Coherence shifts in attribute evaluations.

Abstract

In five experiments, people repeatedly judged individual options with respect to both overall value and attribute values. When required to choose between two snacks, each differing in two attributes (pleasure and nutrition), people's assessments of value shifted from pre-to post-choice in the direction that spread the alternatives further apart so as to favor the winner, thereby increasing confidence in the choice. This shift was observed not only for ratings of overall value, but also for each of the two individual attributes. The magnitude of the coherence shift increased with choice difficulty as measured by the difference in initial ratings of overall value for the two options, as well as with a measure of attribute disparity (the degree to which individual attributes "disagree" with one another as to which option is superior). In Experiments 2-5, tasks other than explicit choice generated the same qualitative pattern of value changes, confidence, and response time (RT). These findings support the hypothesis that active consideration of options, whether or not explicitly related to value, automatically refines the mental value representations for the options, which in turn allows them to be more precisely distinguished when later included in a value-based choice set.

Evidence Accumulates for Individual Attributes during Value-Based Decisions

“…Consistent with this idea, recent work has shown that the disparity of the options' attribute compositions affects multi-attribute decision making (Lee & Holyoak, 2021). A pair of options has high disparity if, for example, one option scores high in the first attribute dimension but low in the second, while the other option scores high in the second dimension but low in the first (Figure 1a, left panel).…”

“…In addition, our model simulation and comparison results show that the representations that determine an option's subjective value differ between rating and binary choice tasks. From the simulation tests, it is clear that decision models based on a unitary overall value do not generate the influence of disparity (Lee & Holyoak, 2021) on choices or response times seen in the empirical data. In terms of quantitative model comparisons, if the overall values of the foods in these six studies were invariant across the ratings and choice tasks, then DDMs using only the overall values to determine the drift rate (Model 1) would be preferred over DDMs using either a subset of the food attributes (i.e., Model 2 with pleasure and nutrition) or individual attributes plus the overall values (Model 3).…”

“…We analyzed the data from six previously published experiments (Experiments 1-5 in (Lee & Holyoak, 2021), plus one unpublished pilot experiment that we label Experiment 0). The total number of participants across the six datasets was 307, using the same exclusion criteria from the original study (41 for Experiment 0, 50 for experiment 1, 48 for experiment 2, 54 for experiment 3, 60 for experiment 4, and 54 for experiment 5).…”

“…Notably, two decisions could be equally difficult in the traditional sense that the overall value ratings of the choice options are equally close together, yet have very different levels of disparity (Figure 1a, right panel). In multiple independent experiments, Lee & Holyoak (Lee & Holyoak, 2021) found that choice behavior differs as a function of disparity, such that higher disparity corresponds to higher choice consistency (i.e., a choice in favor of the option that was previously rated as having the higher overall value) and lower response time, even after accounting for differences in overall values.…”

“…If instead attributes are weighted differently during choices compared to ratings, then the flexibility to estimate attribute-specific decision weights may give Model 2 the advantage even though it is based on only two attributes out of a larger set. Across all six datasets (studies 0-5 from(Lee & Holyoak, 2021)), the participant population was split evenly between supporting Models 1 and 2 (estimated model frequency of 0.50 for each; see Figure3A).…”