Day-to-day fluctuations in motivation drive effort-based decision-making

Abstract

Internal states like motivation fluctuate substantially over time. However, studies of the neurocomputational mechanims of motivated behavior have failed to capture this. Here, we examined how naturalistic ups and downs in state motivation influence the subjective value of reward and effort. In a microlongitudinal design (N = 155, state timepoints = 3,344, decision-making tasks = 845), we captured fluctuations in state and effort-based decision-making using smartphone-based momentary assessments as people went about their daily lives. We found that both state and trait have independent and multiplicative effects on decision-making. State-behavior coupling was particularly pronounced in individuals with higher trait apathy, meaning that their choices were even more state dependent. Using computational modeling, we demonstrate that state motivation prospectively boosted reward sensitivity, making people more willing to exert effort in future. Our results show that day-to-day fluctuations in state and cognition are tightly linked and critical for understanding fundamental human behaviors and mental ill-health.

Keywords: decision-making; effort; longitudinal; motivation; reward.

Fig. 1.

Microlongitudinal study design to determine the link between real-world, rapidly unfolding subjective states and decision-making. The study involved self-report of subjective state “right now” (A) and decision-making (B) over 15 d. This included a baseline session (day 0) with app enrollment, self-reported traits, maximum effort calibration, the first state and decision-making game. Follow-up began the next morning. (A) Participants received smartphone notifications twice per day (1 morning, 1 afternoon) to report subjective social and behavioral motivation right now (blue box, *reverse scored item), along with happiness, fatigue, and sleep quality (Methods). On alternate days (25% of study timepoints, orange marker), they also repeated the decision-making task and reported the state afterward. The order of task assessments was pseudorandomly counterbalanced between morning and afternoon (4 each). (B) Smartphone-based decision-making task. Trials begin with a choice between two routes varying in reward (coins) and effort (Left). Greater effort was manipulated as more button presses within 10 s (% of maximum). After making a choice, the participant presses POWER as fast as possible to fill the power bar within 10 s (Middle). When successful (success μ = 0.98), feedback is displayed and the character flies across the route to collect the coins associated with the chosen option (Right). (C) We required participants to provide valid data at ≥70% of the study notifications. From those included (N = 129), most participants (y-axis) provided valid data at most assessments (x-axis). Timepoints total = independent aggregate state and task measures available after exclusion criteria. Created with BioRender.com (31).

Fig. 2.

Naturalistic, day-to-day fluctuations in state. (A) Spearman rank correlations for mean states (over 15 d; all P < 0.001). People who reported greater motivation also reported greater happiness, better quality sleep, and lower fatigue. (B) Trait apathy (x-axis) was a strong negative predictor of 15-d mean (absolute) state motivation (y-axis). (C) Trait apathy (x-axis) was not related to the SD in state motivation (y-axis). (D) Person mean (normalized within-person) state motivation (gray) and group mean ± SE of mean (black) across weekdays (x-axis). Each gray line is an individual, which gives an idea of the substantial variability. The x-axis follows the study order, which began on a Thursday and y-axis is truncated to −1 and +1 SD from the mean to better visualise the group trajectory. (E) Mean (normalized within-person) state motivation for the first part of the week (Mon–Weds) compared with the second part (Thurs–Sunday) significantly differed, with greater motivation later in the week. (F) State motivation autocorrelation (y-axis, ACF = autocorrelation function) across participants indicating the mean correlation between state fluctuations at a given time and previous timepoints up to lag-4 (bar = group mean, errors = group mean ± SEM). (G) Autocorrelation between state motivation earlier in the day (first notification between 09:00 and 13:00, x-axis) compared with state later on the same day (14:00 and 18:00, y-axis) was typically strong and positive (gray slopes are individuals). The negative moderation effect of trait apathy on the autocorrelation of state fluctuations is overlaid as fitted mean ± SEM at each level state, for trait apathy at −1 and +1 SD from the mean. People with greater apathy (dark blue) experienced a weaker, positive relationship between states within a day meaning that states early in the day were not as strongly sustained to the afternoon. ***paired t test P < 0.001.

Fig. 3.

Reliability and validity of the smartphone assessment of effort-based decision-making. In this plot, filled circles are days and empty circles are participants. (A) Probability of choosing the harder option (y-axis) depended on the difference in the reward value (coloured errorbar = group mean ± SEM, points = game (day) group mean ± SEM), the difference in the effort-values (x-axis) and their interaction. This pattern is consistent with linear reward–effort discounting. (B) The mean probability of successfully receiving coins (y-axis) for each game was consistently near to 1. (C) Maximum effort was standardized to each person through a calibration procedure at baseline (day 0; x-axis, normalized %). The initial maximum effort (x-axis) calibrated at the start of the first game was a strong predictor of the mean 100% effort across the remainder of the study (days 2 to 14, y-axis, normalized %). The ordinary least squared regression (x ~ y) ± SE (blue) overlaps with an exact fit (dashed-line, intercept = 0, slope = 1) indicating excellent calibration of effort-level across the study. For full details of the calibration procedure, see Methods and SI Appendix. (D) Test–retest reliability (ICC) of mean probability of choosing the harder option for each of 28 possible 2-game combinations. (E–G) The reliability of P{harder option} between example game combinations: Baseline (Day 0), Day 2 and Day 14.

Fig. 4.

Fixed effects for predictors of p(harder option) on each trial in the study. (Gray area = state-reward and state-effort interaction effects estimated independently) *P < 0.05, **P < 0.01 ***P < 0.001. Δ = difference on a given trial.

Fig. 5.

State motivation modulated effort-based decision-making by increasing reward sensitivity. Each point represents one game. (A) Main effect of trait apathy (x-axis) on choices [y-axis = conditional fitted mean (% effort chosen–unchosen)]. (B) Main effect of state motivation relative to 15-d person mean (x-axis) on choices (y-axis = conditional fitted mean % effort chosen—unchosen for each game). (C) The relationship between state fatigue (x-axis) and choices was not significant after controlling for state motivation (also the case for state happiness and sleep quality, SI Appendix). (D) State–trait interaction effect on behavior (slope increases at greater levels of apathy). In (A–D), the fixed effect (marginal predicted value) ± residuals are plotted as the smoothed regression line. (E) The latent motivational state (θ, x-axis) was positively associated with the latent decision-making parameter, reward sensitivity on a given day (y-axis). Errors = posterior mean ± SD. Purple solid line = posterior mean coefficient for the effect of current state motivation (β_stateR = 0.19, Fig. 4F, Middle row), with intercept set arbitrarily for visualization as mean (reward sensitivity). (F) Posterior mean coefficients for the effect of state motivation on reward sensitivity at different points in time. The past and current states were associated with greater reward sensitivity on a given day (Top, Middle row), but the future state (Bottom row) was not. Note that the current state effect was modeled first, independently. Post hoc, the past and future states were compared directly within a subsequent model and both assigned the empirical prior mean and SD of the current state effect. This explains the increased certainty (narrower distribution) of the posterior for the effects of past and future states. (A–C) *P < 0.05, **P < 0.01; (E and F) **P(Direction) ≥ 0.97, ***P(Direction) ≥ 0.99.