When several choice options are sampled one at a time in a sequence and a single choice of the best option is made at the end of the sequence, which location in the sequence is chosen most often? We report a large-scale experiment that assessed tasting preferences in choice sets of two, three, four, or five wines. We found a large primacy effect—the first wine had a large advantage in the end-of-sequence choice. We also found that participants who were knowledgeable about wines showed a recency effect in the longer sequences. We conclude with a process model that explains our findings.
We propose that two biases operated within that sequential competitive evaluation process. First, a first-is-best bias accounts for the consistent primacy effect. Second, a bias in favor of each new wine among high-knowledge participants accounts for the recency effect, and for an interesting reason: Compared with the low-knowledge participants, the high-knowledge participants were more persistent in looking for a better wine later in the sequence—a plausible result of greater expertise. Thus, high-knowledge participants were likelier to make a comparison between their current favorite and the new wine when each new wine was sampled. Thus, there was a substantial chance that each new wine would beat the current favorite, and this habit produced the pronounced recency effect in longer sequences, especially for high-knowledge participants. For example, suppose that each new wine has a .30 chance of beating the current favorite, and the current favorite remains the preferred choice with a .70 probability. Note that these values are consistent with the size of the observed primacy effects (e.g., the first wine was chosen with approximately .70 probability in the two-option sets) and with the recency effects for the high-knowledge participants (the last wine was chosen with approximately .30 probability in the four-option and five-option sets).
We account for the lack of recency effects among the low-knowledge participants by proposing that they followed the pair-wise competitive-evaluation strategy less vigorously than the high-knowledge participants, eliminating the potential recency advantage. We speculate that the low-knowledge participants were more likely to be overwhelmed by the cognitive demands of the pair-wise competitive strategy as memory load and interference increased across the sampling trials.
The pair-wise model provides an almost perfect fit to the data if we add one more assumption about the comparison process. Thus far, we have assumed that all current favorites have a .70-versus-.30 advantage in all pair-wise comparisons. But if we suppose that the current-favorite advantage increases for later favorites (e.g., if the third option wins its pair-wise competition, its advantage increases to .75 vs. .25; if the fourth wins, its advantage is .80 vs. .20), then the model fits the data almost perfectly. This pair-wise-competition process model is impressive; its one failing is that it predicts a small recency effect for the three-option set for high-knowledge participants.