Methodological Change in Mortality Forecasting, 1992–2025: A Bibliometric Analysis of 520 Publications

Dataset + analysis code underlying a bibliometric of 520 mortality-forecasting publications indexed in the Web of Science Core Collection (1992–2025). ## Contents | Folder | Files | Description ||---|---|---|| `data/` | `520_papers_classification.xlsx` | Full sample: one row per publication with title, authors, year, journal, DOI, methodological category, and author keywords. Categories: Structural/Statistical (420), Machine Learning (26), Hybrid (74). || `data/` | `classification_evidence.csv` | **Audit trail.** Per-publication evidence for the rule-based classification: matched keywords by dictionary, triggered rule (keyword logic / clinical redirect / DOI or title override), the basis of each Hybrid assignment (`Hybrid_basis`: explicit integration term vs. structural theme + ML vs. manual override), and abstract availability. Allows any reader to verify why each publication received its category. || `data/` | `regression_data.csv` | Analysis dataset for the citation-impact regressions (N = 520). || `data/` | `Total_Table1-4.xlsx` | Source tables of the manuscript: selection criteria, method comparison, methodological classification, OLS results. || `code/` | `generate_regression_data.R` | Builds `regression_data.csv` by merging the classification table with Web of Science citation exports (raw WoS records not redistributed, see below). || `code/` | `regression_results_with_vif.R`, `regression_results_final_without_vif.R` | OLS sensitivity models (raw rate, excluding 2024–2025, log(rate+1)) with HC1 robust errors and VIF diagnostics. || `code/` | `classify_mortality_forecasting.py` | **Classification rules.** Rule-based assignment of the 520 publications to the three methodological categories (keyword dictionaries with word-boundary matching on title, abstract, and keywords; documented manual override tables). Reproduces `data/520_papers_classification.xlsx` exactly (420 / 26 / 74). Requires the raw WoS export, which is not redistributed (see below). || `code/` | `negative_binomial_analysis.py` | **Main model.** Negative binomial (NB2) regression of WoS Times Cited with log(years since publication) exposure offset; HC1-robust standard errors; Poisson overdispersion likelihood-ratio test. Reproduces Table 4 of the manuscript. || `figures/` | `figure1.png` … `figure10.png` | Figures of the manuscript. | ## Variable dictionary (`regression_data.csv`) | Variable | Definition ||---|---|| `No.` | Publication identifier (order in the classification table) || `Title`, `Authors`, `Journal`, `DOI` | Bibliographic metadata || `Year` | Publication year (1992–2025) || `Category` | Structural/Statistical, Machine Learning, or Hybrid || `IsHybrid`, `IsML` | Category indicators (reference: Structural/Statistical) || `NumAuthors` | Number of authors || `Citations` | Web of Science Times Cited at retrieval || `YearsSincePub` | Years since publication (whole years, minimum 1) || `CitationRate` | Citations / YearsSincePub | ## Reproducing the main results ```bash# Negative binomial models (Table 4): requires Python >= 3.10 with# pandas, numpy, statsmodels (>= 0.14)python code/negative_binomial_analysis.py# -> console output plus code/code_output/nb_results.csv and nb_diagnostics.txt``` Expected key estimates: IsHybrid IRR = 1.84 (95% CI 1.30–2.62, p < 0.001) full sample; IRR = 2.52 (1.70–3.73, p < 0.001) excluding 2024–2025; dispersion α ≈ 1.02; Poisson restriction rejected (LR ≈ 8888 / 8826). The OLS sensitivity models (Table 5 of the manuscript) are reproduced with the R scripts in `code/` (`R` with `readxl`, `tidyverse`, `sandwich`, `lmtest`). Note that the log-transformed specification uses log(citation rate + 1). ## Classification rules and robustness The methodological category of each publication is assigned by `code/classify_mortality_forecasting.py`: keyword dictionaries (structural, machine learning, integration markers, clinical contexts) are matched with word boundaries against the title, abstract, Keywords Plus, and author keywords (all 520 records include abstracts). Integration markers require a machine-learning co-occurrence to trigger the Hybrid category; phrases such as "hybrid mortality model" (names of structural models) are explicitly excluded; a small set of documented edge cases is corrected via DOI/title override tables. The per-publication evidence for each decision is recorded in `data/classification_evidence.csv`. Inter-rater reliability of a stratified 119-publication subsample was double-coded by two human raters (raw agreement 91.6%, Cohen's kappa = 0.78), and the rule-based classifier agreed with the manual gold standard at 89.9% (kappa = 0.74, macro F1 = 0.82). A sensitivity check is provided with this deposit (printed by the ROBUSTNESS CHECK block in `code/classify_mortality_forecasting.py`). Restricting the structural dictionary to model-level terms only — removing **all** broad context words such as "mortality forecasting", "longevity", "demographic", "insurance" — reclassifies **18 of 520** publications from Hybrid to Machine Learning; the combined non-structural share (100 publications, 19.2% overall and 35.5% for 2021–2025) and all headline results are unchanged. Removing only the two smallest context words ("actuarial", "insurance") reclassifies **0** publications. The 18 reclassified papers are listed in the robustness output; none changes the category tallies' interpretation because they remain within the non-structural group. ## Excluded content and licensing of source records Raw Web of Science records (`savedrecs*.xls`) are **not** redistributed, in line with the Clarivate Analytics licence agreement. The deposited files contain all derived variables and, where available, the DOI of every publication (11 of 520 records have no DOI, mostly books and older WoS entries); this is sufficient to reproduce all reported results.