Beyond Evidence: How Framing Shapes Public Health Policies During Health Crises — Replication Package (Data, Code, and Models)

Replication package for the article “Beyond Evidence: How Framing Shapes Public Health Policies During Health Crises”, accepted at the Policy Studies Journal. The study asks whether the level of evidence that accompanies policy action depends on how a problem is framed, and on who frames it. It compares the COVID-19 press conferences of Quebec (n = 190; March 2020–March 2022), where political decision-makers held greater authority, and Sweden (n = 160; March 2020–February 2022), where scientists enjoyed more autonomy. Fine-tuned BERT classifiers (CamemBERT for French, KB-BERT for Swedish) annotate more than 100,000 five-sentence segments for COVID-19 relevance, journalist questions, frames (dangerous, moderate, neutral), references to suppression and mitigation measures, and cited scientific evidence. The aggregated series feed OLS models with frame-by-evidence interactions, Granger causality tests with Benjamini–Hochberg correction, a pooled cross-jurisdictional Wald test, and extensive robustness diagnostics. Files manuscript.pdf — the accepted manuscript (English and French abstracts). EVD.COVID_ANALYSIS-code-data.zip — the complete reproducible project: raw press-conference corpora, preprocessed five-sentence segments, manual annotations with the conference-grouped train/validation/test split, machine-annotated corpora, epidemiological series, stringency indices, all analysis scripts, generated tables and figures, Makefile, variable codebook, and full documentation (README). models.zip — the ten fine-tuned BERT checkpoints used in the article; unzip directly under models/. Reproducing the results: unzip the code-data archive, then run make setup followed by make analysis (Python 3.10+). Every statistical table and figure of the article is regenerated from the data. To retrain the classifiers from scratch, run the numbered scripts in Scripts/3_training_and_annotation/ (requires torch and the AugmentedSocialScientistFork package); otherwise use the provided checkpoints. Evaluation protocol: all reported model metrics are computed on a held-out test set composed of whole press conferences that contributed no segment to training, so that no overlapping five-sentence context window spans the partitions. Data sources: Quebec transcripts from a prior study by the author; Swedish transcripts obtained from the Public Health Agency of Sweden through an access-to-information request; epidemiological data from INSPQ, ECDC, and Our World in Data; stringency indices from OxCGRT (Sweden) and IRPP (Quebec). License: the record is distributed under CC-BY-4.0; the code carries the MIT license (included in the archive). This record supersedes the development archive 10.5281/zenodo.12785800.