Source-linked AI summary
Threat Amplified, Blame Restrained: LLM-Assisted Media Framing Analysis of the 2026 Bangladesh Measles Outbreak
Shahan Ahmed
TL;DR
Outbreak-media framing in low- and middle-income countries remains understudied, including whether coverage reflects epidemiological or political dynamics. This paper analyzes archived English-language Bangladeshi headlines using a validated LLM annotation pipeline. Coverage grew more negative and threat-focused as mortality accumulated, while blame remained sparse and mostly systemic.
Problem
Computational studies of outbreak coverage have focused mainly on high-income settings and dominant platforms, leaving LMIC framing dynamics underexamined.
Source: §1 · pp. 1–2Method
The study builds an Internet Archive corpus of 403 headlines and applies a locked sentiment-and-stance codebook with LLM labels validated against human adjudication.
Source: §1 · p. 2, §3.2 · p. 4, §3.3 · p. 4Results
Coverage became significantly more negative and risk-focused as cumulative mortality rose, while 32 of 37 blame headlines (86%) attributed fault to systemic failures.
Source: §4.3 · p. 6, §4.4 · p. 7, §6 · p. 9Takeaways & Limitations
In this outbreak’s English-language coverage, threat amplification substantially outweighed attribution of political fault.
Source: §6 · p. 9The corpus covers seven national English-language outlets, excluding Bangla coverage, and headline-level analysis does not capture article-body nuance.
Source: §5.5 · p. 9
Abstract
from arXiv · showhide
How news media frame and emotionally code a public health emergency shapes public risk perception and trust, yet outbreak-coverage dynamics remain understudied for low- and middle-income countries (LMICs). We examine sentiment and stance in English-language Bangladeshi coverage of the 2026 measles outbreak -- the country's most severe in two decades, with over 97,000 suspected cases and 600 deaths across 61 of 64 districts, unfolding after the 2024 change of government and a 2024-2025 vaccine stockout. Using the Internet Archive, we build a reproducible corpus of 403 headlines from seven national outlets (396 in-window in 2026), label them for binary sentiment and four-way stance via a large language model under a locked codebook, and validate against a two-coder human-adjudicated gold standard (n=153; Cohen's kappa=0.89 stance, 0.75 sentiment). Aligned to the DGHS epidemic curve, coverage grew significantly more negative (56% to 88% negative; Cochran-Armitage z=4.12, p<.001) and risk-amplification framing intensified (44% to 84%; z=4.15, p<.001). Media negativity lagged incidence, tracking cumulative mortality. Contrary to the political backdrop, blame remained a minority frame (~9% overall) and was overwhelmingly systemic (32 of 37, 86%) rather than directed at named actors. The pipeline offers a scalable, transparent method for LMIC outbreak-media analysis; Bangladeshi coverage amplified threat far more than it assigned political blame.
1 Introduction
Computational research has underexamined outbreak coverage in low- and middle-income countries, despite its potential relevance to public fear, blame, and action. This study examines whether politically charged coverage of Bangladesh’s 2026 measles crisis tracked systemic or political explanations and presents a validated, reproducible LLM annotation pipeline.
1 Introduction
- Outbreak-media framing in low- and middle-income countries remains underexamined, although it shapes what publics fear, whom they blame, and whether they act.Source: §1 · p. 1
- The paper asks whether coverage of Bangladesh’s severe 2026 measles crisis foregrounded systemic critique, personal political blame, or technocratic reassurance.The outbreak produced more than 97,000 suspected cases and 600 deaths across 61 of 64 districts.Source: §1 · pp. 1–2
- The politically contested interim administration and 2024–2025 vaccine stockout made programme disruption and procurement delays politically legible explanations for the outbreak.Source: §1 · p. 2
- The study combines analysis of sentiment and framing across outbreak phases with a reproducible, Internet Archive-based LLM annotation pipeline validated against human labels.It reports stance agreement of κ=0.89 and significant intensification of negativity and threat framing.Source: §1 · p. 2
2 Related Work
The related work situates outbreak coverage within issue-attention and framing research, motivates validation when using LLMs as annotators, and establishes the structural-versus-political framing question in Bangladesh’s measles outbreak.
2.1 Issue-Attention and Framing of Health Crises
- Issue-attention research describes outbreak coverage moving from near-absence to an onset-driven volume spike and increasing emphasis on severity as consequences accumulate.Framing research distinguishes threat-focused from reassurance-focused coverage and systemic from individual blame; the latter distinction bears on accountability.Source: §2.1 · p. 2
2.2 LLMs as Annotators for Computational Social Science
- LLM annotations can be comparable to trained coders, but stance detection is harder than sentiment and varies more across models and prompts.The cited best practice is to validate LLM labels against a human gold standard using standard inter-annotator statistics.Source: §2.2 · p. 3
2.3 The 2026 Bangladesh Measles Outbreak
- The Bangladesh outbreak followed declining vaccination coverage and a 2024–2025 supply disruption, while domestic analyses emphasized programmatic failures as proximate causes.Whether English-language media reflected structural or political framing is the paper’s empirical question.Source: §2.3 · p. 3
3 Methods
The study constructs an archived corpus of Bangladeshi English-language measles headlines, codes sentiment and stance using a locked codebook and LLM, validates labels against human adjudication, and tests phase trends against the DGHS epidemic curve.
3.1 Sampling Frame and Search Terms
- The sampling frame comprised seven national English-language newspapers, with measles-only search terms and pre-2026 coverage retained as a baseline.Three candidate outlets were excluded for technical reasons, including opaque URLs or insufficient archival coverage.Source: §3.1 · p. 3
3.2 Archival Retrieval and Extraction
- Internet Archive captures yielded 403 headlines across seven outlets, including 396 in-window in 2026, after URL-based retrieval and headline extraction.Prothom Alo required retrieval of archived story pages because its URLs used opaque identifiers; six extraction stubs and one non-Bangladesh headline were removed.Source: §3.2 · p. 4
3.3 Codebook and Categories
- The locked codebook labels sentiment and four stance classes, treating quantified harm as risk amplification and requiring asserted fault for blame.Sentiment was later collapsed to negative versus non-negative after coders disagreed on the neutral/positive boundary.Source: §3.3 · p. 4
3.4 LLM Annotation
- Claude Haiku 4.5 labeled the corpus with a fixed prompt and five few-shot examples, producing validated JSON outputs for all 403 items.Of 324 initial-corpus items, 321 (99.1%) were labeled on the first pass; the remaining three were labeled on retry, and Prothom Alo items used the same model, prompt, and codebook.Source: §3.4 · p. 4
3.5 Human Validation
- Two independent coders labeled a stratified sample, and a third analyst adjudicated stance disagreements to create a rule-based consensus gold standard.Against that standard, the LLM achieved κ=0.89 on stance and κ=0.75 on binary sentiment; blame was under-detected by approximately 17%.Source: §3.5 · pp. 4–5
3.6 Temporal Analysis
- Weekly counts and stance shares were aligned with the DGHS epidemic curve and analyzed across four outbreak phases.Tests assessed stance-by-phase association and trends in negative sentiment and risk amplification; blame trends were reported qualitatively because phase cells were small.Source: §3.6 · p. 5
4 Results
Coverage rose sharply during the outbreak, while sentiment became more negative and stance shifted toward risk amplification. Blame remained a minority frame and was predominantly systemic rather than directed at named actors.
4.1 Corpus and Coverage Dynamics
- 32 headlines appeared in March, 160 in April, 172 in May, and 32 in partially archived June, after only seven headlines before 2026.The June decline reflects both falling attention and incomplete archiving.Source: §4.1 · p. 5, Figure 1 · p. 6
4.2 Sentiment Became Monotonically More Negative
- Negative headlines rose from 56% in March to 88% in June (Cochran–Armitage z=4.12, p<0.001), while negativity continued climbing as suspected cases receded and cumulative deaths rose.Average daily suspected cases peaked at 1,277 during 7–13 May and declined to 981 by mid-June; the figure compares weekly negativity and cumulative deaths with the case curve.Source: §4.2 · p. 5, Figure 2 · p. 7
4.3 Framing Shifted Toward Risk Amplification
- Risk-amplification framing rose from 44% in March to 84% in June (Cochran–Armitage z=4.15, p<0.001), with stance composition differing significantly across phases (χ2(9)=34.44, p<0.001; Cramér’s V=0.17).Table 1 reports stance composition and negativity by phase, with headline counts for each phase.Source: §4.3 · p. 6, Table 1 · p. 6
- Reassurance rose from 9% in March to 22% in April during the vaccination campaign launch, then fell to 11% in May and 3% in June.The managed-response narrative did not persist as mortality climbed.Source: §4.3 · p. 6
4.4 Blame Was Sparse and Predominantly Systemic
- Blame remained a minority frame across phases and was systemic in 32 of 37 blame headlines (86%), compared with five headlines (14%) naming individual political actors.Blame’s phase trajectory was reported qualitatively because of small cells and LLM under-detection.Source: §4.4 · p. 7
5 Discussion
The discussion interprets coverage as increasingly threat-focused as cumulative costs rose, contrasts that with restrained, largely systemic blame, and presents a validated annotation pipeline alongside important scope limits.
5.1 Interpreting the Trajectory
- The trajectory followed Downs’s issue-attention cycle, while negativity tracked cumulative mortality rather than incidence and continued rising into June as cases fell.Reassurance peaked during the vaccination campaign but did not persist as threat framing displaced it while mortality climbed.Source: §5.1 · p. 8
5.2 Threat Amplification Over Political Blame
- Despite the political backdrop, blame remained a minority, predominantly systemic frame, cautioning against assuming politically salient outbreaks produce personalized coverage.The authors report that only 5 of 37 blame headlines (14%) directed blame at individual actors.Source: §5.2 · p. 8, §5.3 · p. 8
5.3 Interpreting the Restraint on Political Blame
- Possible explanations include editorial preference, visible systemic failures, and press-freedom pressures, but headline-level data cannot adjudicate them or establish causal editorial self-restraint.The authors call for newsroom-level and article-body evidence to investigate the pattern further.Source: §5.3 · pp. 8–9
5.4 Methodological Contribution
- The transferable pipeline combines Wayback-anchored retrieval, a locked codebook, and validated LLM annotation, with strong human agreement on stance (κ=0.89).A headlinematch retrieval adaptation extends applicability to outlets with opaque URLs without changing search scope.Source: §5.4 · p. 9
5.5 Limitations
- The study is bounded by an English-language, single-outbreak corpus, sparse blame counts, headline-only analysis, and binary sentiment coding.June coverage was partially archived; reliability was checked against a two-coder gold standard, while cross-provider and prompt comparisons remain future work.Source: §5.5 · p. 9
6 Conclusion
Across 403 headlines, Bangladeshi coverage became more negative and threat-focused as cumulative mortality rose, while blame remained sparing and systemic; framing tracked cumulative cost more closely than real-time epidemiology.
6 Conclusion
- Coverage became significantly more negative and more threat-focused as cumulative mortality rose, while attribution of blame remained sparing and systemic.The paper concludes that media framing tracked cumulative outbreak cost more closely than the real-time epidemiological course.Source: §6 · p. 9
- The validated LLM-annotation pipeline offers a transparent, transferable approach to analyzing media responses to health emergencies in LMIC settings.The corpus contains 403 headlines across seven outlets.Source: §6 · p. 9
- Institutional reassurance appeared briefly during the vaccination campaign before coverage emphasized threat and systemic rather than personalized blame.Source: §6 · p. 9