There is another and more interesting point raised in conversations on various occasions by Nancy Cartwright, Paul Holland, and others. It is that the full notion of causality requires a sense of experimental manipulation. There are many ways of formulating the idea. Holland likes to say that, from a statistical standpoint, without random assignment of individuals to experimental groups an unimpeachable causal inference cannot be made. My most immediate reply is that ordinary talk and much scientific experience as well does not in any sense satisfy these conditions of experimental design, that is, the causal claims that are made in ordinary talk or in much of science have not arisen from well-designed experiments but from quite different circumstances—in fact, from circumstances in which no experiments have taken place and in many cases are not possible. The great classical example is celestial mechanics. From the time of the Babylonians to the present, we have seen a variety of causal theories to account for the motion of the planets, the moon, and the stars. In the case of some terrestrial phenomena that are not themselves directly subject to experiment but for which an analysis can be built up in terms of experimental data, we are faced with a rather more complicated decision about what we regard as proper extrapolation from experiment. In fact, one underhanded way to meet the objections raised by Cartwright and Holland is to point out that the use of scientific theories outside the experimental domain and the power of the application of science depend upon sustaining causal claims in nonexperimental settings. Are we to conduct experiments on extendability in order to establish a justification of using the results of experiments in nonexperimental settings? It would not be difficult to set up a straw man of infinite regress by literal pursuit of this line of thought. My own view is that, rather than claiming that only in experimental settings can we really make proper causal claims, we should formulate theorems that are applicable to experimental settings but not to others. It seems to me one kind of theorem we might want to insist upon is that for experiments whose theory of design is adequate we should expect to be able to prove within a framework of explicit probabilistic concepts that all prima fade causes are genuine. We would not expect such a theorem to hold in general in nonexperimental settings.

Kreisel has pointed out to me that the general theory of causality is unlikely to be of much scientific significance once specific scientific theories are considered. Indeed, the interest of such theories is to provide a testing ground for the correctness of the general notions. On the other hand, not only ordinary talk but much highly empirical scientific work does not depend on a well-defined theoretical framework, and for these cases the general theory of causality can provide useful analytic concepts.

Foundations of Psychology

I have already remarked on my earliest experimental work in psychology in connection with the test of various concepts and axioms of decision theory. I shall not refer further to that work in this section. Because of my extensive work in psychology over the past two decades, I have organized my remarks under four headings: learning theory, mathematical concept formation in children, psycholinguistics, and behaviorism.

Learning theory

Either in my last months as a graduate student at Columbia or shortly after my arrival at Stanford in the fail of 1950—I cannot remember which—I developed my first interest in learning theory. As might easily be surmised, it began with trying to understand the various works of Clark Hull, not only the Principles of Behavior (1943) but also the relatively unreadable work written earlier in collaboration with the Yale logician Frederick Fitch and others (Hull, Hovland, Ross, Hall, Perkins, & Fitch, 1940). Part of my interest was stimulated by some very bright graduate students in psychology who attended my lectures in the philosophy of science. Probably the most influential was Frank Restle. I was a member of his dissertation committee, but I am sure I learned more psychology from him than he learned from me. My serious interest in learning theory began, however, in 1955 when I was a Fellow at the Center for Advanced Study in the Behavioral Sciences. Restle was there, but even more important for my future interests was the presence of William K. Estes. In his own and very different way, Estes has the kind of intellectual clarity I so much admired in McKinsey and Tarski. We began talking seriously about the foundations of stimulus sampling theory, which really began with Estes’s classical paper (1950). It became apparent to me quite soon that stimulus sampling theory was from a conceptual and mathematical standpoint much more viable and robust than the Hullian theory of learning. No really interesting mathematical derivations of experimentally testable results could be made from Hull’s axioms. The great virtue of stimulus sampling theory was that with variation of experimental conditions new experimental predictions could be derived in an honest way without the introduction of ad hoc parameters and with the hope of detailed experimental test.