KurisutinaJournal of Personality and Social Psychology 99(6), 1042–1060 (2010); received 17 December 2009, accepted 24 May 2010.
DOI 10.1037/a0020962. Peer reviewed. University of Leuven (and Melbourne). Read as the online-first version (20
September 2010, paginated 1–19), posted on co-author Zita Oravecz's university course site; not compared with the
final paginated version. Provenance: papers/base/kuppens2010_dynaffect.provenance.json.
Why this paper. It replaced Eldar et al. 2016 ("Mood as representation of momentum") in the candidate list. B4 needs a model fitted to people's daily-life affect with per-person set-points and recovery; this is one. Eldar 2016 was not read; judged from its abstract, it is a theory and review piece on the reward-learning account of mood that Rutledge 2014 already supplies in fitted form.
What was read
Every line of the pdftotext -layout text of the 19-page PDF (1,171 lines): the theory, the model, both studies
(methods, model comparison, posterior predictive checks, Tables 1–6), general discussion, psychopathology, limits,
footnotes and references. Not read: Figures 1–7 as images (captions read); the cited papers that define the
estimation (Oravecz and Tuerlinckx 2008; Oravecz et al. 2009, in press).
Question
Can three person-level processes, an affective home base, variability, and attractor strength, account for how people's feelings change over days and over minutes, and do they relate to personality and emotion regulation?
Model
A hierarchical Ornstein–Uhlenbeck (diffusion) model, fitted separately to valence and arousal (two dimensions):
- measurement: Y_p(t) = θ_p(t) + ε_p(t);
- dynamics: dθ_p(t)/dt = β_p (μ_p − θ_p(t)) + ξ_p(t).
Per-person parameters: home base μ_p (the set-point, per dimension), attractor strength β_p (how fast affect is pulled back), the noise ξ_p that with β_p sets variability, and measurement error. Parameters are drawn from population distributions (random effects). Continuous time, so uneven intervals and missing beeps are handled. Bayesian estimation by MCMC in custom MATLAB code.
Data
- Study 1: 79 Leuven students (50 female, mean age 24), 10 random beeps a day for 14 days (stratified within waking hours; about every 84 min), 82% of beeps answered. Core affect on a 99 × 99 Affect Grid (valence × arousal); momentary self-esteem; trait questionnaires (NEO-FFI neuroticism and extraversion, PANAS, Rosenberg self-esteem, Satisfaction With Life, emotion regulation, rumination).
- Study 2: 60 students (40 female, mean age 23), 50 beeps a day for 4 days (about every 17 min), 87% answered; the same questionnaires.
Results
-
All three person-level processes are needed (deviance information criterion, lower is better):
Model Study 1 DIC Study 2 DIC Full DynAffect (all parameters vary by person) −3,604 −6,697 Attractor strength the same for everyone 1,966 −5,128 Variability the same for everyone 5,151 −4,378 Both the same 15,872 10,379 No attractor (random walk) 20,408 7,906 Infinite attractor (no time dependence) 40,082 42,258 -
Simulated people look like the real ones. Correlation between how often each person visits nine regions of the affect space and how often their simulated twin does (1,000 replications): mean 0.80 (SD 0.20) in Study 1 and 0.84 (SD 0.14) in Study 2; with a finer 9 × 9 grid, 0.47 and 0.51. Pooled vector fields: little movement near the home base, faster return the further away, as replicated by the model.
-
Forecasting the next report. The held-out last observation fell inside the 95% prediction interval for 80% (valence) and 91% (arousal) of people in Study 1, and 82% and 95% in Study 2: some undercoverage on valence.
-
Population values (rescaled to 0.1–9.9; the printed SDs are posterior uncertainty, not between-person spread, which is not reported):
Valence home base Arousal home base Valence variability Arousal variability Valence attractor Arousal attractor Study 1 5.95 4.30 3.21 4.28 0.026 0.025 Study 2 5.82 4.23 2.67 4.29 0.043 0.049 The home base is slightly pleasant and of medium arousal. Attractor strength is larger at the shorter time scale (95% credibility intervals barely overlap), so the estimated pull depends on sampling interval. The time unit of β is not stated in the text read.
-
Trait correlates (Study 1, N = 79; Study 2 in brackets):
- valence home base: neuroticism −.386 (−.413), positive affect .446 (.303), self-esteem .356 (.257), life satisfaction .247 (.488);
- valence variability: neuroticism .278 (.261), negative affect .283 (.309), self-esteem variability .393;
- arousal attractor strength: reappraisal .310 (.268), rumination −.226 (Study 2: n.s.). The patterns of correlations agree across studies (r = .66 over 54 coefficients; intercorrelations r = .68).
-
Authors' reading. The home base acts as the desired state; reappraisal acts on arousal more than valence; valence instability marks poor adjustment, arousal variability does not. Disorders could be described as extreme home bases, variability or attractor strength (depression: negative, low-arousal home base with weak attractor; bipolar disorder: two attractors).
Limits
- Students; 4 or 14 days; no retest of the person-level parameters across occasions.
- No time-varying inputs: events and appraisals enter only as noise. The authors name event covariates, asymmetric return (from negative and from positive states) and multiple or switching attractors as next steps.
- One fixed attractor per person; recovery is exponential by construction.
- Between-person spread of the parameters is not printed, so the model cannot be used as a population prior from this paper alone.
What it means for Kurisutina
- A fitted form for B4's slow, daily-life channel. Per person: a set-point (home base), a recovery strength and a noise level, on valence and arousal; estimable from roughly 100–200 experience-sampling reports; the population distribution of the three is what B8 would hold. Model comparison says none of the three can be shared by everyone.
- It needs inputs to be a mechanism. As fitted, the model only describes fluctuations; events enter as noise. B4 would add event-driven inputs, for example Rutledge-style expectation and surprise terms, to the drift.
- One time constant is not enough. The pull measured over minutes was stronger than the pull measured over hours. Together with log-time recovery after life events in the panel studies, this suggests B4 needs several time scales (minutes, days, years), each with its own set-point dynamics [inference].
- Set-points can move. The authors note the home base itself can shift after major life events; the slow dynamics of μ are where B4 meets B6.
Cross-references
summaries/base/rutledge2014_momentary_wellbeing.md: event-driven momentary happiness at the minutes scale.summaries/carry_on/beck2018_idiographic_networks.md,fisher2018_group_to_individual.md: within-person affect structure from experience sampling.summaries/carry_on/haehner2024_event_characteristics_swb.md,luhmann2012_adaptation.md: change and recovery after life events.