A Technique for Analyzing the Effects of Group Composition

NEITHER problem we shall discuss nor its general solution is new. Indeed, germ of whole idea is contained in Durkheim's The Rules of Sociological Method.' The problem is this: Given data on individuals within specified aggregates (groups, counties, nations) what inferences can one make about nature of an effect at different levels of aggregation? Thus, for instance, finding that Catholic nations have lower suicide rates is consistent with any of following inferences: (1) Catholics are less likely to commit suicide regardless of religious composition of nation; (2) In countries with many Catholics, both Catholics and Protestants are less likely to commit suicide, but within a given country there is no religious difference in suicide; (3) Catholics are more likely to commit suicide than are Protestants, but both religions have lower suicide rates where there are many Catholics; (4) Regardless of country, Catholics have a constant suicide rate, but suicide rate for Protestants declines with proportion of Catholics. All these statements assert some sort of relationship between religion and suicide, but they differ widely in their substantive implications, because they differ in their assumptions about nature of effect at individual level (religious differences within a given nation) and group level (differences between persons of same religion in different nations). The statistical principle underlying these ambiguities has been recognized for some time. It has been stated negatively as the fallacy of ecological correlation. 2 Patricia Kendall and Paul F. Lazarsfeld, however, in analyzing methodological problems in The American Soldier, suggest that principle can lead to positive gains if data allow for simultaneous group and individual level analyses. They write: