Apparent sunk cost effect in rational agents

Abstract

Rational decision makers aim to maximize their gains, but humans and other animals often fail to do so, exhibiting biases and distortions in their choice behavior. In a recent study of economic decisions, humans, mice, and rats have been reported to succumb to the sunk cost fallacy, making decisions based on irrecoverable past investments in detriment of expected future returns. We challenge this interpretation because it is subject to a statistical fallacy, a form of attrition bias, and the observed behavior can be explained without invoking a sunk cost-dependent mechanism. Using a computational model, we illustrate how a rational decision maker with a reward-maximizing decision strategy reproduces the reported behavioral pattern and propose an improved task design to dissociate sunk costs from fluctuations in decision valuation. Similar statistical confounds may be common in analyses of cognitive behaviors, highlighting the need to use causal statistical inference and generative models for interpretation.

Sunk cost sensitivity in mice, rats, and humans on the Restaurant Row and WebSurf tasks cannot be explained by attrition biases alone

“…Used with permission of the publisher. (Ott et al 2021) already acknowledge in their preprint that their model cannot explain behavior in the variants of Restaurant Row and Web-Surf that have a separate offer phase, and agree with our conclusions that decisions made in the offer and wait phases arise from different decision processes. As we have argued elsewhere Kazinka, MacDonald, and Redish 2021;Sweis, Abram, et al 2018), we agree that behavior in the offer and wait phases arise from different decision processes, consistent with both multiple-decision theories (Redish 2013 and commitment hypotheses (Gollwitzer, Heckhausen, and Steller 1990;Achtziger and Gollwitzer 2007).…”

“…As can be seen from the data analyses above, the Ott et al model (Ott et al 2021) cannot explain the behaviors we see within variants of the Restaurant Row and WebSurf tasks. We therefore conclude that the sunk cost sensitivity is a real effect showing a sensitivity to sunk costs (as defined by an escalation of commitment, (Staw 1976;Staw and Fox 1977;Staw and Ross 1989)) and is not an epiphenomenal consequence of the simple explanation suggested by Ott et al (Ott et al 2021).…”

“…In this paper, we analyze their model and compare it to the extensive data we have from Restaurant Row, Web-Surf, and follow-up variants (such as variants with and without an offer phase (Steiner and Redish 2014;Schmidt, Duin, and Redish 2019;Sweis, Abram, et al 2018), the Movie Row task, in which subjects navigate spatially in virtual reality to view movies (Huynh et al 2021), the Candy Row task, in which subjects wait out delays to receive small pieces of candy (Huynh et al 2021), and the Known-Delay/Randomized-Delay tasks, in which we manipulated the predictability of the upcoming delay in each restaurant (Duin et al 2021)). While we applaud the mechanistic simplicity of the (Ott et al 2021) model, we find that their model does not fit key aspects of the data. We therefore conclude that their model does not explain the sunk cost behavior that we are observing in these tasks.…”

“…In a recent bioRxiv preprint, Ott, Masset, Gouvêa, and Kepecs proposed that this observed behavioral sensitivity to sunk costs could be an epiphenomenon consequence of a combination of attrition statistics, a drift diffusion model, and a changing threshold for quitting (Ott et al 2021). In this paper, we analyze their model and compare it to the extensive data we have from Restaurant Row, Web-Surf, and follow-up variants (such as variants with and without an offer phase (Steiner and Redish 2014;Schmidt, Duin, and Redish 2019;Sweis, Abram, et al 2018), the Movie Row task, in which subjects navigate spatially in virtual reality to view movies (Huynh et al 2021), the Candy Row task, in which subjects wait out delays to receive small pieces of candy (Huynh et al 2021), and the Known-Delay/Randomized-Delay tasks, in which we manipulated the predictability of the upcoming delay in each restaurant (Duin et al 2021)).…”

“…Their model (Ott et al 2021) propose that they can replicate the sunk cost behavior reported in (Sweis, Abram, et al 2018) with a simple model in which they assume (1) that the subject has a willingness to wait W which distributes normally (W = W 0 + N(σ W )), (2) that the decision to enter an offer occurs when W is greater than the threshold for that offer T OZ , (3) that W then wanders with a given variance (σ N ), and the subject quits if W ever crosses a quit-threshold in the wait zone T WZ . Although they do not discuss the importance of the shape of the quit threshold T WZ in their preprint, they assume that the quit-threshold T WZ decreases with time so that it starts at the entry threshold T WZ (t=0) = T OZ and reaches 0 at the time of reward T WZ (t=T OZ ) = 0.…”