¹ A brief discussion of the uses and limitations of this technique is given in the AID program description in Inter-University Consortium for Political Research, OSIRIS II: OS Users Manual (Ann Arbor: Institute for Social Research, University of Michigan, mimeo, 1971), pp. 375–388Google Scholar. For a more detailed discussion of the method, see Sonquist, John A. and Morgan, James N., The Detection of Interaction Effects: A Report on a Computer Program for the Selection of Optimal Combinations of Explanatory Variables, Monograph No. 35 (Ann Arbor: Institute for Social Research, University of Michigan, 1964)Google Scholar.

² An exception is the effort of scholars associated with Columbia's Bureau of Applied Social Research to develop socially defined contexts, especially through the use of “snowball” and “sociometric” sampling techniques. See Barton, Allan H., “Bringing Society Back In: Survey Research and Macro-Methodology,” The American Behavioral Scientist, 12 (November-December, 1968), 1–9CrossRefGoogle Scholar, and the works cited therein.

³ See, for example, Scheuch, Erwin K., “Social Context and Individual Behavior,” in Quantitative Ecological Analysis in the Social Sciences, ed. Dogan, Mattei and Rokkan, Stein (Cambridge: MIT Press, 1969), pp. 133–135Google Scholar; Valkonen, Tapani, “Individual and Structural Effects in Ecological Research,” in Dogan, and Rokkan, , pp. 53–68Google Scholar; and European Politics: A Reader, ed. Dogan, Mattei and Rose, Richard (Boston: Little, Brown, 1971), pp. 144–145CrossRefGoogle Scholar.

⁴ The term “breakage effect” is developed in Berelson, Bernard R., Lazarsfeld, Paul F., and McPhee, William N., Voting: A Study of Opinion Formation in a Presidential Campaign (Chicago: University of Chicago Press, 1954), pp. 98–101Google Scholar.

⁵ Blau, Peter M., “Structural Effects,” American Sociological Review, 25 (February, 1960), 178–193CrossRefGoogle Scholar.

⁶ Davis, James A., Spaeth, Joe L., and Huson, Carolyn, “A Technique for Analyzing the Effects of Group Composition,” American Sociological Review, 26 (April, 1961), 215–225CrossRefGoogle Scholar.

⁷ Ennis, Phillip H., “The Contextual Dimension in Voting,” in Public Opinion and Congressional Elections, ed. McPhee, William N. and Glaser, William A. (New York: Free Press, 1962), pp. 180–211Google Scholar.

⁸ See, for example, Blau, ; Valkonen, ; Davis, , Spaeth, , and Huson, ; Tannenbaum, Arnold S. and Bachman, Jerald G., “Structural Versus Individual Effects,” American Journal of Sociology, 69 (May, 1964), 585–595Google Scholar; Putnam, Robert D., “Political Attitudes and the Local Community,” American Political Science Review, 60 (September, 1966), 640–654CrossRefGoogle Scholar.

⁹ One such effort, however, is Meltzer, Leo, “Comparing Relationships of Individual and Average Variables to Individual Response,” American Sociological Review, 28 (February, 1963), 117–123CrossRefGoogle Scholar.

¹⁰ The survey data were originally collected by David Butler and Donald Stokes. See Political Change in Britain: Forces Shaping Electoral Choice (New York: St. Martin's, 1969)Google Scholar. They were provided by the Inter-University Consortium for Political Research. The census data were originally collected by William Miller and were made available through the Zentralarchiv für empirische Sozialforschung of the University of Cologne. The original collectors and the archives supplying the data, of course, bear no responsibility for the uses made of it in this analysis.

¹¹ For arguments which, at least by implication, condemn such atheoretical approaches, see Holt, Robert T. and Richardson, John M. Jr.,, “Competing Paradigms in Comparative Politics,” in The Methodology of Comparative Research, ed. Holt, and Turner, John E. (New York: Free Press, 1970), pp. 21–71Google Scholar.

¹² See the comments of Mayer, Lawrence C., Comparative Political Inquiry: A Methodological Survey (Homewood, Ill.: Dorsey Press, 1972), pp. 279–280Google Scholar for a discussion of the research strategy appropriate for a pretheoretical discipline such as political science.

¹³ Sonquist and Morgan; Sonquist, John A., “Finding Variables That Work,” Public Opinion Quarterly, 33 (Spring, 1969), pp. 83–95CrossRefGoogle Scholar; and Sonquist, John A., Multivariate Model Building: The Validation of A Search Strategy (Ann Arbor: Institute for Social Research, University of Michigan, 1970)Google Scholar.

¹⁴ Inter-University Consortium for Political Research, p. 375.

¹⁵ Lipset, Seymour Martin and Rokkan, Stein, Party Systems and Voter Alignments (New York: Free Press, 1967), pp. 1–64Google Scholar. The basic cleavages discussed by Lipset and Rokkan are social and economic in nature. Consequently, the individual and contextual indicators in Table 1 are all social and economic variables. It is possible that more variance would be explained if individual psychological orientations (such as issue opinions) and/or contextual political indicators (such as past party voting) were utilized in the analysis. Other studies might find it worthwhile to include such alternative predictors. They were omitted here to maintain manageability of data in an openly exploratory analysis.

¹⁶ Since the independent variables were restricted to indicators of the basic socioeconomic cleavages in Western Europe, there were in fact few problems of availability. All the relevant cleavages could be operationalized by at least one individual and one contextual indicator in the available data sets.

¹⁷ Another way of evaluating the trees in Figures 1–3 is by ascertaining the reduction in the number of prediction errors for individual respondents made when using the information provided by the individuals' final group classifications in the trees rather than the overall mode. Without knowledge of individual characteristics, one's best prediction for each individual voter would be the mode: a Labour vote. Predicting the overall mode for all 1363 respondents leads to 641 errors in prediction. Predicting the mode of each final group for all individuals in the group produces the following numbers of prediction errors: Figure 1, 315; Figure 2, 548; Figure 3, 315. For Figures 1 and 3 the proportionate reduction in error is .51 [(641–315)/641]. For Figure 2, it is only .15. Using these measures rather than the proportion of variance explained leads to identical conclusions.