Proceedings and Addresses of the American Philosophical Association, Vol. 73, No. 2. (Nov.,
1999), pp. 47-76.
Nowadays, there is general agreement about the importance of understanding what it takes for a statement to be confirmed or disconfirmed by an observation. There is also a wide consensus that the design of experiments is an important issue; if somebody wants to test a proposition, it is important to make sure that the experiment that is carried out, actually bear on the proposition in question.
Testing is an inherently contrastive activity; and testing a hypothesis requires that it make a prediction that can be checked by observation. We make observations in order to learn about things that we do not observe; these observations must be “theory neutral”, they should be neutral, relative to the competing theories under test.
An empirically soluble problem is one in which the competing hypotheses make different observational predictions. The hypotheses rarely make observational predictions on their own; they require supplementation by auxiliary assumptions if they are to be tested. The problem here is that usually the auxiliary assumptions are looked for, or chosen because the researcher has good reasons to think they are true. This means that the auxiliary assumptions used in a test and the hypotheses under test differ in their epistemological standing. The observational outcomes favour one competing hypothesis over the others. But the test typically will not test the auxiliary assumptions at all. Typically, the auxiliary assumptions are epistemically independent to the test outcome.
Somebody could think that auxiliary assumptions may include idealizations. Nevertheless what is essential is not that one be able to say that a set of assumption is true, but that it is harmless- that correcting the idealization could not affect the conclusion one draws. However, it is not enough just to assert that the idealization is harmless; one must have evidence that this is so.
Finally the author talks about the probability of the hypotheses, and tells us that there is no probabilistic analog of modus tollens. If a hypothesis deductively entails something false, then the hypothesis is false. But if a hypothesis that what you observe was very improbable, what then? It does not follow that the hypothesis it self is improbable. So based on this, we can judge which hypotheses do better and which do worse in their competition, that is all.