Salience effects in memory-based decisions: an improved replication

ABSTRACT

An experimental report by Platzer and Bröder ([2012]. Most people do not ignore salient invalid cues in memory-based decisions. Psychonomic Bulletin & Review, 19(4), 654–661) claimed that in memory-based decisions, salient attributes are often not ignored even if they are less valid than other cues. When the rank order of cue validities was congruent with their salience hierarchy, people predominantly used a noncompensatory take-the-best strategy (TTB) based on the most valid cue whereas they used more compensatory strategies when hierarchies were incongruent (i.e. the least valid cue was most salient). Given the recent replication crisis in psychology and methodological shortcomings of the original study, a better-controlled replication with new stimuli and a larger sample was conducted. A pilot study established a visual salience hierarchy of the cues used. The main experiment clearly replicated the salience effect at the strategy selection level and the longer response times for compensatory strategies compared to TTB. A response time interaction of strategy and condition did not replicate.

KEYWORDS:

• Decision making
• memory
• visual salience
• spatial attention
• accessibility

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 This consensus notwithstanding, improvement is of course desirable in other areas of psychology as well, e.g. in theory development and theoretical integration.

2 A search on Google Scholar on December 18th 2020 yielded 37 citing papers of authors not affiliated with our group (47 altogether).

3 In the replicability debate, conceptual replications have been criticized for being inherently ambiguous in their interpretation, especially if the replication fails. This leaves room for proponents of an initial result to explain away the failure. However, both kinds of replications pursue different goals, namely a robustness check in the case of direct replications and the generalizability across operationalizations in conceptual replications. The latter should be the ultimate goal when investigating the scope of a hypothesis (e.g. searching for moderators), but admittedly, a generalization test is only worthwhile if the to-be-generalized effect is robust in the first place.

4 Gigerenzer and Goldstein (1996), however, emphasize that the variability on the accuracy dimension is much less pronounced than commonly assumed.

5 However, compensatory strategies can also often be modeled by assuming parallel processing (e.g. Glöckner & Betsch, 2008).

6 We became aware of this result during the revision of the current manuscript.

7 The cartoon characters were taken from the well-known German TV series for children “Die Sendung mit der Maus” (“The broadcast with the mouse”) that has been on air since 1971. Hence, the characters were probably known to most of our participants. Example Screenshots of all experimental phases can be found on https://osf.io/gpsuj/

8 Analyses were conducted with the lme4 (Bates et al., 2015) package and the lmerTest package (Kuznetsova et al., 2017) in R (R Core Team, 2019).

9 The confidence intervals for the latter two cues overlap, however.

10 A difference may emerge even in the verbal condition, however, if participants have preconceptions about the cue validities that interfere with the instructed validities in at least one of the conditions. Nevertheless, the difference should be larger in the pictorial condition if an additional salience effect was at work.

11 With four binary cues, 16 different patterns can be generated, a full comparison of which yields 120 decision situations. Of these, only 9 can differentiate between TTB and compensatory strategies, 27 differentiate between WADD and EQW, and the rest yields identical predictions for all strategies. To keep participants’ memory load manageable, we selected the 10 patterns which yield a maximal amount of eight decision situations differentiating between TTB and the compensatory strategies. These were presented twice. See Bröder and Schiffer (2003), for details.

12 We fell somewhat short from the intended 120 participants for organizational reasons.

13 p-values are based on 100,000 Monte Carlo samples to approximate an exact test.

14 Pooling across the three control conditions like in the original study also yields a significant difference in the expected direction χ²(1, N = 114) = 6.66, p = .012 (1-β = .89 to detect a medium-sized effect of w = .30 and > .99 to detect a large effect w = .50).