tax_revenue_tech_review_replication

# Replication package: Digital technology and tax revenue performance — a systematic review (2020–2025) This package supports the systematic review submitted to the *International Journal of Public Administration in the Digital Age* (IJPADA). It contains the search record, corpus metadata, both reviewers' screening decisions, the study-level extraction matrix with quality appraisal, the NLP and keyword-network code, and the figures. ## Structure| Folder / file | Content ||---|---|| `01_search/search_strings.md` | Databases, dates, exact search strings and filters || `01_search/prisma_counts.csv` | Record counts for every PRISMA 2020 stage || `02_corpus/corpus_684_metadata.csv` | 684 screened records: ID, title, authors, year, DOI, database, keywords || `02_corpus/included_ids_289.csv` | IDs of the 289 included studies || `03_screening/codebook_screening.md` | Inclusion criteria, exclusion codes E1–E7 and v2 rule clarifications || `03_screening/screening_decisions_684.csv` | Reviewer A (human) and reviewer B (AI) decisions, agreement, adjudication and final decision || `03_screening/reviewer_B_rule_changes_v1_v2.csv` | Reviewer B decisions before and after the v2 rule clarifications || `04_extraction/extraction_specification.md` | Variables, codes and proposition (P1–P4) rules || `04_extraction/extraction_matrix_v2.csv` | 487 determinant–outcome relationships from 99 studies (one relationship per row) || `04_extraction/mmat_appraisal.csv` | MMAT 2018 appraisal of the 34 full-text technology studies || `04_extraction/table3_cat.csv`, `table4_tech.csv`, `v2_*.csv` | Consistency tables reported in the article || `05_nlp/nlp_pipeline.py` | Text cleaning, phrase detection, category dictionary, figures, sensitivity analyses S1–S4, dictionary validation || `05_nlp/outputs/` | Dictionary (Appendix B), top terms, category shares, keyword clusters, sensitivity, validation, `summary.json`, VOSviewer thesaurus || `06_figures/` | Figures 1–7 (TIFF, 600 dpi, LZW) | ## Screening variables`A_*` = reviewer A (author); `B_*` = reviewer B (AI assistant, Claude, Anthropic). Decisions: INCLUDE / EXCLUDE / UNSURE. Exclusion codes: E1 evasion or avoidance outcome; E2 tax is the cause, not the outcome; E3 non-tax or non-public revenue; E4 not empirical; E5 off-topic; E6 duplicate; E7 compliance cost or tax acceptance. `agreement = disagree` records were adjudicated; `adjudication_rationale` gives the reason and every final decision was confirmed by the author. `kappa_sample` marks the records used to compute Cohen's kappa. ## Extraction variablesKey fields: `determinant`, `category` (ECO, INST, ADM, DEM, POL, BEH, TEC), `tech_instrument`, `plane` (capacity / conversion), `sign`, `significance`, `role`, `conditioning_factor`, `mechanism`, `proposition`, `evidence_source` (full text / abstract), `mmat_score`, `verified` (Y = checked by the author against the full text). `sign_favourable` harmonizes signs when the outcome is adverse (e.g., a VAT gap), so that consistency = favourable and significant relationships / relationships with known significance. ## Reproducing the NLP analysis```pip install pandas numpy matplotlib networkx scipy scikit-learnpython nlp_pipeline.py --corpus corpus.csv --ids included_ids.txt --out outputs --coding ../04_extraction/extraction_matrix_v2.csv```The pipeline needs a `corpus.csv` with an `abstract` column. Abstracts are not redistributed here because they are copyrighted by publishers and the database licenses (Scopus, Web of Science) do not allow redistribution; they can be retrieved with the DOIs in `corpus_684_metadata.csv` or by re-running the search in `01_search`. No language-model downloads are needed, and seeds are fixed, so results are deterministic. ## Use of AIAn AI assistant (Claude, Anthropic) acted as the second screener (reviewer B), proposed adjudications, and performed the initial extraction and MMAT appraisal. The author screened all records independently, confirmed every final decision and verified a subsample of relationships against the full texts (`verified = Y`). ## LicenseData and documentation CC BY 4.0; code MIT (see `LICENSE.md`).