Guardat en:
| Autor principal: | |
|---|---|
| Format: | Recurso digital |
| Idioma: | |
| Publicat: |
Zenodo
2026
|
| Matèries: | |
| Accés en línia: | https://doi.org/10.5281/zenodo.19967597 |
| Etiquetes: |
Afegir etiqueta
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
Taula de continguts:
- <p>Pre-registration of confirmatory hypotheses H1–H4 for the SLOPDET Path B trust-geometry audit of Bangladesh's Bangla-language news ecosystem (Q4 2019 – Q1 2024).</p> <p>Frozen on 2026-05-02.<br>Git commit at deposit: 1733ad54183d4eac63ad28c0df6401f0ff60907d<br>Git tag: prereg-deposit<br>DOI: 10.5281/zenodo.19967597</p> <p>This deposit binds the confirmatory hypotheses H1–H4 to the analysis pipeline at the named git commit. Any analyses conducted after the deposit timestamp, and any methodological deviation from the parameters specified in §5 of the deposited document, will be reported as exploratory in the resulting manuscript.</p> <p>RESEARCH QUESTION<br>Were the structural conditions for an information crisis detectable in Bangladesh's Bangla-language news ecosystem before the July 2024 crisis, through change-point analysis of monthly trust-geometry metrics computed on outlet co-coverage networks?</p> <p>CONFIRMATORY HYPOTHESES<br>H1 — In-strength distribution of monthly Core-4 outlet co-coverage networks fits a power-law better than a log-normal alternative in at least 12 of 54 months (Vuong likelihood-ratio test on Clauset-Shalizi-Newman MLE; LR > 0 with p < 0.10 against log-normal counts as a power-law win). H1 is statistically underpowered at n = 4 nodes per network and a negative result will be interpreted accordingly; a descriptive heavy-tailedness check on the All-Active sensitivity panel (8 outlets when active) is reported as supporting evidence but cannot trigger PASS or FAIL.</p> <p>H2 — Gini concentration T trends monotonically across the 54-month panel with absolute change of at least 0.05 (Spearman ρ on month-index versus T; residual bootstrap 95% CI on T(2024-03) − T(2019-10) using 10,000 resamples).</p> <p>H3 — At least one of {W, T, S, ECI} has a PELT change-point with permutation p < 0.05 located before October 2023 (PELT with l2 cost, penalty = 2 × variance × log(n); permutation p-values from 1,000 random shuffles; WBS2 cross-check via R breakfast::breakpoints.wbs.thresh).</p> <p>H4 — BNAD ECI trajectory differs from Klein et al. (2022) BLM trajectory by factor at least 2 (absolute log2 ratio at least 1) in at least one month (95% bootstrap CI on log2-ratio with 10,000 resamples per month, Bonferroni correction across months).</p> <p>Decision rules and failure conditions for each hypothesis are specified in §4 of the deposited document. Exploratory analyses outside the H1–H4 scope will be clearly labelled in any resulting manuscript.</p> <p>PRIMARY PANEL<br>The Core-4 primary panel: ittefaq (314,016 articles, 54/54 panel months), janakantha (263,265 articles, 54/54), samakal (226,487 articles, 54/54), and dhaka_tribune_bangla (62,972 articles, 54/54). All four outlets satisfy the preregistered G3 inclusion gate: structural completeness across all 54 panel months and average quarterly volume ≥ 1,000 articles. Article counts and month coverage were verified by Stage 1 of the pipeline at the deposited git commit.</p> <p>SENSITIVITY PANEL<br>A separate All-Active sensitivity panel includes the Core-4 plus four additional BNAD outlets that meet the volume floor but enter the panel after October 2019 and therefore violate structural completeness: ajker_patrika, daily_inqilab, ekattor_tv, manab_zamin. Sensitivity-panel results are descriptive and cannot trigger PASS or FAIL on H1–H4.</p> <p>DATA SOURCES<br>- BNAD V2 (Saad et al., 2024, Data in Brief 57:110874, DOI 10.1016/j.dib.2024.110874, Zenodo 10.5281/zenodo.11111869, CC BY-NC 4.0). Approximately 2.23M articles across 9 Bangla-language outlets.<br>- BanFakeNews (Hossain et al., LREC 2020, https://aclanthology.org/2020.lrec-1.349/) — outlet credibility prior c_v derived as empirical-Bayes-smoothed proportion of authentic items per outlet (Beta(2,2) prior).<br>- Klein et al. (2022) Black Lives Matter trajectory — Anglophone comparator for H4, accessed via the published replication archive.</p> <p>AMENDMENT (2026-05-02)<br>The pre-registration document includes one amendment substituting the Pólya filter (Marcaccioli & Livan 2019, Nature Communications 10:745, https://doi.org/10.1038/s41467-019-08667-3) for the Serrano-Boguñá-Vespignani disparity filter (Serrano et al. 2009, PNAS 106:6483) as the backbone-extraction method on monthly co-coverage networks.</p> <p>Rationale: pre-Stage-4 verification on Q2-2022 BNAD data confirmed that the disparity filter is mathematically degenerate on the Core-4 primary panel. With n = 4 nodes, an edge from any node would need to carry more than 77% of that node's total strength to clear the α = 0.05 threshold; this cut-off is empirically unattainable on Core-4 monthly co-coverage networks. The Pólya filter generalises the disparity filter through a Pólya-urn null model parameterised by reinforcement strength a; the limiting case a → 1 recovers the disparity filter asymptotically, while a = 0 recovers the multinomial null and is more permissive. The primary setting (a = 0) avoids the n = 4 degeneracy; the sensitivity setting (a = 1) allows direct comparison to the original disparity-filter framing.</p> <p>The amendment was made on 2026-05-02 before any Stage 6 (BiCM null model) confirmatory analysis was run. Stage 4 had been run on Q2-2022 only as a smoke test. H1–H4 statements are unchanged by the amendment; the change affects only how monthly co-coverage networks are pruned before metric computation. Full rationale is recorded in §5 and the ## Amendments section of the deposited document.</p> <p>VALIDATION GATES<br>The pipeline enforces four binary gates documented in §6 of the deposited document:<br>- G1 (pre-registration deposit): this deposit and the populated DOI/URL/SHA fields in the pipeline configuration.<br>- G2 (backbone retention): Pólya filter at a = 0 retains ≥ 30% of edges per month averaged across the panel; sensitivity setting at a = 1 retained as fallback.<br>- G3 (outlet structural completeness): every Core-4 outlet has articles in all 54 panel months and averages ≥ 1,000 articles per quarter. Verified at deposit.<br>- G4 (BiCM null convergence): all 1,000 BiCM rewirings converge under bicm_max_iter iterations; chain-Metropolis fallback documented if non-convergence persists.</p> <p>STOPPING RULES<br>Sample size is fixed: 54 monthly time points, 1,000 BiCM null draws per month, 4 outlets in the Core-4 panel. Decision thresholds are pre-specified in §4 and not modified after observing results. Hypotheses are not added or removed after deposit. The Core-4 panel composition is determined by the two pre-registered inclusion criteria applied to BNAD V2 contents and is not adjusted in response to outcomes.</p> <p>REPRODUCIBILITY<br>- Global random seed: 20260501.<br>- Package versions are pinned in requirements.txt at the deposited git commit.<br>- Each pipeline stage writes a manifest recording git commit, seed, parameters, input hashes, and output hashes.<br>- Code is version-controlled. The 40-character SHA above is the canonical reference for the codebase as deposited.</p> <p>CITATION<br>Khan, D. (2026). Structural Vulnerability of Bangladesh's Bangla-Language News Ecosystem in the 4.5 Years Preceding the July 2024 Information Crisis: A Trust-Geometry Audit Using Configuration-Model Null Inference and Wild Binary Segmentation [Pre-registration]. Zenodo. https://doi.org/10.5281/zenodo.19967597</p>