 2 tails - Example used to show why actual finite frequencies cannot define probability in general. Classical test case: 1 in 1000 - Rare-disease example from Kahneman-Tversky used to illustrate base-rate fallacy. Base-rate fallacy structure: True positive rate vs false positive rate - Diagnostic-testing example used to separate confirmation from posterior probability. Quantitative criterion: Bayes factor / likelihood ratio - Defined as the ratio of true-positive to false-positive rates, used as the measure of confirmation. Confirmation dimensions: 2 - Feitelson’s model treats argument strength as both posterior probability and relevance/confirmation. Raven example: All Ravens are black - Universal hypothesis used in the paradox of confirmation discussion. Wason selection task cards: D, K, 3, 7 - Four-card problem used to test the hypothesis that D implies 3. Historical period: Late 1960s - Linda problem background involving Berkeley and anti-nuclear activism.

Pivotal Quotes: "Science never proves things." — Sean Carroll: Opening framing of the episode, contrasting science with deductive proof in mathematics and logic. "No one's entitled to their own likelihoods." — Sean Carroll: Summary of the distinction between subjective priors and objective experimental evidence. "I want to say there's a two-dimensional theory of argument strength." — Brandon Feitelson: Feitelson explains his central thesis that probability and confirmation/relevance are separate axes.

Implications: Listeners should expect scientific claims to be probabilistic, context-dependent, and model-based rather than proven. For science and policy, the key is not certainty but choosing experiments and models that maximize confirmational power.