Data and code for Climate hazard impact research underrepresents vulnerable countries and cascading socioeconomic impacts

Overview This repository contains the datasets underlying the paper "Climate hazard impact research underrepresentsvulnerable countries and cascading socioeconomic impacts", a study of the socioeconomic impacts of climate-related hazards as reported in the scientific literature. Publications were screened with a two-stage, multi-label, transformer-based active-learning pipeline: a first classifier identifies sentences containing impact information (relevance), and a second classifies those sentences into 30 impact subclasses grouped under six main classes (Water, Society, Food production, Health, Infrastructure, Economy). Events were cross-referenced with the EM-DAT international disaster database. The files below include the model training/labelled data, the classified outputs, and the event-level summaries. File descriptions publication_corpus.csv — List of the 11,176 scientific publications analyzed in this study. Each entry is assigned a unique identifier and includes key metadata: publication identifiers (DOI or PMID), full article titles, and abstracts.dataset_events.csv — Publication-to-event records. Each row links a scientific publication to a hazard event at a given location. Columns include publication identifiers (doi, pmid, publicationYear), geographic fields (country, region, state, city, basin, other_location, continent, countryCode), hazard information (hazard_name, hazard, startYear, endYear), the matched EM-DAT record (emdat_event, a serialized dictionary of EM-DAT fields), and contextual attributes of the affected country (incomeGroup, population). Note that a single publication may appear in multiple rows when it concerns several locations or events. dataset_impact_statements.csv — The classified impact statements. Each row is one sentence extracted from a publication full text. Columns include the publication identifier (id), a sentence_id, the sentence text, the relevance classifier's predicted probabilities (prediction_probability_not_relevant, prediction_probability_relevant), and a set of binary indicator columns. The indicators are the 30 impact subclasses (e.g. CFLC, LVLH, DISP, INEQ, AVLB, AGRC, FRST, RECR, ECOG, …) plus the six aggregated main classes (SOCIAL, WATER, FOOD, HEALTH, INFRA, ECONOMY); a value of 1 indicates the sentence was assigned that label. Labels are non-exclusive (a sentence may have several labels). emdat_events_summary.csv — Event-level summary of the EM-DAT disasters cross-referenced in this study. One row per disaster, keyed by DisNo. (EM-DAT identifier). Columns: Number of publications (count of publications matched to the event in this corpus), Event Name, Country, Disaster Type, Total Deaths, Total Affected, and Income group. This file underlies the publications-per-fatality and publications-per-affected-person analyses. labelled_data_sentence_relevance.csv — Initial manually labelled training set for the relevance (impact vs. no-impact) classifier. One row per sentence, with paper_id, article title, the sentence text, a sentence_id, and a binary relevance label. labelled_data_sentence_relevance_active_learning.csv — Additional relevance-labelled sentences acquired during active-learning iterations. Columns: publication id, sentence text, the model's predicted value, and the assigned label. labelled_data_sentence_impact_classes.csv — Initial manually labelled training set for the impact-subclass classifier. One row per sentence (sentence_id, sentence) followed by binary columns for each impact subclass (AGRC, AGRI, ANIM, AVLB, CFLC, …, TRSP). labelled_data_sentence_impact_classes_active_learning.csv — Additional impact-subclass-labelled sentences acquired during active-learning iterations. Same subclass column structure as above, with publication id and sentence_id. labelled_hazard_characterization.csv — Publication-level hazard characterization data. One row per publication–location pairing, with bibliographic fields (doi, year, journal, title, abstract), an indicator of whether a specific hazard was identified (specific_hazard), hazard timing and naming (startYear, endYear, hazard_name, hazard), and location fields (country, iso_3166_code, region, state, city, other_location). codebook.csv — Data dictionary for the impact-classification labels used throughout this repository. Each row defines one of the 30 socioeconomic impact subclasses into which sentences were classified. Columns are: code, the short label used as the indicator-column header in the data files (e.g. CFLC, AVLB, MORT); main_class, the human-readable name of the parent impact class (Society, Water, Food production, Health, Infrastructure, Economy); main_class_acronym, the corresponding code used for the aggregated main-class indicator columns in dataset_impact_statements.csv (SOCIAL, WATER, FOOD, HEALTH, INFRA, ECONOMY); and subclass_name, the full descriptive name of the subclass. run_analyses.ipynb — Jupyter notebook (Python 3.12) that reproduces all statistical analyses and figures in the manuscript from the deposited data files.