Understanding Geographical Trends in Stack Overflow Code Maintenance Discussions – Replication Package

Code maintenance sustains the long-term value of software systems, yet large-scale evidence about how maintenance work is discussed across regions remains limited. This study analyses code-maintenance discussions on Stack Overflow using a geocoded corpus derived from the June 2022 Stack Exchange Data Dump, examining how temporal patterns (RQ1), topic distributions (RQ2), and technology associations (RQ3) vary across six continents. The headline finding is a consistent split: the four ISO/IEC/IEEE 14764 maintenance categories (corrective, adaptive, perfective, preventive) are universally present across continents, while their proportions, sub-topics, and technology associations vary by region. The June 2022 snapshot precedes the mass adoption of LLM coding assistants and so serves as a pre-LLM baseline for maintenance discourse. This replication package accompanies the manuscript and is provided so that others can inspect the methodology, verify the reported results, and extend the analysis. It contains the source-database schema, per-continent analysis outputs for each research question, sensitivity and validation artefacts, and aggregated data series. Package contents Data/: aggregated monthly time series, including the normalised post-share series (Posts Normalised.parquet). The underlying raw source is the public June 2022 Stack Exchange Data Dump (linked below) and is not redistributed here. Figures/: the complete source-database schema (SO Database Schema.html). Results/Evolution (RQ1)/: changepoint segment timing per continent (Segments.xlsx) from the B-spline mixed-effects (NLME) regression combined with PELT changepoint detection. Results/Topics (RQ2)/: per-continent LDA keyword lists and topic visualisations, the topic-to-category rationale (Topics Alignment.docx), and the maintenance-topic distribution (Maintenance Topics.xlsx). Results/Associations (RQ3)/: tag co-occurrence outputs, including Fisher's exact test results with odds ratios (Test Results.xlsx) and network structural properties such as density and Louvain modularity (Network Properties.xlsx). Sensitivity/: robustness and validation artefacts, including the alpha-beta threshold sweep with Jaccard scores (Jaccard.xlsx), LDA topic validation (Topic Validation.xlsx), topic-to-category mapping validation (Topic-to-Category Mapping.xlsx), and per-document permutation distances (Pairwise Distances.xlsx). Interrater/: inter-rater reliability details (Cohen's kappa and Gwet's AC1) for tag and topic validation. Perplexity Results/: per-continent LDA perplexity plots used to select the topic count k. Corpus and provenance. Source: Stack Exchange Data Dump, June 2022 release. User-provided locations were resolved to countries using the MapQuest geocoding service and grouped into six continents. Posts were filtered to code-maintenance content using validated tag-relevance and tag-significance heuristics followed by manual validation, yielding 44 final tags, each mapped to an ISO/IEC/IEEE 14764 maintenance category. The resulting corpus comprises 319,967 posts (108,362 question threads and 211,605 answers) from 173 countries, spanning August 2008 to June 2022 (167 monthly observations per continent). Methods and software. The analysis pipeline uses BeautifulSoup (HTML-to-text parsing); ruptures v1.1.9 and statsmodels v0.13.5 (RQ1 NLME regression and PELT changepoint detection); NLTK and WordNet (tokenisation, stop-word removal, lemmatisation); Gensim v4.3.2 and pyLDAvis v3.4.1 (RQ2 LDA topic modelling and visualisation); and Apriori association-rule mining (RQ3). The key analysis parameters for each research question are documented in the manuscript and reflected across the result files. File formats. Parquet (aggregated series), XLSX (results, sensitivity, and validation tables), HTML (database schema), and DOCX (topic-to-category rationale). Citation. Please cite this dataset by its DOI (10.5281/zenodo.20746992) together with the associated manuscript once published. License. Creative Commons Attribution 4.0 International (CC BY 4.0). External resource. Raw source data: Stack Exchange Data Dump (June 2022), https://archive.org/details/stackexchange_20220606