Lodge and his colleagues have proposed an impression-driven model of candidate evaluation (Lodge, McGraw, and Stroh, 1989) and vote choice (Lodge, Steenbergen, and Brau, 1995) whereby candidate evaluations are updated automatically "on-line" as new information about the candidates is encountered, but then the information itself, the reasons for any change in candidate evaluation, can be discarded. While we generally support the cognitively limited information processing theories upon which this model is based, we argue that it has rarely been tested in an experimental context that comes anywhere close to approximating an actual political campaign-dynamic, lasting over several months, with the voter having great discretion over the amount and nature of information learned about the candidates. Nor, for that matter, has the alternative memory-based moment-of-decision model been properly examined.

We devise a much more realistic test of the on-line model over the course of a 10-week panel study employing a dynamic process tracing program that allows subjects to choose what they wish to learn from a large menu of possibilities. During the first week of the study information about subject's political preferences and general political knowledge/sophistication is gathered. Subjects then choose to vote in either a mock Democratic or Republican primary race leading up to the 2016 U.S. presidential election. Over the next 8 weeks, new information about the competing candidates is released on a weekly basis, but subjects have almost complete discretion over how much or little of that information to examine. In the final week of the study, subjects evaluate both candidates and vote.

We predict that when voters get to choose what they want to learn about candidates, the contents of memory will be a much better indicator or evaluation and choice than any plausible operationalization of online evaluation.