Political Knowledge and Electoral Choice in the 1992–2004 United States Presidential Elections: Are More and Less Informed Citizens Distinguishable?

Abstract

This study estimates an individual-level model of turnout and vote choice in United States presidential elections. Extending a methodology proposed by Bartels (1996), the model allows one to assess whether the electoral choices of citizens would change if they were more knowledgeable about politics, and if such changes have implications for aggregate election results. Simulations of hypothetical electorates under different assumptions about the distribution of political knowledge show that while some citizens would change their votes if more knowledgeable, the primary effects of increasing voter knowledge is to raise turnout levels and to solidify preexisting vote tendencies. The few vote changes that result from increased political knowledge largely average out in aggregation. Increased turnout resulting from a more informed electorate, however, favors Democratic candidates in two of the four studied elections.

[A] people who mean to be their own governors, must arm themselves with the power which knowledge gives.1

-James Madison

Notes

James Madison letter to William T. Barry, August 4, 1822.

I use the term “observed” knowledge to reinforce that K _i is defined at the level of respondent and corresponds to that level of measured political knowledge actually possessed by each respondent. The “average” level of knowledge in the sample is the arithmetic average of K _i over all respondents.

The dependent variable is vote choice as reported in the ANES post-election interview. The majority of items used to construct the knowledge scales are also obtained from the post-election interview. The remaining variables including demographics are obtained from the pre-election survey.

The computational complexity of the estimated maximum likelihood model requires that the sample size be as large as possible. To avoid reducing the sample size by list-wise deletion of observations due to occasional missing values in the explanatory variables, I substitute the mean value of the variable when a missing value is encountered. The variable that returns the largest percentage of missing values is liberal–conservative ideological self-placement. The remaining variables return only a negligible number of missing values relative to the sample size. Bartels (1996) also appears to use this strategy to obtain the estimates that he reports.

To calibrate the model I attempted to replicate Bartels's (1996) findings for the 1992 and 1988 election cycles. Using numerous coding rules, I was unable to recreate datasets corresponding to those used in Bartels's original study. Depending on the independent variable coding rules, the differences in observations were typically modest: generally 50 or 60 observations in a dataset of several hundred observations. This naturally means that our ML parameter estimates differ. In some cases, these differences are minor, but in some cases these were more significant. Despite these differences, I am very confident my ML approach is identical to that proposed by Bartels (1996) and my finding may be interpreted as informing the same questions posed in his original study.

For a detailed discussion of the construction of the political knowledge scale used here, and the disciplinary debate over how best to measure political knowledge, see the discussion in Dow (2009).

The difference between the coefficients is distributed chi-square with one degree of freedom. The test statistic critical value for statistically significant differences at the p = 0.05 level is 3.8.

I estimate the model using multinomial logit (MNL) rather than multinomial probit (MNP) even though MNL, unlike MNP, imposes the restrictive independence of irrelevant alternatives (IIA) property. The IIA property assumes that ones choice between any two alternatives does not depend on the availability of other alternatives. Here this means that the probably that one votes Democratic relative to the probability that one votes Republican is unaffected by the presence or absence of abstention option. Hausman tests (Hausman & McFadden, 1984; Long, 1997) show the IIA property holds in the 1992–2000 specifications, and that MNL estimation is appropriate. However, in 2004 the IIA assumption may be violated. Even here, however, MNL is still preferred to MNP. This is because even though MNP relaxes the IIA restriction, it presents disadvantages that must be balanced against the benefits of relaxing the IIA assumption. The most significant of these is that in small to medium sized samples, the MNP does not always converge to its global maximum, and may produce misleading or inaccurate estimates. Such MNP estimation problems are often difficult to diagnose, and obtaining proper estimates often involves placing much structure on the error covariance matrix reducing the practical advantages of the MNP specification (Dow & Endersby, 2004). Such is the case here. Guidelines for choosing between MNL and MNP in studies where each presents advantages and liabilities are most clearly articulated by McFadden and reiterated by Cheng and Long (2007) who argue that the proper determinant for choosing between each estimation approach centers less on Hausman tests than on whether the researcher is justified in believing that choices “can plausibly be assumed to be distinct and weighed independently in the eyes of each decision maker” (McFadden, 1973: 113). In choosing whether to vote Democratic, Republican or abstain, it seems clear that MNL estimation is justified.

The predicted electoral choice for observed knowledge is that alternative with the greatest probability weight as determined by Equation 2a. The predicted electoral choice for high knowledge is that alternative with the greatest probability weight as determined by Equation 2b. Recalling that P_D + P_R + P_A = 1, if, for example, P_D > P_R and P_D > P_A, then the predicted choice is vote Democratic. The calculation is identical for predicting a Republican vote or abstention.

The remaining probability weight is distributed between the probability of abstention and the probability of voting Republican. In this example, the mean probability of voting Republican is about 0.08 and the mean probability of abstention is 0.44.