Chernobyl ¹³⁷Cs and neoplasm mortality across Türkiye — data and analysis code

Abstract

# Chernobyl ¹³⁷Cs and neoplasm mortality across Türkiye — data and analysis code **Distinguishing spatial confounding from radiation effects in ecological studies through an apparent Chernobyl–neoplasm mortality gradient across Türkiye** This repository is the reproducibility compendium for the study above. It contains the province-level analytic dataset (81 Turkish provinces, pooled 2018–2024, 53 variables) and a single, self-contained Python pipeline that regenerates every table and figure in the article. **Short summary.** An apparent positive association between modelled Chernobyl-derived caesium-137 (¹³⁷Cs) and province-level neoplasm mortality is shown to be better explained by **spatial confounding** than by radiation. The evidence is a falsification framework: dose reconciliation, sequential (hierarchical) confounder adjustment, negative-control outcomes, an order-invariant (Shapley) decomposition of the sequential adjustment, penalized regression with bootstrap, explicit spatial regression (spatial-error and spatial-lag models, with Lagrange-multiplier diagnostics and spatial-lag impacts), analyses of residual confounding by historical smoking, the E-value, and a ⁹⁰Sr collinearity check. --- ## Repository contents ```.├── README.md This file├── CHANGELOG.md Version history├── CITATION.cff How to cite (Citation File Format)├── .zenodo.json Zenodo deposit metadata├── requirements.txt Exact Python dependencies (Python 3.12)├── LICENSE-CODE-MIT.txt License for the code (MIT)├── LICENSE-DATA-CC-BY-4.0.txt License for the data (CC BY 4.0)├── data/│ ├── data.xlsx Analytic dataset (sheet 'data'; 81 × 53)│ └── data_dictionary.csv Codebook: variable, description, unit, source, notes├── code/│ └── chernobyl_neoplasm_analysis.py Single-file analysis pipeline└── results/ Reference outputs (regenerated by the script) ├── ALL_RESULTS.xlsx 39 result sheets ├── Fig2_dose_reconciliation.png Dose reconciliation ├── Fig3_negative_control.png Negative-control outcomes ├── Fig4_sequential_attenuation.png Sequential attenuation ├── FigS1_block_order_shapley.png Order-invariant (Shapley) decomposition of the attenuation ├── FigS2_coefficient_paths.png Penalized-regression coefficient paths ├── FigS3_penalized_R2.png Out-of-sample test R² (repeated splits) └── FigS4_moran_lisa.png Moran scatter + LISA quadrant counts``` ## The dataset `data/data.xlsx` (sheet **`data`**) has one row per Turkish province (N = 81) and 53 columns. Full definitions, units, and sources are in **`data/data_dictionary.csv`**. Key fields: - **`NEOP`** — neoplasm mortality (ICD-10 C00–D48), deaths per 100,000, 7-year pooled mean (primary outcome).- **`UNSCEAR_CS137`, `WHO_CS137`** — the exposure. These are **modelled first-year adult effective dose (mSv yr⁻¹)** from Simsek et al. (2014), under UNSCEAR- and WHO-referenced calibrations. **They are NOT deposition densities (kBq m⁻²).** The two calibrations are almost perfectly collinear (r > 0.99) and are used in separate specifications.- **`CHERNOBYL_DISTANCE`** — population-weighted great-circle distance to the Chernobyl NPP (km).- **Confounders central to the analysis** — age-band shares (`AGE_45_54`…`AGE_75_AND_OVER`), `HUMAN_DEVELOPMENT_INDEX`, `GDP`, `EDUCATION`, `UNEMPLOYMENT`, `SMOKING`, `OBESITY` (both available only as current NUTS-1, 12-region, estimates), physician density (`PHYSICIANS_PER_100K` and components), hospital beds (`BEDS_PER100K`), urbanisation (`URBAN_PCT`), and internal migration (`NET_MIGRATION_RATE`, `GROSS_MIGRATION_RATE`).- **Negative-control outcomes** — cause-specific mortality rates `CIRC`, `RESP`, `NEURO`, `ENDOC`, `EXTERNAL`, `COVID`, `OTHER` (per 100,000); `UNKNOWN_DEATH` (ill-defined-cause) as a registration-quality index.- **`PERSON_YEARS`** — total person-years (Σ 2018–2024 population), the offset for count models. > **Note on two fields.** `UNKNOWN_DEATH` is a registration-quality index whose exact unit (proportion of deaths vs rate per 100,000) should be confirmed against the source; its near-zero correlation with the gradient makes this immaterial to the conclusions. `migration` is a legacy column superseded by `NET_MIGRATION_RATE`/`GROSS_MIGRATION_RATE`. `PM2_5` is incomplete and is not used as a primary covariate. ## How to reproduce Requires **Python 3.12**. ```bash# 1. (optional) create a clean environmentpython -m venv .venv && source .venv/bin/activate # 2. install exact dependenciespip install -r requirements.txt # 3. run the pipelinepython code/chernobyl_neoplasm_analysis.py``` The script auto-locates `data/data.xlsx` and writes outputs to `results/`:`ALL_RESULTS.xlsx` (39 sheets) and eight PNG figures. It prints a detailed log, including a **`Key_numbers`** sheet that lists every headline number for line-by-line cross-checking against the article's tables. **Google Colab.** Upload `data/data.xlsx` to the session, paste the script, and run all. Missing packages (`libpysal`, `esda`, `spreg`) are installed automatically at start-up. ## What the pipeline computes Descriptive statistics; bivariate and exposure-axis correlations; primary OLS (HC3) and negative-binomial models; ΔR² decomposition; **Moran's I + LISA** and **explicit spatial-error / spatial-lag regression**; **penalized regression (Ridge/Lasso/Elastic-Net) with a 1000-replicate bootstrap**; sequential (hierarchical) adjustment; **dose reconciliation** with a derived first-year→cumulative multiplier and multi-specification observed/expected comparison; a radon dose-tracking negative control; the **E-value**; a **⁹⁰Sr collinearity** demonstration; new-covariate impact; a churn (population-turnover) interaction; registration-quality and detection-capacity robustness; **negative-control outcomes** across seven causes of death (with Benjamini–Hochberg correction) and their coupling to the socio-demographic axis; an **order-invariant decomposition** of the sequential adjustment (all 64 block-subset models, exact Shapley values over 720 orderings, leave-one-block-out); **spatial diagnostics** (Lagrange-multiplier tests, likelihood-ratio/AIC comparison, direct/indirect/total spatial-lag impacts); **historical-smoking** analyses (NUTS-1 fixed effects, flexible specifications, bias analysis); a particulate-matter exclusion sensitivity; a **dose-reconciliation sensitivity** to risk coefficient (including the INWORKS low-dose slope) and DDREF; and a scale-reconciliation table (IRR vs rate-scale vs standardized coefficients). ## Data sources Compiled from public sources: the **Turkish Statistical Institute (TÜİK)** (cause-of-death and population statistics, internal migration, degree-of-urbanisation, socioeconomic indicators); the **Republic of Türkiye Ministry of Health** (Health Statistics Yearbooks: physicians, hospital beds, smoking, obesity); modelled ¹³⁷Cs effective doses from **Simsek et al., *Science of the Total Environment* 499 (2014) 74–88**; a Turkish indoor-radon survey; satellite products (Sentinel-5P/TROPOMI, MODIS). See `data/data_dictionary.csv` for per-variable attribution and the article for full citations. ## Licenses - **Code** (`code/`): MIT License — `LICENSE-CODE-MIT.txt`.- **Data** (`data/`): Creative Commons Attribution 4.0 (CC BY 4.0) — `LICENSE-DATA-CC-BY-4.0.txt`. Please also respect the terms of the original data providers listed above. ## How to cite If you use these materials, please cite **both** this deposit and the associated article. A machine-readable citation is in `CITATION.cff`. Example: > Etli, Y., Korkmaz Yalçın, D., Kartal, E., & Aşırdizer, M. (2026). *Distinguishing spatial confounding from radiation effects in ecological studies: Chernobyl ¹³⁷Cs and province-level neoplasm mortality across Türkiye — data and analysis code* (Version 2.1.0) [Data set]. Zenodo. https://doi.org/ ## Contact Yasin Etli — yasinetli@yyu.edu.tr — Department of Forensic Medicine, Faculty of Medicine, Van Yüzüncü Yıl University, Van, Türkiye. ## Disclaimer This is an ecological (province-level) analysis and does not support individual-level causal inference. The study finds **no evidence** that Chernobyl fallout has measurably increased neoplasm mortality in Türkiye and provides no basis for changes to cancer screening or clinical practice.