Contribution
A theoretical and methodological proposal. The authors take one property from each of three broad classes of emotion theory and ask which psychometric model allows all three. They argue that the two standard latent-variable models do not and that a psychometric network model does. On that model an emotion is the system of causal interactions among its components, not a latent cause or index of them.
Key insight
The question “what is an emotion?” is partly a question about the measurement model, since “any psychometric model depends on a (possibly minimal) theory of the construct” (pp. 444–445). Choose a network model and an emotion becomes a stable configuration of mutually reinforcing components. That gives a mechanism for why some emotions last: “the components continuously reactivate each other, sustaining the activity of the network” (p. 460).
Assumptions
- That emotions have components (feelings, cognitions, action tendencies, expressions, physiology), as all three classes of theory agree (p. 445).
- That the edges of a pairwise Markov random field (Ising or Gaussian graphical model) are “indicative of causal structures” (p. 458), though not causal estimates.
- Weak emergence in Bedau’s sense: “The traffic jam does not develop causal powers independent of the cars” (p. 457).
- That valence alone is not the construct. “measuring only the valence of an affective state implies that there are no distinct emotions, or at least that these are not considered relevant” (p. 445).
Key results
Argument, not data:
- Property audit (Table 1, pp. 448–458). The reflective model fails on distinct emotions (no common-cause neural signatures, and no separately identifiable emotion) and on causal relations (local independence is contradicted). The formative model allows variation but makes component relations uninformative. The network model allows all three.
- Distinct emotions as communities (pp. 456–457). Components cluster (feelings with cognitions, motivations with expressions). Appraisals link clusters, and the whole forms a small-world community that is the emotion.
- Emergence (p. 457). A stable system emerges whatever component was activated first, and then “force[s] each emotion component into a specific state”. This is weak downward causation.
- Duration and hysteresis (p. 460). In a strongly connected network activation is self-sustaining and hard to reverse. In weakly connected ones hysteresis does not occur. “situations appraised as very relevant could strengthen the causal connections between emotion components, explaining why relevant situations increase emotion duration.”
- Different networks, different signs (p. 461). An edge between anger and aggression may be positive in cross-sectional and contemporaneous networks but negative in the temporal one: “it may be the case that when the person feels angry earlier, this will predict less aggression at a later time point.”
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Reflective latent-variable models cannot represent emotions consistently with the evidence | moderate | review of neural-signature and local-independence evidence |
| C2 | A psychometric network model accounts for distinct emotions, variation, and causal relations among components | moderate (as compatibility) | conceptual argument; no emotion network estimated here |
| C3 | Distinct emotions are weakly emergent small-world communities | weak | analogy to attitude and psychopathology networks |
| C4 | Network connectivity predicts emotion duration through hysteresis | weak (a prediction) | simulation work in psychopathology (Cramer et al. 2016), not tested on emotions |
Concepts
- psychometric network model: nodes are indicators, edges are conditional dependencies; the system of their interactions is the construct.
- weak emergence: a macrostate autonomous of which microstates produce it, with downward influence fixed entirely by local interactions.
- hysteresis: in a strongly connected network, the difficulty of returning to the deactivated state once it is activated.
- cross-sectional, contemporaneous and temporal networks: three estimands, between persons, within a person at one time, and within a person across lags, that need not agree.
Connections
The emotion-domain application of Borsboom’s network approach, whose theory of mental disorders the record holds (LIT-547, NOTE-435, THEORY-058). It cites Lewis (2005, LIT-1455) and Scherer (2009) as dynamic-systems precursors that stayed theoretical or neural. It cites Kuppens, Oravecz and Tuerlinckx (2010, LIT-1469) for continuously varying core affect, and Moors for the typology of theories. Dalege et al.’s attitude networks are the closest worked model.
Bearing on the record
On potential. Lange et al. supply one of the three marks with a mechanism: persistence, as hysteresis in a strongly connected network (p. 460). The weakly emergent state that “force[s] each emotion component into a specific state” (p. 457) is Lewis’s “enslave”, made statistical. Neither is the record’s potential, for three reasons.
- The nodes are not perspectives, and nothing in the network ranks anything. A component (frowning, arousal, aggressive thoughts) has a level, not an ordering of the options the system faces. What becomes self-sustaining is a configuration of components. In CLAIM-440, what prevails is one perspective’s ranking of options in the conflicts where perspectives disagree. The network model has no “conflict” in that sense. It has edges, positive or negative, but no pairs of options on which two rankers disagree.
- No sign. The authors set valence aside on purpose. A network-state reading of valence would have to add what the proposal adds, an ordering of the system’s own states and a reference course along it.
- No priority mark. Persistence and resumption have network analogues: hysteresis, and return to the activated attractor after a perturbation. Priority, winning conflicts against other rankings, has none, because the network has no rival rankings.
On measurement (QUESTION-268). The paper’s own limits are the record’s. “the psychometric network model primarily provides information about relationships between variables (i.e., conditional dependencies) and not about to what extent variables are endorsed” (p. 460). A measure of potential needs exactly what this omits: which ranking’s verdicts were followed, episode by episode. The paper does point to a usable tool. A temporal personalized network (p. 461) over nodes that are perspectives’ verdicts, not emotion components, would show whether one perspective’s verdict at t predicts conduct at t+1 against the others. That is out-prediction (CLAIM-427’s grade) estimated as a lagged network. Connectivity predicting duration (p. 462) is a candidate measure of persistence strength.
On Borsboom (THEORY-058). The paper extends the record’s network reading from disorders to emotions with no change of machinery. A mood disorder and an emotion become the same kind of thing, a self-sustaining network state, differing in connectivity and duration. If the proposal is right, a disorder of valence would be a fault in which ordering prevails, or in its reference course, and not only an over-connected network. The two readings make different predictions. A strongly connected depressive network with no stable potential ranking would be ambivalent on the proposal, not negative.
On synchronization. The network model is not a synchronization model. Its edges are conditional dependencies, estimated cross-sectionally or at lags, not phase relations. It therefore complements Scherer’s synchronization of components (LIT-1451) and Lewis’s phase synchrony (LIT-1455) rather than measuring the same thing. All three remain, as QUESTION-268 says of the first two, correlates that potential would produce. They are not measures of which ranking prevails.
Candidly. This is a weaker rival to the proposal’s novelty than Moors (LIT-1434), because it is about components, not about selection. Its main value is a ready statistical vocabulary for persistence and lagged prediction. The proposal can borrow that vocabulary only after it says what the nodes would be. That is the individuation of perspectives, which CLAIM-440 sets aside.
Limitations
No emotion network is estimated, so the case is for compatibility, not fit. The duration prediction rests on simulations from psychopathology (Cramer et al. 2016). The authors grant that edges are not causal estimates and that estimation is exploratory (pp. 458, 461). The claim that the network model integrates affect-program theory concedes the property but not the theory: it is “inconsistent with the perspective that distinct brain modules activate all emotion components simultaneously” (p. 460).
Open questions
Which nodes an emotion network should have, since “Behind certain items, other network structures can hide” (p. 459). Whether hysteresis in emotion networks can be shown in within-person data. How a network account would represent conflict between emotions, as distinct from co-activation.