16

In the Table 4 estimates of the LAST POLL variable has been regressed on vote

intention. The same controls used in estimating favorability have been used in the

estimates of vote intention. The models control for sex, race, direction of party

identification and the 3 variables for general media exposure. Table 4 is subdivided by

education levels and Table 5 by the respondent’s subjective willingness to follow politics.

As one might suspect looking across the Pseudo R^2 values of the models in

Table 4, the better educated an individual is, the better our model will predict the

variance in vote preferences. Looking back on original claims however, in this case our

expectation for the necessary sophistication of minimal cognitive effort utility

maximizers has been met exactly as predicted. The coefficients are unimodal and

centered on the middle of the educational scale (some college education). Moreover, in

both the lowest levels and the highest levels of educational sub-groupings the LAST

POLL variable does not successfully predict a respondent’s vote intention.

In the case of the three middle categories of education some further discussion is

required. While both the categories of those with “High School diplomas” and those with

“4 year degrees” in the table are annotated as having p-values less than .1, the statistical

significance merits further discussion. The P-Value

4

for “High School” is .057, or we

have a 5.7% chance of making an invalid inference from the Annenberg sample to the

population of all people with a high school education and no more. Critics might say that

a 5.7% chance of a Type I error in a sample of 7544 is too high to give further

consideration. It is respectfully submitted that in this case it is not. The substantive

impact of the variable suggests that a 10 point shift towards John Kerry in the value of

LAST POLL would lead this demographic to express a 3.7% greater likelihood to support

4

P-Vaues not shown in Tables, but can of course be calculated from standard errors.