Ben Meuleman, Agnes Moors, Johnny Fontaine, Olivier Renaud & Klaus R. Scherer (2019), Emotion 19(3):425–442 — DOI-10.1037/emo0000449
Key takeaways
- The target (abstract, ms p. 3). Appraisal research has mostly fitted linear effects of appraisal on feelings, but “a linear model is not compatible with the idea held by many appraisal theorists that appraisal criteria interact to produce emotional responding.”
- The data (§2.1). In the GRID study 4,948 respondents “rated the typicality of 144 items for 24 common emotion words” (ms p. 16): how likely, when a speaker uses the word, is each appraisal, action tendency, bodily change or expression. Averaged by sample and word, this gives 816 cases. These are semantic profiles, not episodes.
- The model (§2.2). PCA gives 7 orthogonal appraisal factors and 11 oblique response factors. MARS fits them jointly with piecewise-linear hinges and automatically selected interactions.
- Results (§3.2, §4). Third-order interactions were selected. Goal compatibility is the main effect throughout. Agency and norm compatibility “appeared exclusively in interactions, without having a corresponding main effect” (ms p. 34). Explained variance (cross-validated) runs from 89.0% (appetitive vs defensive) to 7.3% (intervention vs disengagement). Arousal, temperature, jaw drop and vocal energy are fitted as well by a linear additive model.
- Hierarchy only in part (ms p. 34). “Nevertheless, the strong hierarchical interactivity proposed by Roseman (2001) and Scherer (2001) could not be substantiated.” Instead “models were found to be partly additive and partly hierarchical.”
- Thresholds (ms p. 34). “In some cases the spline effects suggested that an appraisal criterion must reach a certain threshold value before it can affect emotional responding”, and in others a slope reverses at the hinge.
- No time (ms p. 38). “A second limitation of the current study is the omission of time from our emotion model.”
Standing in the record
Filed on 2026-10-23 and read against two of the record’s questions, as set out in NOTE-1104. It gives no measure of component synchronization; “synchronization” appears only in its review of Scherer. Its threshold effects do not bear on CLAIM-435’s continuous-versus-threshold contrast. They are kinks in a continuous fitted function, MARS cannot represent a jump, and the inputs are appraisals of word meanings, not costs and resources traded in choice. What it does show is that in emotion knowledge an additive structure and a gated one (pure modifiers) coexist in one mapping.
It concerns emotion psychology and uses machine-learning regression as a tool. It is not an instruction for machine-learning practice.