TL;DR
This paper argues that Forster's purported counterexamples to the Likelihood Theory of Evidence (LTE) are not convincing. The authors demonstrate that Forster's arguments rely on questionable intuitions and that the "consilience" he favors can actually be understood and accounted for within the framework of LTE.
What Problem Does This Paper Address?
The Likelihood Theory of Evidence (LTE) posits that the evidential value of data for comparing hypotheses lies solely in the likelihood of the data under each hypothesis. A key tenet of LTE is the "relativity of evidence," meaning evidence is always assessed relative to an alternative hypothesis. Forster (2006) proposed counterexamples to LTE, particularly in the context of distinguishing causal direction between variables, arguing that his "consilience of inductions" methodology could not be captured by LTE. This paper aims to refute Forster's counterexamples and show that consilience is, in fact, compatible with LTE.
Methods
The authors tackle Forster's counterexamples by:
- Carefully examining Forster's arguments: They dissect one of Forster's key examples involving two variables, X and Y, and two hypotheses about their causal relationship (A: Y depends on X, B: X depends on Y).
- Highlighting the "relativity of evidence": They emphasize that LTE requires comparing hypotheses against alternatives. Forster's initial examples are critiqued for comparing "incomparable" hypotheses (those concerning different random variables), which violates this core LTE principle.
- Reinterpreting "consilience": They propose that the "consilience" Forster champions (how well different parts of a model or different pieces of evidence "jump together") can be systematically connected to likelihood comparisons. They show that a lack of consilience in a model corresponds to statistical dependence between variables, which in turn affects the likelihood.
- Using mathematical formulations: They employ likelihood calculations and introduce the concept of mutual information to quantify statistical dependence and demonstrate its link to likelihood differences.
Datasets
Not discussed in the paper. The paper uses hypothetical examples and simulated data based on these examples.
Key Insights and Ideas
- Forster's counterexamples are not convincing: The authors argue that Forster's examples fail because they either compare incomparable hypotheses or rely on intuitions that LTE proponents have long criticized.
- The "relativity of evidence" is crucial: The paper strongly defends the idea that evidence is always relative to an alternative. Forster's initial examples falter because they don't adhere to this principle.
- Consilience is compatible with LTE: The authors show that the concept of "consilience" (or lack thereof) can be understood as a measure of statistical dependence between variables within a model. This dependence directly impacts the likelihood of the data under different hypotheses, thus fitting neatly within the LTE framework.
- Mutual information as a measure of consilience: They propose mutual information as a quantitative way to measure the "lack of consilience," demonstrating its relationship to likelihood differences.
Key Results
- The paper demonstrates that Forster's initial counterexample, involving hypotheses about different random variables, is invalid because the hypotheses are incomparable under LTE.
- A revised version of Forster's example, using comparable hypotheses, shows that the difference in likelihoods between competing models directly reflects the degree of "consilience" (or lack thereof) as defined by Forster. Specifically, the difference in log-likelihoods between two models (A" and B") is shown to be directly proportional to the difference in mutual information between key variables (Î(Y, V) – Î(X, U)).
- The authors conclude that Forster's consilience-based methodology can be vindicated and understood within the framework of the Likelihood Theory of Evidence.
Limitations and Future Work
- Focus on one example: The paper primarily focuses on one of Forster's examples involving discrete variables, although the authors state their points extend to his continuous variable example.
- "Ad hoc" marginals: In one version of the example, the authors note that the constructed marginal distributions are "objectionable and useless in practice" but set them aside to focus on the core argument.
- Future work: The authors suggest that the connection between likelihood and consilience holds more generally for causal models and warrants further exploration in various contexts beyond causal inference.
Summary updated