CPA Award for Distinguished Contribution to Psychology in Education and Training (1999) -- Prix de la SCP pour contribution remarquable a l'education et la formation en psychologie (1999)
It is well known to anyone who has taken even an introductory course in statistics that "correlation does not mean causation." This is a truism! The present article examines four truisms about correlation, demonstrating that they are not always true. For example, under certain conditions, correlation can imply causation, though these conditions are seldom satisfied in most applications. Nonetheless, there are many of us who are interested in investigating individual differences, and in making inferences of the type that this individual difference variable is related to, mediates, moderates, or even causes or influences that individual difference variable. Generally speaking, the analytic procedures we use involve the correlation coefficient in one form or another. I propose four steps that researchers can follow to accumulate evidence that increases one's confidence in the validity of a particular causal model. These steps are illustrated by reviewing research on individual differences in second language acquisition. This approach is not conclusive, of course, but it does force one to examine the implications of the model, thus leading to further insights and research. Although the focus here is on second language acquisition, the generalizations apply to other areas of research that are concerned with individual differences.
When I was told that I was to receive the CPA Education and Training Award and was to give a talk in conjunction with it, I thought long and hard about a possible topic. My teaching interests involve statistics and data analysis, and my research interests are concerned primarily with the role of attitudes and motivation in second language acquisition. I decided, therefore, to focus attention on a statistical and conceptual issue that has troubled me over the years, and to discuss how I have resolved the issue in my own mind in the context of my research interests. The issue is that of the correlation coefficient and causation. To some, this is a nonissue. You cannot infer causation from correlation. Case closed!
To those interested in individual differences, however, such a fatalistic conclusion is tantamount to concluding that there is no possible way of ever drawing a causal inference based on individual differences. One approach is to accept the canon that correlation does not imply causation, then go on to talk about prediction as opposed to causation (though as we shall see this still implies causation), and rely on causal (i.e., structural equation) modelling, and the like. The point is that individual difference research involves covariation, and regardless of which analytical procedure one adopts (i.e., multiple regression, factor analysis, discriminant function analysis, or even structural equation modelling, etc.), the basic statistic involves co-relation in one form or another. In the end, many of us believe that we have identified causal associations, even though we will concede that other interpretations are possible. That is, we believe that personality causes, or accounts for some behaviours, intelligence plays a role in academic achievement, anxiety disrupts performance, etc.
The solution I propose is to direct attention not so much to the relationship but to the underlying process, accepting the causal interpretation that seems most appropriate to the relationship and then expanding the implications, continually refining and evaluating them in the context of a research program. This is similar to the notion of construct validity, but more inclusive since the focus is not so much on the validity of a test or measure, but rather the elaboration of a conceptual model that is based on research sometimes using different instruments in different contexts. The focus in this instance is on the validity of the causal hypothesis explaining the relationships among a series of variables. …