Speed-accuracy trade-offs in sample-based decisions.

Abstract

Success on many tasks depends on a trade-off between speed and accuracy. In a novel variant, a speed-accuracy trade-off with sample-based decisions in which both speed and accuracy jointly depend on (self-truncated) sample size, we found strong accuracy biases. On every trial of a sequential investment game, participants chose between 2 investment funds based on binary samples of the funds’ past outcomes. Participants could stop sampling and decide whenever they felt sufficiently informed. Total payoff was the product of choice accuracy and number of choices completed within the available time (speed). Participants’ failure to understand the dominance of speed over accuracy—that speed decreases more than accuracy improves with increasing sample size—led to dramatic oversampling. Our research aimed to examine to what extent metacognitive functions of monitoring and control could correct for the accuracy bias. Experiments 1a through 1c demonstrated similarly strong accuracy biases and payoff losses in psychology and economics students, depressed, and control patients. In Experiments 2 through 4, the accuracy bias persisted despite several manipulations (feedback, sample limit, choice difficulty, payoff, sampling truncation as default) that underlined the speed advantage, reflecting a conspicuous metacognitive deficit. Even when participants faced no risk of losing on incorrect trials but could still win on correct trials (Experiment 3) and when sampling was contingent on the active solicitation of every new element (Experiment 4), participants continued to sample too much and failed to overcome the accuracy bias. The final discussion focuses on psychological reasons and possible remedies for the metacognitive deficit in trade-off regulation.

On the difficulty of overcoming one’s accuracy bias for choosing an optimal speed–accuracy tradeoff.

“…That errors were largely avoided in our experiment, although this lowered the reward, supports the hypotheses that people have an accuracy bias. A plausible reason for this bias is the negative connotation of errors (Fiedler et al, 2020), which evokes negative emotions after each response error (Johnson et al, 2017). Accordingly, it seems that the strategy of most participants was to optimize their well-being during the experiment rather than their monetary reward.…”

“…That is, they try to avoid errors, even if this is disadvantageous for their reward. A possible reason for this bias is that errors have a negative connotation (Fiedler et al, 2020) and, therefore, elicit negative emotions (e.g., Dignath et al, 2020). Thus, it seems that, rather than optimizing reward, people try to optimize or, at least, maintain their wellbeing, that is, to gain some reward while limiting error-related negative emotions.…”

“…Already Swensson (1972a) argued that fast responses might be experienced as fast guesses and, therefore, considered as cheating, which evokes unpleasant feelings. Recently, Fiedler et al (2020) proposed a similar account. They argue that the accuracy bias is due to the social surplus meaning of accuracy, which implies carefulness and responsibility.…”

Toward Predicting Human Performance Outcomes From Wearable Technologies: A Computational Modeling Approach

“…Raw data to define marginal distributions Brunyé et al, 2013;Giles et al, 2014;Mathiassen et al, 2014;Solis et al, 2015;Inoue et al, 2016;Harris et al, 2017;Hutchinson et al, 2017;Klein et al, 2017;Rosenbaum et al, 2017;Heuberger et al, 2018;Sandra and Otto, 2018;Schumacher et al, 2018;Angelidis et al, 2019;Bock et al, 2019;Goldfarb et al, 2019;Okano et al, 2019;Pyke et al, 2019;Sanabria et al, 2019;Wei et al, 2019;Barrett et al, 2020;Baumert et al, 2020;Fiedler et al, 2020;Holgado et al, 2020;Johnson et al, 2020;Knelange and López-Moliner, 2020;Larsen et al, 2020;Lin H. et al, 2020;Madore et al, 2020;Pahwa et al, 2020;Rodas and Greene, 2020;Rodeback et al, 2020;Timme and Brand, 2020;Tsukahara et al, 2020;Vine et al, 2020;von Helversen and Rieskamp, 2020;Pavlov and Kotchoubey, 2021 Effect size estimates to define xy relationships Lisper and Kjellberg, 1972;Glenville et al, 1978;Larsson, 1989;…”

Self-Regulation

“…Owing to their limited capabilities, human agents need to balance selfregulatory processing rate and schematic complexity, because both consume limited resources. Put simply, owing to limited capabilities, the higher the processing rate, the lower the schematic complexity, and vice versa (Fiedler et al, 2020). This results in two major options.…”