Contribution
Opinion dynamics had modelled one issue at a time on a fixed influence network. This paper models a standing group working through a sequence of issues, with each member’s openness to influence set by how much they influenced the last outcome, and reports the first experiment long enough (15 issues, triple the longest before) to look for the predicted evolution. It finds the individual-level mechanism and several group-level trends consistent with it.
Key insight
Influence feeds back on closure. “Success in influencing group issue outcomes elevates the individual’s self-weight (encourages opinion assertiveness, confidence, and intransigence)”, and failure dampens it (p. 445). Because a self-weight scales every flow of influence through that person, a small individual adjustment, iterated over issues in a strongly connected group, concentrates influence on one member: “Inequalities of influence centrality are reinforced” (p. 470).
Assumptions
- Single-issue dynamics are Friedkin–Johnsen (Eqs 1–2, p. 450). x(k+1) = AWx(k) + (I − A)x(0), with a_ii = 1 − w_ii. Members stay anchored to their initial opinions, so exact consensus needs some self-weights at 0 or 1. This is not the DeGroot model.
- Influence centrality (Eq. 8, p. 451). C_i(s) is the mean, over members, of the total relative contribution of i’s initial opinion to their settled opinions, from V = (I − AW)⁻¹(I − A).
- Reflection (Eq. 9, p. 452). w_ii(s+1) = C_i(s), or (Eq. 11) the time-average T_i(s) of past centralities. Interpersonal weights are w_ij(s+1) = a_ii(s+1) r_ij(s+1), with R free to change from issue to issue.
- Veridical self-appraisal. “Such adjustments assume that each group member has a veridical appraisal of their own relative influence centrality on each issue of the sequence” (p. 454). This is assumed, not measured.
- Scope. “The key scope condition of the postulated reflection mechanism is a consensus seeking group in which failures to reach consensus are viewed as social problems” (p. 470).
- Measurement (pp. 454–456; S7–S8). Undergraduates at a West Coast university, randomly assigned to 30 groups of four, each discussing 15 of 18 Stoner–Wallach choice dilemmas in random order, under high or low pressure to agree. Initial and final opinions are private 0–100 ratings. W is each member’s post-discussion allocation of 100 chips between the extent to which the discussion did and did not influence their final opinion, the first pile split among the others. So w_ii, and every centrality derived from W, comes from the same self-report.
Key results
- Simulation (pp. 446–449, 453; S1–S3). With a fixed strongly connected R, one self-weight goes to 1 and the rest to 0. With R redrawn at random on every issue, in configuration and values, all 14,000 simulated systems (n = 4 to 50, 1,000 each) still converged to a single closed member, though disturbances can change who. The attractor is a DeGroot system that reaches consensus on the closed member’s initial opinion (S2).
- Specification check (Table 1, p. 455). The F-J prediction of final opinions, with no fitting, correlated 0.872 with observed final opinions and 0.811 with observed changes.
- Reflection (Table 2, p. 457). Prior centrality C_i(s) predicted next-issue closure w_ii(s+1); the effect of C_i(s) on w_ii(s+1) “is constant along the issue sequence”, while the effect of time-averaged centrality rose as 0.404 + 0.095s. Prior centrality also predicted next centrality (Table 3), the T_i effect rising as 0.176 + 0.049s. Issue closure and centrality correlated at 0.769.
- Concentration (Figs 4–7, pp. 458–463). Time-averaged centralities stabilized at unequal levels; the chance of complete closure on the next issue rose with prior time-averaged centrality; complete closure became more frequent along the sequence.
- Structure (Tables 4–6, Figs 8–11, pp. 463–467). Discussion time fell along the sequence (about 0.9 minutes per issue, with a late rise). Network density fell from 84% to 67% by issue 11, then rose to 70%: “the sustained engagement along an issue sequence appears to have the opposite effect on interpersonal influence allocations” to the usual expectation that interaction strengthens ties (p. 465). Density of about 67% is the threshold below which unilateral topologies become common in four-person groups.
- Consensus (pp. 467–468, Figs 12–13). Of 450 discussions, 36.9% ended in unreconciled disagreement, 31.8% in consensus on a member’s initial opinion and 31.3% in compromise consensus. “In our data, the probability of reaching consensus does not alter along the issue sequence” (p. 467). The trace of AW fell toward 0.5, and a lower trace went with consensus on an initial opinion.
- The authors’ summary (p. 471). “Our group-level findings are signals, in the midst of structural disturbances”, toward an attractor that “expeditiously reaches consensus”; without disturbances “the trajectory is toward a system with one dominant individual (a directive leader) whose initial opinions are accommodated”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Under the reflection mechanism in an F-J system, influence concentrates on one member, even when who-influences-whom is redrawn at random each issue | strong within the model: simulation, 14,000 runs | S1 |
| C2 | A member’s self-reported closure on the next issue rises with their prior influence centrality | moderate: 1,680 observations in 30 groups, but closure and centrality come from one self-report | Table 2 |
| C3 | The effect of cumulative centrality grows along the sequence | moderate | Tables 2–3 |
| C4 | Group influence networks thin and discussions shorten along the sequence | moderate: small R² (about 0.05) | Tables 4–5, Figs 9–11 |
| C5 | Groups move toward a single closed member and consensus on that member’s opinion | weak: the consensus rate did not change; only “signals” | pp. 467–471 |
| C6 | Reflected appraisal is the mechanism | weak: no alternative explanation of the persistence is tested | the design |
Concepts
- Influence centrality: “the total (direct and indirect) relative influence of each individual’s initial opinion to the settled opinions” of the group on an issue (p. 444).
- Self-weight, closure-openness: w_ii, the weight a member keeps on their initial opinion; a_ii = 1 − w_ii is openness.
- Reflection (reflected appraisal) mechanism: closure on the next issue set by realized influence on prior issues, after Cooley.
- Attractor system: the DeGroot system with one closed member and the rest fully open, which yields consensus on the closed member’s view.
- Reserved authority (S10): a nominal leader who voices no opinion and enforces “a full display and vetting of group members’ opinions”.
Connections
It is the empirical companion of Jia et al. 2015 (LIT-1146, read today in NOTE-1135), which analyses the same mechanism with DeGroot dynamics on each issue and a fixed relative interaction matrix, and of Friedkin’s earlier formal theory (LIT-1404, unread). The record’s reading of Cooley (NOTE-1080) is the source of the reflected-appraisal idea. The supplement connects it to Michels’ iron law of oligarchy and to Janis on groupthink.
Bearing on the record
- The exchange’s description. It cited this paper as “the laboratory
support for DeGroot–Friedkin social-power dynamics”. Two corrections.
- The experiment’s model is Friedkin–Johnsen, not DeGroot. On each issue members stay anchored to their initial opinions, and the predicted attractor is one fully closed member, an autocracy. The DeGroot–Friedkin model of LIT-1146 predicts instead an interior equilibrium with power ranked by centrality, and autocracy only for a star network. LIT-1146 notes that the F-J simulations show “a more pronounced tendency to autocracy”.
- What the laboratory supports is the individual step. Closure tracked prior centrality, and network structure shifted. The group-level endpoint, a directive leader and consensus on that member’s opinion, was not reached; consensus did not become more frequent.
- CLAIM-448 (coherence is not integration). The claim cites
this paper with LIT-1146 as the baseline for self-reinforcing authority.
The reading supports that, and bears on the claim in three further ways.
- Supports: efficiency is not quality. The groups got faster and their networks thinner as influence concentrated. The authors say their account does not address “whether the evolution of a group toward greater efficiency is associated with a decrease or increase of collective intelligence” (S11), and connect the trajectory to groupthink. That is the claim’s point, made by the authors about their own data: faster settlement is not evidence of better integration.
- Supports: what thins is who is listened to. Density fell as members stopped allocating influence to some others. This is a loss of uptake rather than of admission: the dissenters spoke and were discounted. It is nearer the claim’s “access, recognition, credibility, uptake” ladder (CLAIM-403) than to a regime that keeps contests off the record, and it shows a coherence that comes from concentration without any chair deciding what is admitted.
- Complicates the protected-channel prediction. The claim predicts that a protected challenge channel can reveal hidden conflict. The supplement argues, untested, that a “reserved authority” who enforces full display and vetting of every member’s opinion still yields concentration: “the reticence of the nominal leader of group encourages a competition (explicit or implicit) for influence among the other group members” (S10). If so, admitting every view does not by itself stop the record from becoming coherent by concentration; the reflection step acts after admission.
- Filed as THEORY-902.
- CLAIM-451 (cycles, tension and dissonance). Indirectly, through the claim’s “entrench” route: one perspective learns to dominate and choices become consistent. This paper is a mechanism for that route among people. It says nothing about cycles or dissonance.
- CLAIM-277 and CLAIM-364 (durable authority; low supervisory cost). A social instance of authority made durable by its own success, and of rising efficiency that says nothing about domination.
- No instruction for machine-learning practice.
Limitations
- One self-report carries both sides of the key regression. Closure w_ii(s+1) and prior centrality C_i(s), which depends on w_ii(s), both come from the chip allocations. Closure is persistent across issues, and Table 2 does not control for w_ii(s). Some of the “reflection” effect could be stable individual differences in self-weight. This is the record’s reading; the paper does not discuss it.
- Veridical appraisal is assumed, and self-reported influence is the only measure of influence.
- Small, homogeneous sample. 120 undergraduates, groups of four, one session, hypothetical choice dilemmas with no correct answer.
- Weak group-level fits. R² of about 0.04–0.07 for the trends over issues; a rise near the end of the sequence that the authors attribute to fatigue.
- A table hard to read as printed. In Table 2, the s × C_i(s) coefficient of 0.171 with standard error 0.012 carries no significance mark, while the text says the C_i effect is constant along the sequence. The column layout may be at fault; it does not change the authors’ reading.
- Consensus did not change, so the predicted end state is extrapolated.
Open questions
- Does closure follow centrality once each member’s own prior closure is controlled for, or with influence measured by observers rather than self-report?
- Do groups whose members cannot see their own influence show the same trend? Manipulating feedback about influence would test reflection directly.
- Does a reserved authority who enforces full vetting slow concentration, as the claim would hope, or not, as the supplement suggests?