A Computational Model of the Citizen as Motivated Reasoner: Modeling the Dynamics of the 2000 Presidential Election

Our model is developed within the ACT-R cognitive architecture (The A daptive C haracter of T hought – R ational; Anderson et al. 2004), which is a leading theoretical and modeling framework used in cognitive science for a wide range of learned behaviors, among them language comprehension, the recognition and recall of information, inferencing, the formation of beliefs, and the learning of complex skills. However, while ACT-R provides comprehensive, integrated sets of cognitive mechanisms for learning, it lacks affective mechanisms, which are essential to current theories of political information-processing, and it lacks specific mechanisms for preference updating. Consequently, much of our work developing JQP was devoted to building affective and updating mechanisms and integrating them with the cognitive processes in ACT-R.

This notation for random noise is conventional in computational modeling because it makes transparent the manipulation of the normal density function.

S _ji represents the strength of association from node j to node i. It is an increasing function of the number of times node j has sent activation to node i, and it is not symmetrical (Anderson 1993). C _ji  = 1 when nodes j and i share the same valence (positive or negative), C _ji  = −1 when they have different valences, and C _ji  = 0 when either of them is neutral.

The number of buffers could be manipulated to represent greater or lesser cognitive limitations, but we do not explore this implication here.

To test the internal validity of the model (i.e., whether the model behaves as expected in our theory), we conducted a series of purely formal computational experiments (Kim 2005) in which the model successfully reproduced practice, recency, and spreading of activation effects on recall; cognitive and attitude priming effects; question order and wording effects in the survey response; both on-line and memory-based processing; and the ability to learn by adjusting beliefs and attitudes in response to campaign events. These tests were essential to establish that the model is in fact consistent with a wide range of “well-known phenomena” from the empirical cognitive literatures. These tests are available from the first author.

For the trait perceptions to be useful, we need to assign a value to each trait. That is, when a respondent says “Bush is trustworthy,” we need to have some idea how positive this attribution is. Fortunately, most human traits and other general concepts have been normed by large samples of respondents, so we can scale the evaluative implications of a given trait (positivity, negativity, intensity, etc.) by consulting the Affective Norms for English Words (ANEW; Bradley and Lang 1999) and N. Anderson (1968), which provide means and standard deviations for a large number of trait concepts.

Recall also that transient random noise is added to the activation levels of all objects in memory each time an object is retrieved (Eq. 1), which also approximates individual differences in processing.

The average of correlation between the actual and simulated evaluations over time and that across groups was used as a fit measure. However, the same qualitative results were obtained with wide ranges of parameter values.

The parameter search space was extensive; the values within the interval [0, 10] with stride 0.1 for P ₀ , h, and q, and those within [0, 1] with stride 0.01 for γ were first examined. Then, the values within [0, 1] with stride 0.01 for γ, h, and q were examined in a refined search. For very small values of h and q (0, 0.01, and 0.02), simulation results and model fits were almost identical. With the optimized parameter values, the weight, K _t , did not converge during the simulation.

The values within [0, 1] with stride 0.01 were examined.

Formally, JQP will be a motivated reasoner as long as γ > 0, 0 < δ < 1, 0 < ρ < 1 and its belief structure is reasonably consistent.