Measuring cognitive effort without difficulty

Abstract

An important finding in the cognitive effort literature has been that sensitivity to the costs of effort varies between individuals, suggesting that some people find effort more aversive than others. It has been suggested this may explain individual differences in other aspects of cognition; in particular that greater effort sensitivity may underlie some of the symptoms of conditions such as depression and schizophrenia. In this paper, we highlight a major problem with existing measures of cognitive effort that hampers this line of research, specifically the confounding of effort and difficulty. This means that behaviour thought to reveal effort costs could equally be explained by cognitive capacity, which influences the frequency of success and thereby the chance of obtaining reward. To address this shortcoming, we introduce a new test, the Number Switching Task (NST), specially designed such that difficulty will be unaffected by the effort manipulation and can easily be standardised across participants. In a large, online sample, we show that these criteria are met successfully and reproduce classic effort discounting results with the NST. We also demonstrate the use of Bayesian modelling with this task, producing behavioural parameters which can be associated with other measures, and report a preliminary association with the Need for Cognition scale.

Keywords: Anhedonia; Cognitive effort; Computational psychiatry; Depression; Individual differences; New measures.

Fig. 1

Number Switching Task trial structure. Participants chose whether to perform an effortful task depending on the points and effort level offered. If they accepted the offer, they were shown a random sequence of the digits 1-9 and had to indicate (by pressing the “f” or “j” keys) whether each of the digits was even or odd. Sequences with more frequent switching between odd and even were more effortful. To win the points on offer, participants had to categorise at least 8 of the 9 digits correctly and complete the sequence within the allowed time (which was calibrated to each individual). In the above figure, the “alternative outcomes” show screens that participants saw if they passed an offer or if they failed the trial because of too many errors or timing out

Fig. 2

Overall structure of the different phases of the task. See main text for detailed description of each phase

Fig. 3

Number Switching Task: proportion of offers accepted. Mean, standard error, and distribution of the proportion of offers accepted for each combination of reward and effort level. See Fig. S9 for the same plot without faceting

Fig. 4

Number Switching Task: proportion of trials completed successfully. Mean, standard error, and distribution for each combination of reward and effort level. Trials were marked as “correct” if they were completed within the allowed time, with no more than one error. See Fig. S9 for the same plot without faceting

Fig. 5

Number Switching Task: completion time. Mean, standard error, and distribution of the completion times (expressed as a proportion of each participant’s allowed time) for each level of reward and effort level. See Fig. S9 for the same plot without faceting

Fig. 6

Subjective task load ratings for each effort level. Plots show (from left to right within each plot) the individual data points, the means and standard errors, and the distributions of scores for each of the six scales of the index

Fig. 7

Differences in WAIC relative to the best performing model (Model 9). Also shown is the standard error of this difference (black intervals). For simplicity, only the three best scoring models are shown above; the full plot of all eight models is given in Fig. S6. Model 9, the best performing model, contained a varying intercept, varying linear effects of reward and effort and a varying quadratic effect of reward

Fig. 8

Posterior distributions of the population level parameters from Model 9 (vI + vR + vR² + vE). The vertical line indicates the mean of each distribution and the shaded region the 89% quantile interval

Fig. 9

Relationship between the probability of success and effort sensitivity. The correlation was nonsignificant (p = 0.09), implying that effort sensitivity is not confounded by probability discounting in this task. Note the extreme point on the left of the graph corresponds to a participant who accepted (and failed) only one trial overall. In a sensitivity analysis, removing this participant increased the p-value of the correlation to p = 0.14