For many years I have been thinking about scientific explanation, especially statistical explanation. From the beginning I disagreed with Carl G. Hempel on this subject. He claimed that high probability is a requirement for acceptable statistical explanations; 1 I argued that we need, instead, relations of statistical relevance. 2 At the same time, I was perfectly aware that statistical relations by themselves are not sufficient; in addition, we need to appeal to causal relations. For example, there is a strong correlation between the reading on a barometer and the occurrence of a storm, but the falling reading on the barometer does not produce the storm and does not explain it. The explanation demands a cause. Since the explanation is statistical, the cause has to be of a probabilistic sort. Around 1971 I hoped that it would be possible to define probabilistic causality in terms of such statistical concepts as Hans Reichenbach's conjunctive fork and his screening-off relation, but by the end of that decade I no longer saw any possibility of doing so. I gave my reasons in the 1980 essay "Probabilistic Causality", 3 and again in my book Scientific Explanation and the Causal Structure of the World. 4 For example, Reichenbach used the conjunctive fork to define the relation of a common cause; however, there are events that constitute a conjunctive fork which does not contain a common cause. 5 It was necessary to define the concept of a causal process as a way to distinguish conjunctive forks in which there are bona fide common causes from those in which they do not exist. Causal processes are the key because they furnish the links between the causes and their effects? I found, moreover, that there are two types of causal forks. In addition to the conjunctive fork we must define the interactive fork, in which two causal processes intersect one another. In this intersection both processes are modified, and the changes persist in these processes beyond the point of intersection. An interactive fork constitutes a causal interaction. It is not possible to define this type of causal fork in statistical terms. It is extremely important to understand the profound difference between the screening-off relation and conjunctive forks on the one hand, and causal processes and causal interactions on the other. The former can be defined in statistical terms; the latter cannot be defined in this way. Causal processes and causal interactions are physical structures whose properties cannot be characterized in terms of relationships among probability values alone. Let us consider a few examples. 1. There is a very small probability that an American man, selected at random, will contract paresis. There is a somewhat higher probability that an American man who has had sexual relations with a prostitute will contract paresis. The sexual relation with the prostitute is statistically relevant to paresis. In addition, there is a still higher probability that an American man who has contracted syphilis will develop paresis. Syphilis is also statistically relevant to paresis. However, for a man who has contracted syphilis, the relation with the prostitute is no longer statistically relevant. Syphilis screens off the visit to the prostitute from the paresis. Let