Structural Dependence, Imitation or Herding? A Multi-Agent Simulation of Global Intermedia Agenda Setting

1.1 Following without orders

• Peripheral outlets’ dependence on Western agencies and their similarity to core coverage leave open whether following reflects newsroom constraints or editorial choice.The structural account emphasizes limited resources; choice-based accounts emphasize copying prestigious peers or following whoever moves first.Source: §1.1 · p. 1, §1.2 · p. 1

1.2 Structure or choice

• Because dependence, prestige imitation and herding all predict correlated agendas, observational studies cannot identify which mechanism produces cross-outlet following.A simulation can switch mechanisms off while holding others fixed, although it can establish only that a mechanism is sufficient to produce a pattern.Source: §1.2 · p. 1

1.3 This study

• The study measures 18 outlets in six countries across three issues and three following dimensions, then tests the competing mechanisms in a language-model simulation.The issues span low, mixed and high local stake; the simulation evaluates how well each mechanism reproduces observed following.Source: §1.3 · pp. 1–2
• The study combines world-system, institutional and cascade accounts with salience, framing and timing measures to compare simulated mechanisms against observed coverage.Its contribution is a multi-agent simulation of editorial decision-making across 18 newsrooms.Source: §1.3 · p. 2

2.1 Dependence on the core’s agencies

• Research documents Western agencies’ role in foreign coverage, but structural position has been assumed rather than tested against rival explanations for why outlets follow the core.The study measures dependence using wire credits and correspondent presence.Source: §2.1 · p. 2

2.2 Choice built by structure

• Shared newsroom norms can reflect structures as well as choice: agency dependence may transmit Western formats, while state agencies can shape information flows in peripheral settings.The study does not measure how norms travel and records national-agency credits separately from Western wires.Source: §2.2 · p. 2

2.3 Three rival accounts

• National interest predicts less following on high-stake issues, prestige imitation predicts copying core frames without wire copy, and herding predicts faster following for new, uncertain stories.Prestige imitation follows peers for their standing; herding follows whoever moved first.Source: §2.3 · pp. 2–3

2.4 Three levels of following

• The study measures following as salience, framing and timing; dependence predicts wire-linked following across all three, while the rivals make narrower predictions.Prestige imitation predicts frame-following without wire copy, and herding predicts strongest timing following for new, uncertain stories.Source: §2.4 · p. 3

2.5 Why a simulation

• Prior research measures agency reliance, peer imitation or national interest separately, but none tests dependence against prestige imitation and herding for the same outlets, issues and period.Because the accounts share observational predictions, a correlation cannot distinguish them; a simulation can switch mechanisms off, though it shows sufficiency only.Source: §2.5 · p. 3

2.6 Research questions and hypotheses

• The hypotheses predict greater following in the periphery, stronger following with wire reliance, and a smaller local-stake reduction in the periphery; the simulation compares dependence with its rivals.The study excludes national interest and shared normative isomorphism from the simulation because they cannot be selectively switched off in its design.Source: §2.6 · p. 3

3.1 Type of study

• The study pairs computational content analysis with an agent-based simulation because neither agenda correlations nor simulation alone can separate the proposed mechanisms.Language models serve as both coders and agents; mechanisms are explicit rules, fitted on one period and scored on another.Source: §3.1 · pp. 3–4

3.2 Data and sampling

• The sample covers 18 English-language national outlets in six countries and four core outlets, with AP, AFP and Reuters tracked as wire sources.Outlets were selected from national Media Cloud collections based on story volume and archive coverage; sampled issues were Ukraine, Gaza and COP27/COP28.Source: §3.2 · p. 4

3.3 Measures

• Following is measured in salience, framing and timing, while wire reliance is the share of sampled foreign stories credited to AP, AFP or Reuters.Wire reliance counts credits rather than text reuse, so uncredited agency copy is missed; correspondent presence fell below the reliability floor for confirmatory tests.Source: §3.3 · p. 4

3.4 Simulation design

• Language-model agents receive daily event, wire and core-outlet feeds and propose stories; dependence, prestige imitation and herding rules then alter those proposals.Each rule has one free parameter, and the study tests each alone, in pairs, together and against a no-rule baseline.Source: §3.4 · p. 5

3.5 Validation

• Coder labels were tested for stability across repeated runs, while human spot-checking was absent; simulation rule parameters were fitted on one half and scored on the other.The authors state that stability demonstrates consistency, not validity.Source: §3.5 · p. 5

3.6 Analytic strategy

• Confirmatory hypotheses use prespecified procedures, corrected significance tests and confidence intervals; simulation mechanisms are compared by out-of-sample error and a bootstrap margin.A mechanism wins only if its single-mechanism configuration beats rivals and baseline by the seed margin and that margin holds in at least 95 percent of outlet bootstrap resamples.Source: §3.6 · pp. 5–6

3.7 Reproducibility

• The repository releases the simulation environment, rules, prompts, proposals and scripts, allowing the simulation analysis to be rerun without a language model.Article texts are not released because they belong to the outlets.Source: §3.7 · p. 6

4 Results

• Both tiers followed core attention and timing and sat closer to its frames than the benchmark, but the mechanism comparison remained undecided.The herding test was supported, while the isomorphism test was not.Source: §4 · p. 6

4.1 Validation

• Every confirmatory variable met the stability floor, but simulated framing used coder labels because agent frame labels agreed with the coder at only .232.No human coded any story; the contamination check found no sampled headline recalled at similarity 0.8 or more.Source: §4.1 · p. 6

4.2 Following by tier (RQ1)

• Salience correlations were 0.38 (SD = 0.29) in the periphery and 0.27 (SD = 0.31) in the semi-periphery, above mismatched-issue benchmarks; framing remained below zero.Median first-story lags were −0.07 and −0.03 days, respectively, within two hours of the core’s first mention. AP, AFP, or Reuters credits covered 0.57 of peripheral and 0.40 of semi-peripheral foreign stories. Alternate GDELT title-based shares yielded tier means of 0.28 and 0.30, removing the descriptive gap.Source: Figure 1 · p. 6, §4.2 · pp. 6–7, Table 1 · p. 7

4.4 Rival tests

• The isomorphism test was not supported, whereas the herding test was supported: outlets followed sooner on events covered by fewer core outlets.The herding coefficients were −0.1316 days per additional core outlet in Ukraine and −0.1640 in Gaza; the authors note wire co-delivery and the first-core-mention measure as untested alternative readings.Source: §4.4 · p. 7

4.3 Hypothesis tests (H1–H3)

• Tier did not separate groups on any level; wire reliance was associated with closer timing only, while local-stake interactions were not distinguishable from zero.The wire-reliance timing coefficient was 0.175 (95% CI [0.084, 0.267]); the periphery × stake intervals included zero on salience, framing, and timing. The climate-cell salience falsifier did not hold, but the framing falsifier did.Source: §4.3 · p. 7, Table 2 · p. 8, Figure 2 · p. 8

4.5 Simulation (RQ2)

• RQ2 remained undecided: dependence’s 0.83 error beat the baseline’s 1.00 and the best other single rule by 0.08, but its margin held in only 0.45 of outlet resamples, below 0.95.At w = 1.00, dependence reduced to publishing wire copy: 17 of 18 outlets kept no own story, and its timing error rested on two cells. Imitation alone lowered salience error to 0.85 from the baseline’s 1.13.Source: §4.5 · pp. 7–9, Table 3 · p. 9, Figure 3 · p. 9

5.1 Structural dependence

• A cell with wire credits on every foreign story reached events about four hours sooner than one with none, but same-hour wire delivery could explain the lag without editorial following.Wire reliance showed no supported effect on attention or framing, and the study establishes association rather than how newsrooms decide.Source: §5.1 · p. 9
• Tier was a distant resource proxy, and wide intervals around tier and stake coefficients mean the nulls are absences of evidence, not evidence of equality.The tier difference was 0.03 index points, while wire reliance varied from none to nearly all of an outlet’s foreign stories; the share tracked timing where the country score did not.Source: §5.1 · p. 9, §5.3 · p. 10

5.2 The rival accounts

• Periphery outlets followed the core’s attention on the high-stake climate conferences, but not its framing; the prestige-imitation test was uninformative and herding remained consistent with a cascade.The no-wire frame comparison could not support the predicted contrast, and wire delivery or an order-statistic artefact remained possible readings of the herding result.Source: §5.2 · p. 9

5.3 Limitations

• The evidence has important measurement and sampling constraints: no human coded or spot-checked stories, timing existed for 37 of 52 cells, and the sample included only English-language outlets.Wire credits miss uncredited copy, while wide intervals around tier and stake coefficients leave the nulls inconclusive.Source: §5.3 · p. 10
• As fitted, the simulation’s dependence rule reduced to publishing wire copy, so it did not test a resource that varies across outlets or decide between mechanisms.Agents proposed once from a daily feed, making simulated lags whole days while observed lags ran in hours; their frame labels also agreed too little with the coder to use.Source: §5.3 · p. 10

6 Conclusion

• One of nine confirmatory tests held: timing following rose with reliance on AP, AFP and Reuters copy, while tier and local stake showed no effect distinguishable from none.Periphery outlets followed core attention on the issue with the highest local stake; the herding test was supported in both wars, with two readings left open.Source: §6 · p. 10
• The simulation did not decide between dependence and cascade because its fitted dependence rule failed the bootstrap condition and reduced to publishing wire copy.The study concludes that wire copy sets timing, while whether dependence or cascade sets the agenda remains to be separated.Source: §6 · p. 10
• For world-system theory, the observed dependence appeared through wire copy rather than measured tier position; the study also offers a mechanism-testing approach using language-model editors.Each proposed mechanism can be switched off while the others remain on.Source: §6 · p. 10

Declaration of generative AI use

• A retrieval-augmented system searched a curated source library and linked prior-work claims to verbatim excerpts verified against their sources.The authors describe generative AI chiefly as a retrieval instrument.Source: Declaration of generative AI use · p. 10
• Language models assisted with literature screening, drafting, evidence checking, manuscript review, corpus coding, and simulation agents.The declaration identifies different models for these tasks and says models used to score the corpus are reported in the Method.Source: Declaration of generative AI use · p. 10
• The authors reviewed, edited, and approved all output, confirmed content and reference accuracy, and take responsibility for the entire content.Source: Declaration of generative AI use · p. 10

• Stories qualify for an issue when the title or text matches its query terms, which were tested on a pilot week per outlet.Source: A Design records · p. 13
• Table S2 records the three mechanism rules applied to agents’ proposals.Source: Table S2 · p. 13
• The simulation note defines new stories as appearing within two days of the first core story and uncertain cases as having fewer than three core stories so far.Its baseline is proposals published with no rule; combined rules are applied in dependence, imitation, then herding order, with parameters fitted on the first validation half and scored on the second.Source: A Design records · p. 13

• Three fresh-context passes of the model coder agreed on 248 of 300 sampled frames, with stability measured by nominal Krippendorff’s α.No human coder participated; the foreign-story wording was revised after its first α was .782, and agent framing below the .80 floor was computed from coder frames on written-out proposals.Source: B Validation and checks on the measures · p. 14
• Tables S4 and S5 report outlet-level foreign-story and wire measures and checks on following measures by tier.Source: Table S4 · p. 14, Table S5 · p. 15
• Wire reliance counts AP, AFP, or Reuters credits in bylines, datelines, or end notes, without a text-reuse rule because agency full texts were unavailable.The 49 Washington Post syndicated copies count as wire copy by URL rule; staff foreign byline is exploratory because its model self-consistency was .749.Source: B Validation and checks on the measures · p. 14
• The following-measure checks compare kept outlet–issue cells across halves, mismatched-issue benchmarks, shuffled core first-mention times, and framing after removing wire-credited core stories.The wire-free core check applies to framing only; the story census lacks credit codes for recalculating salience and timing, and own-country share uses titles because article text is unavailable.Source: B Validation and checks on the measures · p. 15
• Table S6 checks Media Cloud series against GDELT title-based series, including source correlations and salience-following comparisons by tier and cell.The note specifies that GDELT weeks with fewer than five articles are dropped and describes the source-specific weekly issue-share calculations.Source: B Validation and checks on the measures · p. 16, Table S6 · p. 16

• The test note defines the isomorphism decision as a positive intercept and the herding decision as a negative coefficient, with Holm-adjusted p < .05 across the pair.The isomorphism test removes wire-credited stories from outlet and core framing samples; the herding test relates first-mention lag to the number of core outlets mentioning an event that day.Source: C Test output and simulation read-outs · p. 17
• The tier models measure following as event-adjusted salience correlation, bias-corrected framing similarity, or negative median timing lag, with positive coefficients indicating closer periphery following.The note points to nine Holm-adjusted confirmatory decisions and describes the index check as replacing tier with Gunaratne’s 2022 score.Source: C Test output and simulation read-outs · p. 17
• Wire reliance is the share of a cell’s coded foreign stories crediting AP, AFP, or Reuters, so its coefficient spans cells from no wire credits to all foreign stories credited.The confirmatory model omits correspondent presence because it fell below the stability floor; the first falsifier compares the net tier-gap confidence bound with half the H1 estimate.Source: C Test output and simulation read-outs · p. 18
• The nine confirmatory decisions cover H1–H3 across salience, framing, and timing; the second falsifier compares high-stake periphery following with mismatched-issue benchmarks.H3 codes stake as Ukraine 0, Gaza 1, and COP 2; the note specifies paired one-sided tests for the high-stake periphery cells.Source: C Test output and simulation read-outs · p. 19
• Simulation errors average across cells with an available value and across 20 seeds, with dependence, imitation, and herding rules applied in that order.A cell needs at least eight qualifying weeks for salience, five coded published items for framing, or three published core or event items for timing; wire items carry no lag.Source: C Test output and simulation read-outs · p. 19
• At fitted w = 1.00, the dependence-rule cap is n × (1 − W) × C; 17 of 18 outlets keep no own story, and an outlet with C = 0.00 publishes wire copy only.Correspondent presence was exploratory, and C was fixed as the rule’s correspondent-presence input before the data.Source: C Test output and simulation read-outs · p. 20