Past studies have offered diverse estimates of the role of policy preferences, party loyalties, candidate personalities and other factors in voting decisions. Most have postulated recursive (that is, one-way) causal relationships among the central variables. This study specifies a non-recursive simultaneous equation model and estimates its parameters for the 1972 and 1976 elections using CPS data. The estimates differ markedly from those of simple recursive models. Policy preferences appear to have much more influence on voting decisions, and party attachments much less, than was previously thought. Candidate evaluations strongly affect voters‘ perceptions of closeness to candidates on policy issues. Party identification may be influenced by short-term factors. Differences between 1972 and 1976 reflect the issue-oriented McGovern candidacy. Simultaneous equation models offer no cure-all; in the absence of accepted theory many specifications are open to controversy. But future research must take account of reciprocal causal paths.
Nor is The Changing American Voter immune to this criticism (Nie, Verba and Petrocik, 1976). Comparisons between groups or candidates or over time are also subject to error because the biases in estimates are not necessarily constant from group to group or candidate to candidate or year to year. We have made the sweeping claim that virtually all studies of policy orientations, partisanship and the vote-certainly including the primitive efforts reported in Figures 2 through 5 — are subject to simultaneity bias and are potentially quite misleading. It is fair enough to ask whether we have anything constructive to add to this work of destruction. The most appropriate way to handle the problem, we would argue, is through the use of non-recursive, simultaneous equation models which explicitly allow for the possibility of causal processes operating in both directions between variables (Johnston, 1972; Theil, 1971; Hanushek and Jackson, 1977; Duncan, 1975). We can begin to apply such techniques to the voting problem by specifying the central set of variables which we believe to be mutually endogenous — that is, which reciprocally affect each other: comparative candidate evaluations, relative policy distance between the voter and candidates, and current subjective partisanship, We will continue to treat measures of reported vote as direct consequences of overall candidate evaluations.
Among the most noteworthy aspects of Figure 9 is the nearly total absence of any effect of party attachments upon the other two endogenous variables. Our estimates reveal that in 1972 — quite unlike 1976 — party loyalties played no part in the formation either of voting decisions or perceptions of closeness to the candidates on policy matters. McGovern dramatically dissociated himself from the Democratic party’s mainstream; and, by the same token, the central core of the Democratic party abandoned McGovern to his own devices. Party per se had no independent effect at all on the vote. Furthermore, the absence of a significant effect of policy distances upon partisanship, together with the very strong impact of policy distances upon intended vote, indicates that the policy choices around which voting decisions revolved (the Vietnam War, urban unrest, campus disturbances, alternative life styles) cut across the grain of older party cleavages.
There was, on the other hand, apparently some effect of intended votes upon partisanship. That is, although party loyalties did not affect votes, intended votes did affect partisanship — presumably because some Democratic defectors to Nixon felt their Democratic affiliations to be weakened. There are, however, some doubts about this finding. In spite of the moderate size of this coefficient (.29), the conventional significance test (which is asymptotically valid for two- and three-stage least squares estimation) indicates that it is not significantly different from zero at the .05 confidence level. Examination of the intermediate two-stage calculations reveals a high degree of collinearity between the „decontaminated“ versions of the policy distance and candidate evaluations variables. This results in high standard errors of estimates for the respective coefficients (i.e., lowers their precision), which substantially raises the difficulty of disentangling the independent effects for these two variables. Since we presently lack any satisfactory way to deal with multicollinearity in the second stage of two-stage least squares, it is necessary to reserve judgment on the significance or insignificance of the 1972 impact of intended vote upon partisanship.