The autonomy‐validity dilemma in mechanical prediction procedures: The quest for a compromise

Abstract

A robust finding in psychological research is that combining information with a mechanical rule results in more valid predictions than combining information holistically in the mind. Nevertheless, information is typically combined holistically in practice, resulting in suboptimal predictions and decisions. Earlier research showed that decision makers are more likely to use mechanical prediction procedures when they retain autonomy in the decision-making process. However, it remains largely unknown how different autonomy-enhancing features affect predictive validity. Therefore, in two preregistered studies (total N = 342), we investigated if and how prediction procedures can be designed such that they satisfy decision makers' autonomy needs and acceptance without reducing predictive validity. Based on archival

“Adding an Egg” in Algorithmic Decision Making: Improving Stakeholder and User Perceptions, and Predictive Validity by Enhancing Autonomy

“…We investigated (1) when and why stakeholders would appreciate decision-makers' algorithm use (Study 1) and (2) when and why decision makers would be less worried about negative stakeholder evaluations when using algorithms (Study 2). We expected that stakeholders would evaluate decision makers more positively if they retain autonomy in algorithm use, either by holistically adjusting predictions from a prescribed algorithm (clinical synthesis, Kuncel, 2018) or by using self-designed algorithms (i.e., by choosing predictor weights, Neumann, Niessen, Tendeiro, & Meijer, 2021). We henceforth refer to holistically adjusted predictions from a prescribed algorithm and self-designed algorithms as autonomy-enhancing algorithmic procedures (AEAPs).…”

“…Similarly, we expected that when decision makers retain autonomy in algorithm use, they would be less worried about negative stakeholder evaluations and more likely to use algorithmic procedures compared to a prescribed algorithm (no autonomy). A third important aim of this paper was to investigate whether AEAPs would result in more valid predictions than holistic predictions (Dietvorst et al, 2018;Neumann, Niessen, Tendeiro, & Meijer, 2021). This is crucial to investigate since improving decision-makers' meta-beliefs by enhancing their autonomy in algorithm use is only useful if the resulting predictions are more valid than holistic predictions.…”

“…This is expected because hiring decisions would not exclusively be determined by an external decision aid (i.e., a prescribed algorithm), but partly by the decision-maker's internal cognitive processes. Some studies showed that decision makers perceived to have much more autonomy in, and were more likely to use, AEAPs compared to a prescribed algorithm (Dietvorst et al, 2018;Neumann, Niessen, Tendeiro, & Meijer, 2021;Nolan & Highhouse, 2014). Moreover, AEAPs can also be more valid than holistic prediction, although studies regarding predictive validity are rather scarce and have produced mixed results (Dietvorst et al, 2018;Neumann, Niessen, Tendeiro, & Meijer, 2021).…”

“…Some studies showed that decision makers perceived to have much more autonomy in, and were more likely to use, AEAPs compared to a prescribed algorithm (Dietvorst et al, 2018;Neumann, Niessen, Tendeiro, & Meijer, 2021;Nolan & Highhouse, 2014). Moreover, AEAPs can also be more valid than holistic prediction, although studies regarding predictive validity are rather scarce and have produced mixed results (Dietvorst et al, 2018;Neumann, Niessen, Tendeiro, & Meijer, 2021). What remains unknown is whether stakeholders evaluate decision makers who use AEAPs more positively than decision makers who use a prescribed algorithm.…”

“…Although holistically adjusting predictions from a prescribed algorithm usually decreases predictive validity compared to strict algorithm use (Dawes, 1971;Hoffman et al, 2017;Neumann, Hengeveld, et al, 2022), some evidence suggests that such adjusted predictions are still more valid than holistic predictions because decision makers anchor on the algorithm's predictions (Dietvorst et al, 2018;Neumann, Niessen, Tendeiro, & Meijer, 2021). Moreover, in one of two studies, Neumann, Niessen, Tendeiro, and Meijer (2021) found that predictions from a self-designed algorithm were slightly more valid than pure holistic predictions, but only when participants knew predictor validities. Moreover, some very intriguing research has also shown that predictions from algorithms with randomly chosen predictor weights were more valid than holistic predictions (Dawes & Corrigan, 1974;Yu & Kuncel, 2020), due to increased judgment consistency.…”