Contribution
The first full-componential test of two assumptions of appraisal theory: that appraisal changes motivation, physiology and expression and not only feeling, and that appraisal criteria interact. It uses the GRID data on the meaning of emotion words, on the “lexical sedimentation” assumption that a word’s meaning encodes knowledge of the emotion’s typical mechanisms (§1.3).
Key insight
The interactions appraisal theory predicts are real but shallow. One criterion (goal compatibility) does most of the work and splits positive from negative. Two others (agency, norm compatibility) matter only as modifiers of it, and the rest add. That is neither the additive model of regression studies nor the deep decision tree of appraisal theories.
Assumptions
- That culturally shared semantic profiles reflect the structure of real episodes, supported by the similarity of the result to Meuleman and Scherer’s (2013) recalled episodes. The authors say “we do not claim to have examined actual, day-to-day emotion experiences” (ms p. 37).
- That PCA factors, fixed before cross-validation, are acceptable inputs and outputs (they concede a small bias, ms pp. 37–38).
- That a reduction in generalized cross-validation error is a better criterion of relevance than significance with N = 816.
Key results
- Structure (Table 1). Goal compatibility appears at main, second and third order and needs four hinges. Suddenness, familiarity, coping and relevance appear as main effects, and suddenness and relevance also in interactions. Agency and norm compatibility appear only in interactions.
- Fit (Figure 5). Cross-validated R² averages 42.0% for motivation, 41.6% for physiology and 48.7% for expression. Bipolar factors gain from splines. Attack vs submit, tears and speech perturbation gain from interactions.
- The three-way surface (Figure 7). When the self is the agent there is no attack tendency at any combination of goal and norm compatibility, only submission graded by goal incompatibility. When another is the agent, attack appears for high goal incompatibility with high norm incompatibility.
- Matching (Figure 9). Simulated patterns matched by nearest neighbour to GRID words form a structured map: goal compatibility separates positive from negative words, and agency and norms differentiate within each region. The average majority-label probability is 41.8%, against a chance level of 4.2%.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Appraisal affects motivation, physiology and expression through interactions and nonlinear effects | moderate | MARS selection on semantic data; 7 of 11 factors improve |
| C2 | Goal compatibility is the main effect; agency and norm compatibility are pure modifiers | moderate | Table 1; agrees with Meuleman & Scherer (2013) on recalled episodes |
| C3 | The strong hierarchy of appraisal decision trees is not supported | moderate | no single main effect with pure interactions below it |
| C4 | Some appraisals must reach a threshold before affecting responses | weak | a hinge in one fitted curve (Figure 6), on semantic data |
| C5 | Semantic profiles reflect the mechanisms of real episodes | weak | similarity to one earlier study |
Concepts
- pure modifier: an appraisal criterion whose effect exists only in interaction with another, with no main effect of its own.
- hinge: the knot of a MARS spline, where a piecewise-linear function changes slope.
- semantic profile: the typical componential pattern a language community attaches to an emotion word (GRID).
Connections
Builds on Meuleman and Scherer (2013) and on Fontaine, Scherer and Soriano (2013, the GRID sourcebook). The two later Geneva-related studies filed with it use the CoreGRID items to collect ratings of episodes rather than words (LIT-1440, LIT-1443). Its synchronization paragraph restates Scherer’s (LIT-1451).
Bearing on the record
A per-episode measure of synchronization? None, and the data could not give one. Each case is a sample’s average rating of what a word implies. There is no episode and no time. The paper states the hypothesis, “If there are indeed strong causal relations among components, they should show a high degree of synchronization” (ms p. 7). The only coherence it reports is the correlation of factors across the 816 word-by-sample cases (Figure 4). That is coherence in shared knowledge across emotion words, not coupling within an episode. For QUESTION-268 it is not a candidate.
Signed, or which tendency prevails? Its response factors are bipolar and signed: appetitive vs defensive, intervention vs disengagement, attack vs submit. A fitted profile therefore does say which pole of each tendency a word carries. But that is a prediction of meaning, not the outcome of a conflict. A caution on words: its “ATTACK VS SUBMIT factor corresponds to tendencies to dominate (positive values) versus to be dominated (negative values)” (ms p. 22). That is the social dominance of the PAD tradition, the homonym the record renamed its own notion to avoid, and it should not be read as potential.
The threshold effects and CLAIM-435. They do not bear on the claim’s contrast, for three reasons.
- MARS cannot produce a discontinuity. Its basis functions are hinges, zero on one side of a knot and linear on the other, so every fitted curve is continuous. A “threshold” in this model is a kink, a change from a flat slope to a steep one. On the claim’s terms that is a continuous currency whose price rises sharply, Veit’s regime (THEORY-101), not ALARM’s override (THEORY-090). Even if the true relation had a jump, MARS would fit it as a steep segment.
- The “threshold” is the authors’ gloss on one panel, and their own description of that panel is a steepening, not an onset: “increasing familiarity simply leads to decreasing physiological arousal. This decrease occurs even faster for strong appraisal of familiarity” (ms p. 27). The discussion then cites that panel for “an appraisal criterion must reach a certain threshold value” (ms p. 34).
- Nothing is traded off. The inputs are appraisals and the outputs responses. There is no cost, no resource and no choice between pairs, so the claim’s indifference curve does not exist here.
The paper does sharpen one of the claim’s cautions. Following Kappas, it treats a sigmoid and a step as near-equivalent: “Sigmoidal curves can also be viewed as continuous approximations of binary threshold functions” (ms p. 10), and “they share the property that output cannot increase substantially until input reaches a certain value” (ms p. 10). CLAIM-435 says that a finite set of tested pairs can always be fitted by a currency, so the test needs enough pairs to tell a steep curve from a discontinuity (after CLAIM-436). This paper is an example of a field that has not drawn that line. Here “threshold” covers both, and a smooth-spline method decided the matter by construction. A trade-off study meant to test the claim should not fit splines of this kind. It should fit a model with a free jump and compare the two.
It cannot speak to hysteresis either (“the omission of time”, ms p. 38).
One structural point. The result that the mapping is “partly additive and partly hierarchical” (ms p. 34) is a case, in emotion knowledge, of compensatory and gated structure in one function. Additive criteria trade off. A pure modifier changes the effect of goal compatibility without having any effect of its own, which is how a gate acts. CLAIM-435 expects both regimes in one animal, as one ordering changing regime. This is at most a weak analogue: it concerns appraisal-to-response maps in word meaning, not an ordering’s potential in conduct. It is worth noting because it is the same two-regime shape found in a different domain.
Limitations
The data are semantic. The authors argue this is a strength, since semantic profiles are stable across languages (alpha .92). But it limits any inference about episodes. PCA fixes the factor structure before modelling, and there is no feeling component (dropped by design), no time and no feedback. The matching map cannot be scored against ground truth.
Open questions
Whether the pure-modifier structure survives in episode data with a model that allows discontinuities (a tree with free jumps, or a change-point regression), so that a kink and a jump can be told apart.