18
Campaign variables are considered separately from mobilization variables in Table 2 so
as to draw attention to the different expectations that apply to these variables when measures
relating to the character of elections are introduced into the analysis. Campaign variables may
well carry some influences from the character of elections if important elections serve to
stimulate the activities and concerns that campaign variables measure.
All variables have been recoded where necessary to ensure that larger values correspond
to higher probabilities of voting (except in the case of variables whose values have substantive
meanings that call for negative correlations) and rescaled so that all of them range from 0 to 1.
Thus a 1-point shift in the value of any variable corresponds to a shift from the minimum to the
maximum value found in the data, permitting estimated effects to be given a straightforward
substantive interpretation that is comparable across variables.
21
This table provides some preliminary (if very indirect) support for the main hypothesis. If
district-level measures of competitiveness influence turnout but nationwide measures do not (as
indicated by the correlations with voting for variables at the top of the table), this suggests
variations in expressive engagement from district to district according to the closeness of the race
in specific districts. This would make sense, as will be explained below (see footnote 28). Yet to
be seen is whether such mobilizing efforts have particularly strong effect on new cohorts, as
hypothesized. This is the question to which we now turn.
Findings
Table 3 shows the results of an analysis that attempts to classify respondents into those who
voted and those who did not on the basis of logistic regression analysis. The reason OLS
regression cannot be employed for this analysis is because the dependent variable is a dichotomy
(voted or not) and regression analysis can yield biased estimates with such a dependent variable,
especially if (as in this case) the proportion turning out to vote in particular subgroups of respon-
dents sometimes approaches 0 or 1.
22
Because the coefficients that are produced by a logistic
regression analysis are uninterpretable by mere mortals, they have been converted in this table to
21
In interactions with new cohorts, some of the implied shifts might not occur in practice. In particular,
age does not range from 18 to 99 among new cohorts. Consequently a separate age variable was generated
for use in the interaction with new cohorts, whose range was rescaled to run from 18 to 32 (see Table 3).
22
As in earlier tables, the data are weighted to reproduce the actual historic turnout levels at each election.
Convention