Rating–text inconsistency, purchase verification and online review helpfulness

This repository contains supporting data, Python analysis code, a fitted text-rating model, computational outputs, tables and figures for “Rating–text inconsistency, purchase verification and online review helpfulness”. The analysis uses 479,493 reviews from six categories in Amazon Reviews’23, covering 2019–2022. Text-model development uses 140,665 product-disjoint review records. Files analysis_code.zip: Executable Python scripts, pinned dependencies, model settings, data dictionaries, licences and repository setup instructions. Includes the four exploratory analyses and code drawing Figures 1–2. analysis_outputs.zip: Estimates and covariance matrices for 17 core models and four additional exploratory specifications, contrasts, diagnostics, test predictions, Tables S1–S3, manuscript tables and figures, verification outputs and the adapted STROBE reporting checklist. analysis_derived.csv and analysis_derived.csv.gz: Numerical dataset containing the 479,493 analysis records. source_matching_manifest.csv and source_matching_manifest.csv.gz: Source-matching metadata for 620,158 development and analysis records, including product identifiers and SHA-256 hashes of review bodies. development_split_manifest.csv and development_split_manifest.csv.gz: Training, validation and test assignments for 140,665 text-model development records. text_model.joblib: Fitted text vectorizer, classifiers, mixture weight, isotonic calibrator and training configuration. F1000Research_Figures_Tables_20261030.zip: Manuscript figures in EPS, SVG, PNG and TIFF formats and tables in CSV format. README.txt: File descriptions, source access, setup and reproduction instructions, licensing and component checksums. CSV and GZIP formats Each compressed/uncompressed pair contains the same records. Do not combine the two formats as additional observations. Use the original .csv.gz files with the supplied setup and analysis scripts without renaming or recompression. Source data and privacy Raw review text and direct user identifiers are excluded. Product identifiers and text hashes in the manifests support linkage to the source collection; the manifests are not described as irreversibly anonymous. Original source files are available from https://amazon-reviews-2023.github.io/ under the provider’s terms. Licensing Author-created derived data, metadata, model artefacts, documentation and outputs are provided under CC BY 4.0. Python source code is provided under the MIT licence. Original third-party content remains subject to the source provider’s terms. Version 2.0.0 This version completes the code package, aligns Figures 1–2 with the plotting script, standardises archive filenames and adds the repository instructions. The numerical datasets, fitted text model and statistical estimates are unchanged. The study was not preregistered. The additional specification checks are exploratory. Source code: https://github.com/nebiseren/review-helpfulness-inconsistency