¹ To be sure, the work of the Federal Convention has been supplemented in two critical respects. The first of these is the practice, virtually-universal since the mid-19th Century, of choosing Representatives from single-member districts of limited geographic area. The second is the practice, which has also become virtually universal In our own century, of selecting party nominees for the House by direct primary election.

² In the language of Eulau, Wahlke, et al., we speak here of the “style,” not the “focus,” of representation. See their “The Role of the Representative: Some Empirical Observations on the Theory of Edmund Burke,” this Review, Vol. 53 (09, 1959), pp. 742–756 Google Scholar. An excellent review of the mandate-independence controversy is given by Pitkin, Hanna Fenichel, “The Theory of Representation” (unpublished doctoral dissertation, University of California, Berkeley, 1961)Google Scholar. For other contemporary discussions of representation, see de Grazia, Alfred, Public and Re¬public (New York, 1951)Google Scholar, and John A. Fairlie, “The Nature of Political Representation,” this Review, Vol. 34 (April–June, 1940), pp. 236–48, 456–66.

³ The sampling aspects of this research were complicated by the fact that the study of representation was a rider midway on a four-year panel study of the electorate whose primary sampling units were not congressional districts (although there is no technical reason why they could not have been if the needs of the representation analysis had been foreseen when the design of the sample was fixed two years before). As a result, the districts in our sample had unequal probabilities of selection and unequal weights in the analysis, making the sample somewhat less efficient than an equal-probability sample of equivalent size.

It will be apparent in the discussion that follows that we have estimated characteristics of whole constituencies from our samples of constituents living in particular districts. In view of the fact that a sample of less than two thousand constituents has been divided among 116 districts, the reader may wonder about the reliability of these estimates. After considerable investigation we have concluded that their sampling error is not so severe a problem for the analysis as we had thought it would be. Several comments may indicate why it is not.

To begin with, the weighting of our sample of districts has increased the reliability of the constituency estimates. The correct theoretical weight to be assigned each district in the analysis is the inverse of the probability of the district's selection, and it can be shown that this weight is approximately proportional to the number of interviews taken in the district. The result of this is that the greatest weight is assigned the dis¬tricts with the largest number of interviews and, hence, the most reliable constituency estimates. Indeed, these weights increase by half again the (weighted) mean number of interviews taken per district. To put the matter another way: the introduction of differential weights trades some of our sample of congressional districts for more reliable constituency estimates.

How much of a problem the unreliability of these estimates is depends very much on the analytic uses to which the estimates are put. If our goal were case analyses of particular districts, the constituency samples would have to be much larger. Indeed, for most case analyses we would want several hundred interviews per district (at a cost, over 116 districts, of several small nuclear reactors). However, most of the findings reported here are based not on single districts but on many or all of the districts in our sample. For analyses of this sort the number of interviews per district can be much smaller.

Our investigation of the effect of the sampling variance of the constituency estimates is quite reassuring. When statistics computed from our constituency samples are compared with corresponding parameter values for the constituencies, the agreement of the two sets of figures is quite close. For example, when the proportions voting Democratic in the 116 constituencies in 1958, as computed from our sample data, are compared with the actual proportions voting Democratic, as recorded in official election statistics, a product moment correlation of 0.93 is obtained, and this figure is the more impressive since this test throws away non-voters, almost one-half of our total sample. We interpret the Pearsonian correlation as an appropriate measure of agreement in this case, since the associated regression equations are almost exactly the identity function. The alternative intraclass correlation coefficient has almost as high a value.

Although we believe that this analysis provides a textbook illustration of how misleading intuitive ideas (including our own) about the effects of sampling error can be, these figures ought not to be too beguiling. It is clear that how close such a correlation is to 1.0 for any given variable will depend on the ratio of the between-district variance to the total variance. When this ratio is as high as it is for Republican and Democratic voting, the effect of the unreliability of our constituency estimates is fairly trivial. Although the content of the study is quite different, this sampling problem has much in common with the problem of attenuation of correlation as it has been treated in psychological testing. See, for example, Guilford, J. P., Fundamental Statistics in Psychology and Education (New York, 1956), pp. 475–78Google Scholar.

⁴ See Campbell, Angus, Converse, Philip E., Miller, Warren E., and Stokes, Donald E., The American Voter (New York, 1960), pp. 194–209 Google Scholar.

⁵ This conclusion, fully supported by our own work for later Congresses, is one of the main findings to be drawn from the work of Duncan MacRae on roll call voting in the House of Representatives. See his Dimensions of Congressional Voting: A Statistical Study of the House of Representatives in the Eighty-First Congress (Berkeley and Los Angeles: University of California Press, 1958)Google Scholar. For additional evidence of the existence of scale dimensions in legislative behavior, see Gage, N. L. and Shimberg, Ben, “Measuring Senatorial Progreseivism,” Journal of Abnormal and Social Psychology, Vol. 44 (01 1949), pp. 112–117 CrossRefGoogle Scholar; Belknap, George M., “A Study of Senatorial Voting by Scale Analysis” (unpublished doctoral dissertation, University of Chicago, 1951)Google Scholar, and “A Method for Analyzing Legislative Behavior,” Midwest Journal of Political Science, Vol. 2 (1958), pp. 377–402 CrossRefGoogle Scholar; two other articles by MacRae, , “The Role of the State Legislator in Massachusetts,” American Sociological Review, Vol. 19 (04 1954), pp. 185–194 CrossRefGoogle Scholar, and “Roll Call Votes and Leadership,” Public Opinion Quarterly, Vol. 20 (1956), pp. 543–558 CrossRefGoogle Scholar; Farris, Charles D., “A Method of Determining Ideological Groups in Congress,” Journal of Politics, Vol. 20 (1958), pp. 308–338 CrossRefGoogle Scholar; and Rieselbach, Leroy N., “Quantitative Techniques for Studying Voting Behavior in the U. N. General Assembly,” International Organization, Vol. 14 (1960), pp. 291–306 CrossRefGoogle Scholar.

⁶ The content of the three issue domains may be suggested by some of the roll call and interview items used. In the area of social welfare these included the issues of public housing, public power, aid to education, and government's role in maintaining full employment. In the area of foreign involvement the items included the issues of foreign economic aid, military aid, sending troops abroad, and aid to neutrals. In the area of civil rights the items included the issues of school desegregation, fair employment, and the protection of Negro voting rights.

⁷ Because this point has been so widely discussed it has inevitably attracted a variety of terms. Dahl denotes the acts of a whose performance A is able to influence as the scope of A's power. See Dahl, Robert A., “The Concept of Power,” Behavioral Science, Vol. 21 (07 1957), pp. 201–215 Google Scholar. This usage is similar to that of Lasswell, Harold D. and Kaplan, Abraham, Power and Society (New Haven: Yale University Press, 1950), pp. 71–73 Google Scholar. Dorwin Cartwright, however, denotes the behavioral or psychological changes in P which O is able to induce as the range of O's power: “A Field Theoretical Conception of Power,” Studies in Social Power (Ann Arbor: Research Center for Group Dynamics, Institute for Social Research, The University of Michigan, 1959), pp. 183–220 Google Scholar.

⁸ That the Representative's roll call votes can diyerge from his true opinion is borne out by a number of findings of the study (some of which are reported here) as to the conditions under which agreement between the Congressman's roll call position and his private attitude will be high or low. However, a direct confirmation that these two sets of measurements are not simply getting at the same thing is given by differences in attitude-roll call agreement according to the Congressman's sense of how well his roll call votes have expressed his real views. In the domain of foreign involvement, for example, the correlation of our attitudinal and roll call measurements was .75 among Representatives who said that their roll call votes had expressed their real views fairly well. But this correlation was only .04 among those who said that their roll call votes had expressed their views poorly. In the other policy domains, too, attitude-roll call agreement is higher among Congressmen who are well satisfied with their roll call votes than it is among Congressmen who are not.

⁹ During the analysis we have formed constituency scores out of the scores of constituents living in the same district by several devices other than calculating average constituent scores. In particular, in view of the ordinal character of our scales we have frequently used the median constituent score as a central value for the constituency as a whole. However, the ordering of constituencies differs very little according to which of several reasonable alternatives for obtaining constituency scores is chosen. As a result, we have preferred mean scores for the greater number of ranks they give.

¹⁰ The meaning of this procedure can be suggested by two percentage tables standing for hypothetical extreme cases, the first that of full agreement, the second that of no agreement whatever. For convenience, these illustrative tables categorize both Congressmen and their districts in terms of only three degrees of favor and assume for both a nearly uniform distribution across the three categories. The terms “pro,” “neutral,” and “con” indicate a relative rather than an absolute opinion. In Case I, full agreement, all districts relatively favorable to social welfare action have Congressmen who are so too, etc. ; whereas in Case II, or that of no agreement, the ordering of constituencies is independent in a statistical sense of the ranking of Congressmen : knowing the policy orientation of a district gives no clue at all to the orientation of its Congressman. Of course, it is possible for the orders of legislators and districts to be inversely related, and this possibility is of some importance, as indicated below, when the policy position of non-incumbent candidates as well as incumbents is taken into account. To summarize the degree of congruence between legislators and voters, a measure of correlation is introduced. Although we have used a variety of measures of association in our analysis, the values reported in this article all refer to product moment correlation coefficients. For our hypothetical Case I a measure of correlation would have the value 1.0; for Case II, the value 0.0. When it is applied to actual data this convenient indicator is likely to have a value somewhere in between. The question is where.

¹¹ A word of caution is in order, lest we compare things that are not strictly comparable. For obvious reasons, most non-incumbent candidates have no roll call record, and we have had to measure their policy agreement with the district entirely in terms of the attitudes they have revealed in interviews. However, the difference of coefficients given here is almost as great when the policy agreement between the incumbent Congressman and his district is also measured in terms of the attitudes conveyed in confidential interviews.

¹² A third type of connection, excluded here, might obtain between district and Congressman if the Representative accedes to what he thinks the district wants because he believes that to be what a representative ought to do, whether or not it is necessary for re-election. We leave this type of connection out of our account here because we conceive an influence relation as one in which control is not voluntarily accepted or rejected by someone subject to it. Of course, this possible connection between district and Representative is not any the less interesting because it falls outside our definition of influence or control, and we have given a good deal of attention to it in the broader study of which this analysis is part.

¹³ It scarcely needs to be said that demonstrating some constituency influence would not imply that the Representative's behavior is wholly determined by constituency pressures. The legislator acts in a complex institutional setting in which he is subject to a wide variety of influences. The constituency can exercise a genuine measure of control without driving all other influences from the Representative's life space.

¹⁴ We have done this by a variance-component technique similar to several others proposed for dealing with problems of this type. See especially Simon, Herbert A., “Spurious Correlation: A Causal Interpretation,” Journal of the American Statistical Association, Vol. 49 (1954), pp., 467–479 Google Scholar; Blalock, Hubert M. Jr., “The Relative Importance of Variables,” American Sociological Review, Vol. 26 (1961), pp. 866–874 CrossRefGoogle Scholar; and the almost forgotten work of Wright, Sewall, “Correlation and Causation,” Journal of Agricultural Research, Vol. 20 (1920), pp. 557–585 Google Scholar. Under this technique a “path coefficient” (to use Wright's terminology, although not his theory) is assigned to each of the causal arrows by solving a set of equations involving the correlations of the variables of the model. The weight assigned to a full path is then the product of its several path coefficients, and this product may be interpreted as the proportion of the variance of the dependent variable (roll call behavior, here) that is explained by a given path.

A special problem arises because influence may flow in either direction between the Congressman's attitude and his perception of district attitude (as noted above, the Representative may tend both to perceive his constituency's view selectively, as consistent with his own, and to change his own view to be consistent with the perceived constituency view). Hence, we have not a single causal model but a whole family of models, varying according to the relative importance of influence from attitude to perception and from perception to attitude. Our solution to this problem has been to calculate influence coefficients for the two extreme models in order to see how much our results could vary according to which model is chosen from our family of models. Since the systems of equations in this analysis are linear it can be shown that the coefficients we seek have their maximum and minimum values under one or the other of the limiting models. Therefore, computing any given coefficient for each of these limiting cases defines an interval in which the true value of the coefficient must lie. In fact these in-tervals turn out to be fairly small; our findings as to the relative importance of alternative influence paths would change little according to which model is selected.

The two limiting models with their associated systems of equations and the formulas for computing the relative importance of the three possible influence paths under each model are given below.

¹⁵ By “least favorable” we mean the assumption that influence goes only from the Congressman's attitude to his perception of district attitude (Model I) and not the other way round. Under this assumption, the proportions of the variance of roll call behavior accounted for by the three alternative paths, expressed as proportions of the part of the variance of roll call votes that is explained by district attitude, are these:

Inverting the assumed direction of influence between the Congressman's own attitude and district attitude (Model II) eliminates altogether the effect that the Representative's attitude can have had on his votes, independently of his perception of district attitude.

¹⁶ Under both Models I and I I the proportion of the variance of roll call voting explained by the influence path involving the Representative's own attitude is twice as great as the proportion explained by influence, passing through his perception of district attitude.

¹⁷ What is more, the electorate's awareness of Congress as a whole appears quite limited. A majority of the public was unable to say in 1958 which of the two parties had controlled the Congress during the preceding two years. Some people were confused by the coexistence of a Republican President and a Democratic Congress. But for most people this was simply an elementary fact about congressional affairs to which they were not privy.

¹⁸ For a more extended analysis of forces on the congressional vote, see Stokes, Donald E. and Miller, Warren E., “Party Government and the Saliency of Congress,” Public Opinion Quarterly, Vol. 26 (Winter 1962), pp. 531–546 CrossRefGoogle Scholar.

¹⁹ For an account of this episode see Silverman, Corinne, “The Little Rock Story,” Inter-University Case Program series, reprinted in Bock, Edwin A. and Campbell, Alan K., eds., Case Studies in American Government (Englewood Cliffs, 1962), pp. 1–46 Google Scholar.

²⁰ The sample of this constituency was limited to twenty-three persons of whom thirteen voted. However, despite the small number of cases the probability that the difference in awareness between this constituency and the country generally as the result only of sampling variations is much less than one in a thousand.

²¹ In view of the potential nature of the constituency's sanctions, it is relevant to characterize its influence over the Representative in terms of several distinctions drawn by recent theorists of power, especially the difference between actual and potential power, between influence and coercive power, and between influence and purposive control. Observing these distinctions, we might say that the constituency's influence is actual and not merely potential since it is the sanction behavior rather than the conforming behavior that is infrequent (Dahl). That is, the Congressman is influenced by his calculus of potential sanctions, following the “rule of anticipated reactions” (Friedrich), however oblivious of his behavior the constituency ordinarily may be. We might also say that the constituency has power since its influence depends partly on sanctions (Lasswell and Kaplan), although it rarely exercises control since its influence is rarely conscious or intended (Cartwright). In the discussion above we have of course used the terms “influence” and “control” interchangeably.

²² The factors in American electoral behavior that encourage such a departure are discussed in Stokes and Miller, loc cit.