Explanation and Theoretical Virtues

This, however, is just part of the story: scientists do not look simply for theories that are compatible with the phenomena, but that are true, explanatory and fecund. So, theories must be empirically testable, plausible (i.e., consistent with the largest possible background of accepted beliefs), simple, non-ad hoc, and provide models of the underlying causes, which at the same time explain the phenomena, unify them, and predict new phenomena as effects of the same causes.

In other words, the evidence for a theory does not coincide with its empirical consequences: for instance, trivial and ad hoc consequences don’t confirm (Laudan 1996, ch. 3; Laudan and Leplin 1991, 1993; Psillos 1999, 169-176), and certain phenomena can confirm a hypothesis even if not entailed by it (Psillos 1999: 170). In Bayesian terms, different hypotheses have different prior probabilities, so even if supported by the same data, they get different posterior probabilities (Fano 2005: 166; Psillos 1999: 163). Theoretical virtues have confirmatory power (Glymour 1980; Kosso 1992; Psillos 1999: 171-176) and supply the necessary guidance to all ampliative inferences: even the mere inductive projection presupposes the unifor­mity of nature, and follows the rule of simplicity. The same holds for analogical reasoning, while abductive reasoning exploits the explanatory power of hypotheses.

Antirealists deny the confirmatory power of theoretical virtues: simplicity is evidence for truth only if nature is simple, or uniform; but this can be shown only by circularly assuming that our theories, which depict it as simple, are true. Assuming that consistency with other accepted theories is evidence for truth pre­supposes the petitio principii that accepted theories are themselves true: the entire body of our scientific beliefs might be coherent, but false.

This reminds us of the Humean criticism of ampliative reasoning: its validity cannot be proved a priori, since it is a factual question, nor a posteriori, since an inductive proof (from the past success of these methods) would be circular. Yet, this is not enough, if scientific antirealism must differ from general Humean skepticism (Alai 2006: 228-230). Moreover, we use basic patterns of deductive inference, like modus ponens and modus tollens, without requiring a non-circular proof of their validity; so, why shouldn't we do the same for equally basic forms of inductive inference? Besides, many antirealists use them in inferences from observed facts to yet unobserved ones; in particular, only ampliative methods allow us to claim that a theory is empirically adequate—as opposed to merely compatible with known data (Lipton 1991, 154 ff.). Some authors reply that ampliative methods are reliable when inferring from observed to unobserved-but-observable entities, but not when infer­ring to unobservable entities. However, this discrimination is unjustified, since the validity of inferential patterns does not depend on the subject-matter (Alai 2010: § 3).

Some antirealists deny that we need explanations, or that explaining is a task for science (Duhem 1906). However the search for explanations is natural for human beings and a spur to inquiry (Aristotle Metaphysics, 982b); it is typical of ordinary knowledge, and science is just a development of ordinary knowledge (Rescher 1987: 36-41). The desire to explain is not mere curiosity, but a search for con­sistency and self-correction: not for any event we ask why or how it happened, but only for those which in the light of our preconceptions should not or could not have happened. Hence, finding an explanation involves correcting some mistake or learning some new crucial information. In fact, the search for explanations produces new discoveries and is a key to the self-corrective character of science.

Van Fraassen (1980: 25) claimed that, when explaining observable regularities through unobservable ones, we leave the unobservable regularities unexplained; therefore we should limit our search for explanation to the observable level. However unobservable regularities may in turn be explained by deeper unobserv­able regularities, and these by even deeper ones, etc. At each level explanation gains generality, simplicity and information: fewer laws, entities and properties are employed. This does not necessarily launch an infinite regress, since wherever we stop we have a correct, informative and interesting explanation even if the ex- planans is left itself unexplained. The longest the explanatory chain gets, the more it becomes informative and interesting. Furthermore, often observable regularities are not explained by further regularities, but by postulating entities, such as genes, viruses, etc. (Aronson 1984).

4