A computational and neural model of momentary subjective well-being

PNAS 111(33), 12252–12257 (2014); received 30 April 2014, accepted 2 July 2014, edited by Wolfram Schultz (guest editor). DOI 10.1073/pnas.1407535111; PMC4143018; "freely available online through the PNAS open access option". UCL (Wellcome Trust Centre for Neuroimaging, Max Planck UCL Centre, Gatsby). Peer reviewed. Provenance: papers/base/rutledge2014_momentary_wellbeing.provenance.json.

What was read

Every line of the publisher PDF as posted on the first author's site (6 pages, pdftotext -layout, 531 lines): main text, significance statement, figure captions (Figs 1–5), Methods and references. Not read: the Supporting Information (SI Methods, model comparison Tables S1–S2, out-of-sample Table S3, Figs S1–S6). Its PMC and PNAS copies sit behind bot checks; the author-site link to a version "with supplement" returned 404. Figures were not read as images. The SI holds the model comparison and the out-of-sample table, so those are known only through the main text's summary of them.

Question

How do recent events add up to a person's momentary happiness? Is it current earnings, or recent rewards and the expectations and prediction errors that come with them?

Model

Happiness at trial t:

Happiness_t = w₀ + w₁ Σⱼ γ^(t−j) CRⱼ + w₂ Σⱼ γ^(t−j) EVⱼ + w₃ Σⱼ γ^(t−j) RPEⱼ

• CR is a certain reward when chosen; EV the expected value of a chosen gamble; RPE its reward prediction error (outcome minus EV). Unchosen terms are zero.
• Per-person parameters: a constant w₀, three weights w₁–w₃, and a forgetting factor 0 ≤ γ ≤ 1. Fitted to each person's ratings by nonlinear least squares; compared by Bayesian model comparison with alternatives (no exponential constraint, unchosen options, utility-based variants), reported in the unread supplement.

Data and design

• fMRI study: 26 healthy right-handed adults (aged 20–40, 7 male), 150 trials choosing between a certain amount and a 50/50 gamble (mixed, gain and loss trials), "How happy are you at this moment?" after every 2–3 trials (63 ratings per person, 0–100 slider). £20 endowment, paid by performance. Two excluded from fMRI for head movement.
• Two lab replications: "earnings always shown" (n = 22: 11 returning from the fMRI study 14–17 months later plus 11 new) and "only some gamble outcomes shown" (n = 21 returning, median 53 days later, range 3–162).
• Smartphone replication: The Great Brain Experiment app; 18,420 anonymous unpaid adults (8,557 male), 30 trials and 12 ratings per play, over 200,000 ratings.
• Life happiness asked once before each experiment ("Taken all together, how happy are you with your life these days?").

Results

• Earnings do not raise happiness. Participants earned £28.51 ± 7.60, yet happiness went from 60 ± 18 to 54 ± 20 (t(25) = −1.0, P = 0.33); earnings did not predict the change (P = 0.16, r² = 0.079). Successive ratings still moved a lot: root-mean-square change 17 ± 8 points (range 6–43).
• The model fits individuals. CR, EV and RPE weights positive on average (all t(25) > 4.6), EV weights below RPE weights (t(25) = 4.3). γ = 0.61 ± 0.30 (mean ± SD across people). r² = 0.47 ± 0.21 per person (example people 0.79 and 0.41).
• Short memory. At γ = 0.61, a reward five trials ago has 8% of the effect of one just received; at γ = 0.8, 33%; at γ = 0.4, 1%. Rewards more than about 10 trials back have essentially no effect. Fitting the first 140 trials and predicting the final rating from the last 10 left residuals unrelated to earnings (P = 0.81, r² = 0.003).
• Replications. Earnings always shown (n = 22): same pattern (all t(21) > 3.0); earnings now related to the change (P = 0.042, r² = 0.19) but not beyond the model (P = 0.17). Some outcomes hidden (n = 21): median r² 0.43 against 0.39 for a model with raw gamble rewards instead of RPE; expectations raised happiness even when no outcome was shown (t(20) = 2.8, P = 0.011). Splitting RPE into reward and EV at outcome: reward positive (t(20) = 6.6), EV at outcome negative (t(20) = −4.3), their weights anticorrelated across people (r = −0.58), as RPE coding predicts.
• At scale. In all 92 subsets of 200 consecutive app users the three weights were positive (t(199) > 2.0); EV < RPE in all but one subset. A single rating per participant (n = 18,420) gave the same pattern (all P < 0.005). Happiness again did not rise with earnings; among 2,211 users who gained and started below the scale midpoint, 42 ± 18 → 43 ± 19.
• Out of sample. Weights from the fMRI study predicted ratings in the other experiments with median r² > 0.23 (Table S3, not read).
• Brain. Ventral striatal activity at task events predicted later ratings (t(23) = 5.1 left, 3.9 right) and carried the same CR, EV and RPE weights; formal mediation was not significant (P > 0.2). At the moment of rating, right anterior insula activity tracked the reported happiness (t(23) = 4.3), the striatum did not (P = 0.97), and the insula effect did not differ with life happiness (r = −0.20, P = 0.36).
• Authors' reading. Momentary happiness reflects "not how well things are going but instead whether things are going better than expected", varies around a hedonic set point, and in changing environments would show the hedonic treadmill.

Limits

• Gambles for small stakes over about 30 minutes; the time unit is a trial, so γ says nothing about hours, days or life events.
• The constant w₀ is a session baseline, not an estimated long-run set point; no dynamics of w₀ are modelled.
• Some participants did the task twice, months apart, but no test–retest reliability of the parameters is reported in the main text.
• r² ≈ 0.47 is in-sample per person; the out-of-sample check (median r² > 0.23) is in the unread supplement.
• Supplement not read, so the model comparison was not checked.

What it means for Kurisutina

• A fitted, per-person form for B4's fast channel. Momentary affect as a leaky integrator of recent expectations and surprises, with a personal baseline, personal sensitivities and a personal decay. Five parameters per person, identifiable from about 60 ratings in a structured task, and the same form held in 18,420 people.
• What it does not give. Recovery after life events (log time), adaptation of the baseline, and reactivity that depends on load all live at longer time scales; this model has none of them. It covers minutes, not months.
• It ties affect to the learning machinery. The inputs are the same expectations and prediction errors that a learning model computes, so B4's fast channel can read them from B6 instead of being a separate system.
• Earnings are not happiness. A replica whose mood tracked accumulated good fortune would be wrong in a way this study measures directly.

Cross-references

• summaries/base/kuppens2010_dynaffect.md: affect dynamics over days, with per-person home base and regulation.
• summaries/carry_on/luhmann2012_adaptation.md, haehner2024_event_characteristics_swb.md: adaptation after life events at the scale of years.
• summaries/carry_on/mancini2011_hedonic_treadmill.md: set points and their exceptions.