Reading: Sober, Elliott (1993): "Epistemology for Empiricists." In H. Wettstein (ed.), Midwest Studies in Philosophy. Notre Dame: University of Notre Dame Press. Pages 39-61.

Concerns the importance of experience in shaping our beliefs about the world—a normative claim, and not of psychology. The core doctrine of empiricism is that experience is the sole source of information about the world.

Definition: A theory is empirically adequate, roughly, if all of what it says about observable aspects of the world (past, present, and future) is true.

Van Fraassen: Science aims to give us theories which are empirically adequate; and acceptance of a theory involves as belief only that it is empirically adequate.

• Van Fraassen's constructive empiricism is about the content of our theories to which we have epistemic access.

• This statement about the aim of science is a thesis about what the methods of science are able to achieve. It should not be understood as asserting what scientists believe to be the aim of science.

Van Fraassen's Thesis: If a theory meets certain criteria, then we should conclude that it is empirically adequate, but we should remain agnostic about what they say about unobservables.

• We may form such beliefs, but there is no reason why they should be reliable.

• In Reichenbach's cubical world, for example, to believe that the birds exist is unjustified. We should always remain agnostic about the existence of the birds.

Criticism: Constructive empiricism gives the distinction between what is observable and what is unobservable an epistemological significance it does not possess.

Van Fraassen's Definition: X is observable if and only if there are circumstances which are such that, if X is present to us under those circumstances, then we observe it.

• The moons of Jupiter are observable because a properly situated astronaut could see them without the aid of instrumentation.
• The AIDS virus is not observable because it detection will always require the use of instruments.

Why should this difference matter? The distinction between observed and unobserved does matter. But why should the distinction between something unobserved but observable be more believable that something unobserved and unobservable?

We should be guided by the information we actually possess.

Example: We have a fallible test of diabetes and a fallible test for small pox, of equal reliability. We run these tests, and get a positive test result in each case. Everyone agrees that we should be equally confident in our conclusion.

Suppose we are now told that there exists a completely reliable and infallible test for diabetes, which has not been conducted. Should we revise our judgment in light of this new information? No, it is irrelevant.

• Actualism is a principle upon which realists and antirealists could agree.
• Constructive empiricism violates actualism. We detect facts about Jupiter's moons and facts about the AIDS virus using instruments. The counterfactual possibility that Jupiter's moons could be observed, while the virus could not, is irrelevant to the reliability of our conclusions in the two cases.

However, the difference between 'observed' and 'unobserved' is relevant according to actualism.

1. If a theory is true, then it is empirically adequate. The converse does not hold. So, empirical adequacy is the weaker claim.
2. Where there is a choice, we should prefer a claim with the greater chance of being true.
3. Therefore, we should stop short of asserting the truth of theories, and be agnostic about what they assert over and above the claim that they are empirically adequate.

Objection: Premise 2 is false. The claim of empirical adequacy is not the weakest claim possible. The claim that theories make true predictions about events on Mondays, Tuesdays, and Wednesdays is weaker still. Or maybe we should make predictions at all, and limit ourselves to what a theory says about the past? Or maybe we should make no claims for science at all (you can't get weaker than that)?

The inductive logicians talked about the probability of an inductive conclusion (translate this as 'hypothesis' in what follows) given the evidence. We denote this by P(H/E). This should be contrasted with the question of how probable the evidence is given the hypothesis.

Definition: P(E/H) is called the likelihood of H relative to E, whereas P(H/E) is the probability of H given E.

Example 1: Let H = I'm always right, E = I was right when I said that Hume was a philosopher. P(E/H) is high, but P(H/E) is low.

Example 2: Suppose that H = Carbon-14 has a half life of 5,730 ± 40 years and E = this atom of C-14 decayed yesterday. P(E/H) is low and P(H/E) is high (because P(H) is already high, and P(H/E) is not much different in value).

• Reichenbach's principle of common cause appeals to the notion of the likelihood of alternative hypotheses, not their probabilities. "If an improbable coincidence has occurred, then there must exist a common cause."
• There is a sense in which likelihood can be thought of as fit with data. If the likelihood is high, then the fit is high.
• The likelihood can be thought of as the strength of expectation that the hypothesis confers on the data.

The Likelihood Principle: Observational evidence O favors H₁ over H₂ if and only if P(O/H₁) > P(O/H₂). O favors the hypothesis that best fits with O.

Remark: The likelihood principle does not include simplicity as factor in the judgment of theories. So, it needs some revision, but is a good first formulation of empiricism, which provides many valuable insights. In particular, if the competing hypotheses are equal in simplicity, then the likelihood principle will apply.

Argument: The likelihood principle entails actualism. P(O/T is empirically adequate) = P(O/T is true), so that what theories says about unobservables is not relevant to whether O favors the truth or the empirical adequacy of a theory.

1. Empiricism should be formulated in terms of some matter-of-degree notion such as empirical support, or predictive accuracy.
2. Empiricism should claim that observations are special sources of information about which theory is favored; it should not claim that observables are the sole subject matter about which we have reasonable opinions.
3. Empiricism should conform to the principle of actualism.

Conclusion: The likelihood principle meets these three desiderata. Constructive empiricism does not meet these desiderata.

Constructive empiricism was motivated, in part, by the failure of the principle of common cause (PCC) in QM. But that principle, in turn, was motivated by the likelihood principle: Given a choice between a common causal explanation and the cosmic coincidence explanation, the common cause explanation is favored because it makes the observations more probable. PCC breaks down in QM because the common cause explanation makes false predictions, and therefore makes the observed data highly improbable. So, the likelihood principle explains why PCC breaks down here.

This is important because one of the aims of philosophy of science is to find general methods or patterns of inference that apply to many different areas of investigation, if such general methods exist. PCC looked like such a principle, but the Bell argument showed that it is not completely general. However, there is now hope that the likelihood principle does the same work as PCC, but is more general.

So, what does likelihood principle say in the Bell example?

It will rank the possible explanations as follows.

• The common causal explanation and the cosmic coincidence explanations are ranked at the bottom, because they make the observed data highly improbable.
• QM does better because it at least says that the correlations are to be expected.
• The conspiracy theory and the action-at-a-distance theories also fail to predict the exact outcomes of each measurement, because they can not say which outcomes will be changed. But they do at least as well as QM because they do predict the correlations. However, they postulate entities that we cannot measure, and that seems to could against them.

Sober 1993, p. 53: "when something (e.g., an atom) is unobservable, the only way we can find out about it is by finding out about something observable." (p. 54) "Our knowledge of unobservables is necessarily indirect."

The distinction between direct and indirect knowledge is epistemologically relevant, and is connected in some way to the observable/unobservable distinction. But not in any simple way.

• Much of our knowledge of observables is also indirect.

• Some of our indirect knowledge may be more secure than the direct knowledge on which it is based if the errors of observation cancel out. E.g., it is possible to obtain knowledge of the mean daytime temperature on a particular day of the year (something not directly observable) using inaccurate thermometers if we have data over sufficiently many years.

• What should an empiricist say about Bell argument? Is this a case in which we should believe in unobservables, or not? It seems to be neutral on this question.

• How do non-empirical virtues like simplicity figure into an empiricist picture of science? Does the extension of the likelihood principle discussed in Forster and Sober (1994), in which hypotheses are judged in terms of likelihood as well as simplicity, conform to the 3 desiderata? (It certainly does not conform to the likelihood principle.)
• Van Fraassen still has a good case for one thesis: that it not possible to discriminate between predictively equivalent theories. For example, how do we discriminate between T and the hypothesis that T is false but an evil demon makes it appear to us as if T is true? This is not a scientifically respectable problem.