Estimating the Perinatal Health Benefits of Hypothetical Pollution Interventions in Santiago, Chile Using Parametric G-Computation 🏭 👶

💰 Funding

FONDECYT Nº 11240322: Climate change and urban health: how air pollution, temperature, and city structure relate to preterm birth

Additional support: (CR)², Chile, FONDAP/ANID 1523A0002

👥 Research Team

📬 Estela Blanco (estela.blanco@uc.cl) — Principal Investigator / Corresponding Author

📬 José Daniel Conejeros (jdconejeros@uc.cl) — Research Assistant / Repository Manager

Research Collaborators: Ismael Bravo, Felipe Cornejo, Axel Osses, & Tarik Benmarhnia

📌 Publication

Work in progess.

---

🎯 Project Overview

Background

Ambient air pollution is a major environmental risk factor for adverse perinatal outcomes. Preterm birth (delivery before 37 completed weeks of gestation) remains a leading cause of neonatal mortality and long-term morbidity. While epidemiological evidence links PM₂.₅, NO₂, and O₃ to preterm birth, fewer studies have translated observed exposure–response associations into population-level estimates of the health benefits of realistic pollution-reduction scenarios, particularly in Latin American urban settings with heterogeneous exposure patterns.

Objective

To estimate the population-level impact of hypothetical reductions in PM₂.₅, NO₂, and O₃ on the cumulative risk of preterm birth among singleton births in urban Santiago, Chile (2010–2020), using distributed-lag Cox models combined with parametric g-computation.

Methods

We conducted a population-based retrospective cohort study using:

• Study population: Singleton live births to women residing in the urban conurbation of Santiago (33 municipalities of the Province of Santiago plus Puente Alto), 2010–2020
• Sample size: 713,918 births after exclusion criteria (51,081 preterm; 7.2%)
• Exposure: Weekly municipality-level PM₂.₅, NO₂, and O₃ from the national air-quality monitoring network, spatially interpolated to municipal centroids using ordinary kriging (primary) and inverse distance weighting (IDW) (sensitivity)
• Additional exposures: Daily mean temperature (CR2MET), NDVI (MODIS MOD13Q1 via Google Earth Engine), and municipal social vulnerability index (SOVI)
• Exposure windows: Weekly gestational exposure series (weeks 1–44); period-specific averages for descriptives (full pregnancy, trimesters, last 30 and last 4 days)
• Outcome: Preterm birth (<37 weeks), with subcategories:
  • Very preterm (28–31 weeks)
  • Moderate preterm (32–33 weeks)
  • Late preterm (34–36 weeks)
• Distributed-lag models (DLM): Week-specific Cox proportional hazards models with time-weighted lagged exposure and delayed entry at week 28
• G-computation: Parametric g-formula applied to natural-course Cox models (weeks 28–36) under counterfactual exposure histories
• Intervention scenarios:
  • Threshold (cap): PM₂.₅ capped at 5, 10, 15, 20 µg/m³; NO₂ capped at 5, 10, 15, 20 ppbv
  • Proportional reduction: 20% reduction applied to PM₂.₅, NO₂, or O₃ across all gestational weeks
  • Single-week interventions: 20% reduction in one gestational week at a time (critical-window map)
• Covariates: Newborn sex; maternal and paternal age, education, and occupation; month and year of last menstrual period; COVID-19 period; SOVI; weekly temperature; NDVI
• Inference: Parametric bootstrap (250 replicates in manuscript; configurable in code via GFORM_BOOT_ITER)

Key Findings

• NO₂ interventions showed the most consistent protective effects on cumulative preterm birth risk. A 20% reduction across gestation was associated with a risk difference of −0.12 percentage points (PAF −1.77%). Stronger benefits were observed under lower caps (e.g., NO₂ < 5 ppbv: RD −0.48 pp; PAF −6.98%).
• O₃: A uniform 20% reduction was associated with a risk difference of −0.26 percentage points (PAF 3.80%), suggesting a meaningful share of observed preterm risk may be attributable to ozone exposure under the natural course.
• PM₂.₅: Threshold and proportional-reduction scenarios showed small and statistically uncertain effects, with confidence intervals spanning both protective and adverse directions.
• Timing matters: Single-week intervention heatmaps (Figure 5) identified gestational windows in which exposure reductions would yield the largest expected impact on cumulative preterm risk.

---

Code Structure

Setup Scripts

• 00_Code/0.1 Settings.R — Global settings and locale
• 00_Code/0.2 Packages.R — Package installation and loading (main pipeline)
• 00_Code/0.2 Packages_gform.R — Packages for g-computation pipeline
• 00_Code/0.3 Functions.R — Custom helper functions

Data Processing Scripts

• 00_Code/1.0 Pollution_process_data.R — Load and clean interpolated PM₂.₅, NO₂, O₃ series
• 00_Code/2.0 Births_process_data.R — Birth data cleaning, cohort definition, exclusions
• 00_Code/3.0 NDVI_EarthEngine_commune_extraction.py — NDVI extraction (Google Earth Engine)
• 00_Code/3.1 Temp_NDVI_data.R — Temperature and NDVI processing
• 00_Code/4.0 Climate_data_generate.R — Climate data generation
• 00_Code/5.0 Exposure_data_births.R — Weekly gestational exposure histories
• 00_Code/6.0 Join_full_data.R — Merge pollution, climate, and birth data
• 00_Code/8.0 Correlation_pollulants.R — Pollutant correlation analysis

Descriptive Analysis

• 00_Code/7.0 Descriptive_births.R — Birth and preterm trends
• 00_Code/7.1 Descriptive_exposition.R — Exposure descriptives

Statistical Models

• 00_Code/9.0 DLM_pollution.R — Distributed-lag Cox models (PM₂.₅, NO₂, O₃)
• 00_Code/9.1 DLM_plots.R — DLM visualization

G-Computation Pipeline

• 00_Code/10.0 G-Form_build_interventions.R — Build counterfactual exposure histories (Stage 1)
• 00_Code/10.1 G-Form_functions.R — Core g-formula functions
• 00_Code/10.2 G-Form_models.R — Run interventions, bootstrap, and heatmaps (Stage 2)
• 00_Code/10.3 G-Form_plots.R — Publication figures (cumulative risk, heatmaps)
• 00_Code/10.4 G-Form_table.R — Summary tables (Table 3)

---

📈 Principal Findings

Figure 1. Flowchart of Analytical Sample Construction (2010–2020)

Starting from 2,557,140 singleton births in Chile (2010–2020), sequential exclusions yielded a final analytic sample of 713,918 births in urban Metropolitan Santiago.

Note: Flowchart included in manuscript (03_Paper/CTA_PTB_Gf_Manuscript.docx).

Figure 2. Distribution of Daily PM₂.₅, NO₂, and O₃ (Kriging)

Note: Municipality-day pollutant concentrations interpolated by ordinary kriging. Histograms summarize overall, seasonal, and pollutant-specific distributions across the study period.

Preterm Birth Trends in Santiago (2010–2020)

Note: Annual prevalence of preterm birth and subcategories per 1,000 births. Analytic cohort N = 713,918.

Figure 3. Hazard Ratios for Preterm Birth by Gestational Week (DLM)

Note: Week-specific hazard ratios (HR) and 95% CIs from distributed-lag Cox models. Each point represents the acute effect of weekly exposure at gestational week t, conditional on time-weighted lagged exposure through week t−1. Models adjusted for sex, parental characteristics, calendar time, SOVI, temperature, and NDVI. Kriging-based exposures; N = 713,918.

Figure 4. Cumulative Preterm Birth Risk Under 20% Pollution Reduction Scenarios

Note: Cumulative risk of preterm birth (weeks 28–36) under the natural course (observed exposure) vs. a uniform 20% reduction in weekly PM₂.₅, NO₂, or O₃. Estimates from parametric g-computation with distributed-lag Cox models.

Figure 5. Critical-Window Heatmap of Risk Differences (Single-Week 20% Reduction)

Note: Risk difference (RD) in cumulative preterm birth risk for a 20% reduction applied in a single gestational week (columns) and evaluated through follow-up weeks 28–36 (rows). Marginal single-week interventions; not directly comparable to simultaneous full-history interventions.

---

📁 Data Availability

Input Data Sources

1. Birth Records: Chilean Ministry of Health (DEIS) vital statistics (2010–2020)

   • Location: 01_Data/Input/Nacimientos/
   • Variables: Gestational age, birth weight, parental characteristics, municipality of residence

2. Air Pollution Data: National air-quality monitoring network

   • Location: 01_Data/Input/Clime_series/
   • Pollutants: PM₂.₅ (beta-attenuation), NO₂ (chemiluminescence), O₃ (UV photometry)
   • Interpolation: Ordinary kriging (primary) and IDW (sensitivity)
   • Spatial unit: Municipality centroids (34 comunas)

3. Temperature Data: CR2MET gridded climate product

   • Processed in: 00_Code/4.0 Climate_data_generate.R
   • Variable: Daily mean ambient temperature (TAD)

4. NDVI: MODIS MOD13Q1 (250 m, 16-day composite)

   • Extraction: 00_Code/3.0 NDVI_EarthEngine_commune_extraction.py
   • Gaps imputed with Kalman smoother

5. Socioeconomic Vulnerability Index (SOVI)

   • Location: 01_Data/Input/SOVI/
   • Categories: Low, medium-low, medium-high

6. Municipal Boundaries

   • Location: 01_Data/Input/district_geo/

Processed Datasets

Main analytical datasets are stored in 01_Data/Output/:

• births_2010_2020.RData — Cleaned birth records
• Contamination_Climate_Data_2010_2020.RData — Merged pollution and climate series
• births_2010_2020_exposure_weeks.RData — Weekly gestational exposure histories
• births_2010_2020_exposure_weeks_lagged.RData — Weekly data with DLM lag terms

G-computation outputs are stored in 02_Output/G-Form/:

• Summary_results/ — Point estimates and bootstrap CIs by scenario
• Interventions/ — Counterfactual exposure histories (RDS)
• WeeklyEffects/, PopulationEffects/ — Detailed effect objects

Note: Individual-level birth records cannot be publicly shared due to Chilean data protection regulations. Aggregated results and analysis code are available in this repository.

---

💻 Reproducibility

System Requirements

• R ≥ 4.0.0
• Python 3 (for NDVI extraction via Google Earth Engine)
• Recommended: ≥ 16 GB RAM; Linux server for parallel g-computation

Required R Packages

Automatically installed via 00_Code/0.2 Packages.R and 00_Code/0.2 Packages_gform.R:

• Data manipulation: tidyverse, data.table, janitor, rio
• Spatial analysis: chilemapas, sf, rnaturalearth
• Survival analysis: survival, flexsurv, survminer
• Distributed lag / splines: dlnm, splines, mgcv
• Parallel computing: future, furrr, doParallel
• Visualization: ggplot2, patchwork, ggpubr, RColorBrewer
• Imputation: imputeTS, zoo

Running the Analysis

1. Setup:

   source("00_Code/0.1 Settings.R")
   source("00_Code/0.2 Packages.R")
   source("00_Code/0.3 Functions.R")

2. Data processing (run in order):

   source("00_Code/1.0 Pollution_process_data.R")
   source("00_Code/2.0 Births_process_data.R")
   source("00_Code/3.1 Temp_NDVI_data.R")
   source("00_Code/4.0 Climate_data_generate.R")
   source("00_Code/5.0 Exposure_data_births.R")
   source("00_Code/6.0 Join_full_data.R")

3. Descriptive analysis:

   source("00_Code/7.0 Descriptive_births.R")
   source("00_Code/7.1 Descriptive_exposition.R")
   source("00_Code/8.0 Correlation_pollulants.R")

4. Distributed-lag models:

   source("00_Code/9.0 DLM_pollution.R")
   source("00_Code/9.1 DLM_plots.R")

5. G-computation (two stages):

   # Stage 1: build intervention objects (run once)
   source("00_Code/10.0 G-Form_build_interventions.R")
   
   # Stage 2: run models, bootstrap, and heatmaps
   source("00_Code/10.2 G-Form_models.R")
   source("00_Code/10.3 G-Form_plots.R")
   source("00_Code/10.4 G-Form_table.R")

   For server/parallel execution:

   GFORM_EXEC_MODE=server Rscript "00_Code/10.2 G-Form_models.R"

Notes on Computation Time

• Birth data processing (2.0): moderate (depends on raw file size)
• Weekly exposure expansion (5.0, 6.0): several hours (large longitudinal dataset)
• DLM Cox models (9.0): ~20–30 minutes per pollutant/method
• G-computation bootstrap (10.2): several hours to days (250–500 bootstrap replicates; parallelized on server)
• Total pipeline: plan for multi-hour to overnight runs on a modern workstation or Linux server

Detailed methodological notes: 02_Output/Notas_G-Formula_resultados.md

---

📖 Codebook

Birth Variables

• id: Unique birth identifier
• com: Municipality code
• name_com: Municipality name
• weeks: Gestational age at delivery (weeks)
• date_nac: Date of birth
• sex: Infant sex (Boy/Girl)
• tbw: Birth weight (grams)
• birth_preterm: Preterm birth indicator (<37 weeks)
• birth_very_preterm: Very preterm (28–31 weeks)
• birth_moderately_preterm: Moderate preterm (32–33 weeks)
• birth_late_preterm: Late preterm (34–36 weeks)

Parental and Context Variables

• age_group_mom, educ_group_mom, job_group_mom: Maternal age, education, employment
• age_group_dad, educ_group_dad, job_group_dad: Paternal age, education, employment
• month_week1, year_week1: Month and year of last menstrual period
• covid: COVID-19 period indicator
• vulnerability: SOVI category (Low, Medium-low, Medium-high)

Exposure Variables

• pm25_krg, no2_krg, o3_krg: Weekly kriging-interpolated concentrations
• pm25_idw, no2_idw, o3_idw: Weekly IDW-interpolated concentrations (sensitivity)
• tad: Weekly mean ambient temperature
• ndvi_full: Municipality-level NDVI (full pregnancy average)
• Lag term (Liw): Time-weighted cumulative lag through prior gestational weeks

---

🔬 Methods Detail

Distributed-Lag Exposure

For gestational week (w \geq 2):

[ L_{iw} = \sum_{s=1}^{w-1} \frac{X_{is}}{w - s} ]

Exclusion Criteria

Births were excluded if:

• Outside urban Metropolitan Santiago (33 comunas + Puente Alto)
• Missing date of birth, gestational age, or municipality
• Maternal age <12 or >50 years
• Gestational age <28 weeks
• Multiple births
• Missing covariates
• Implausible birthweight-for-gestational-age (Alexander et al., 1996)
• Fixed-cohort bias: gestational window not fully observed within 2010–2020

G-Computation Interventions

Threshold (cap):

[ X'{iw} = \min(X{iw}, c) ]

Proportional reduction (20%):

[ X'{iw} = X{iw} \times (1 - 0.20) ]

Single-week reduction: Apply 20% reduction only in week (j); all other weeks remain at observed values.

Population metrics at week 36: prevalence, expected cases, risk ratio (RR), risk difference (RD), attributable risk (AR), and population attributable fraction (PAF).

---

🗄️ Repository Structure

Contamination_PTB_G-Formula/
├── 00_Code/                        # Analysis scripts
│   ├── 0.1–0.3                     # Settings, packages, functions
│   ├── 1.0–8.0                     # Data processing and descriptives
│   ├── 9.0–9.1                     # Distributed-lag Cox models
│   ├── 10.0–10.4                   # G-computation pipeline
│   └── old_code/                   # Archived scripts
├── 01_Data/
│   ├── Input/                      # Raw data (not publicly available)
│   └── Output/                     # Processed analytical datasets
├── 02_Output/
│   ├── Descriptives/               # Tables and descriptive plots
│   ├── Models/                     # DLM results and figures
│   ├── G-Form/                     # G-computation outputs
│   └── idw_vs_kriging/             # Interpolation comparison
├── 03_Paper/                       # Manuscript and supplementary material
│   ├── CTA_PTB_Gf_Manuscript.docx
│   └── CTA_PTB_Gf_Supplementary_Material.docx
├── 04_Conference/                  # Conference abstracts
└── README.md

---

⚠️ Important Notes

Data Privacy

Individual-level birth records are confidential and cannot be shared publicly. Researchers interested in data access should contact the Chilean Ministry of Health (DEIS).

Air Quality and Climate Data

• National air-quality network: Chilean Ministry of Environment
• CR2MET: Center for Climate and Resilience Research (CR²)

Citation

If you use this code or methodology, please cite:

Blanco, E., Conejeros, J.D., Bravo, I., Cornejo, F., Osses, A., & Benmarhnia, T. Estimating the perinatal health benefits of hypothetical pollution interventions in Santiago, Chile using parametric g-computation. Under Review. 2025.

---

📧 Contact

For questions about the code or methodology:

• Estela Blanco: estela.blanco@uc.cl
• José Daniel Conejeros: jdconejeros@uc.cl

For data access inquiries:

• Chilean Ministry of Health: https://www.minsal.cl

---

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

🤝 Acknowledgments

This research was supported by FONDECYT de Iniciación en Investigación Nº 11240322 and the Center for Climate and Resilience Research (CR²), FONDAP/ANID 1523A0002. We thank the Chilean Ministry of Health (DEIS) for access to birth records, the national air-quality monitoring network for pollution data, and CR² for climate data.

Data sources:

• Birth records: DEIS, Chilean Ministry of Health
• Air pollution: National air-quality monitoring network (SINCA)
• Temperature: CR2MET v2.5, Center for Climate and Resilience Research, Universidad de Chile
• NDVI: MODIS MOD13Q1 via Google Earth Engine