Statistical_Analysis_and_Predictive_Modeling

This project implements a modular data pipeline and a predictive model to analyze the Pima Indians Diabetes dataset. The system is designed following software engineering principles: decoupled, parameterized, featuring virtual environment management (renv), automated logging, unit testing, and an inference script.

📁 Project Structure

├── data/                        # Preprocessed datasets (Git ignored)
│   ├── .gitkeep
│   └── pima_diabetes_clean.csv  # CSV with zeros converted to NA
├── scripts/                     # Pipeline scripts (sequential execution)
│   ├── utils.R                  # Shared functions (VIF, AUC, metrics, Brier, HL)
│   ├── 01_data_ingestion.R      # Download, validation, zero → NA cleaning
│   ├── 02_exploratory_analysis.R# EDA: statistics, NAs, class balance, boxplots
│   ├── 03_predictive_modeling.R # MI PMM → GLM pooled → VIF → metrics → CV → LASSO → ROC
│   └── 04_predict.R             # (Optional) Inference on new data
├── tests/                       # Automated unit tests (testthat)
│   └── test_pipeline.R          # 5 test blocks, 31 assertions
├── results/                     # Generated artifacts (Git ignored)
│   ├── logs/
│   │   ├── eda_processing.log
│   │   ├── model_training.log
│   │   └── pipeline_qa.log
│   ├── plots/
│   │   ├── boxplot_glucosa_vs_outcome.png
│   │   └── roc_curve.png        # Includes optimal Youden threshold marker
│   ├── modelo_logistico.rds     # List of 5 GLM models (multiple imputation)
│   └── medianas_train.rds       # Training set medians
├── main.R                       # Main orchestrator + global CONFIG
├── renv.lock                    # Library version lockfile
└── README.md

Installation and Execution

To guarantee scientific reproducibility, this project uses renv to isolate the required libraries.

1. Clone the repository to your local machine:

   git clone https://github.com/CGARCOR88/Analisis_estadistico_y_modelado_predictivo.git
   cd Analisis_estadistico_y_modelado_predictivo

2.Open the project in your R environment (VS Code or RStudio).

3. Restore the virtual environment by running the following in the R console

   renv::restore()

4. Run the complete pipeline from the root orchestrator:

   source("main.R")

Pipeline Design

Script	Responsabilidad
utils.R	Shared functions: calc_vif, calc_auc, calc_metricas, calc_brier_score, hosmer_lemeshow_test
01_data_ingestion.R	Downloads CSV, validates dimensions and columns, converts physiological zeros to NA
02_exploratory_analysis.R	EDescriptive statistics, missing values analysis, class balance, boxplots
03_predictive_modeling.R	Stratified 70/30 split → MI PMM (m=5) → Rubin's Pooling → VIF → Youden → Brier/HL → Cook's D → 10-fold CV → LASSO → ROC
04_predict.R	Loads model list + medians, preprocesses new data, averages predictions across the m models

Design Decisions:

• True Multiple Imputation (m=5): 5 GLM models are trained (one per imputation). Coeficients are combined using Rubin's Rules to obtain pooled estimates with correct uncertainty propagation.
• **Averaged Predictions:**In testing and production, probabilities from the 5 models are averaged (Rubin's approach for predictions), reducing imputation variance.
• No Data Leakage: mice only operates on the training set. The test set and production data are imputed using medians calculated exclusively from the training set.
• Centralized Parameters in CONFIG: proporcion_train, umbral_clasificacion, m_imputaciones and k_folds are defined in main.R and are globally accessible by all scripts.
• Logs are managed via sink() + tryCatch/finally to guarantee that connections close properly even if an error occurs.

Model Results

GLM with PMM Multiple Imputation (m=5, Rubin's Rules) — Pima Indians Diabetes dataset, stratified 70/30 split:

Test Set Metrics (Default Threshold vs. Optimal Youden Threshold)

Metric	Threshold 0.50	Optimal Threshold 0.28
Accuracy	74.03%	73.59%
Sensitivity (Recall)	50.62%	79.01%
Specificity	86.67%	70.67%
Positive Predictive Value (PPV)	67.21%	59.26%
F1-Score	0.5775	0.6772
AUC-ROC	0.8244	—

The optimal Youden threshold (0.28) improves Sensitivity from 50.6% to 79.0% and the F1-Score from 0.58 to 0.68, at the cost of lower Specificity. For diabetes screening, maximizing Sensitivity is the priority.

Robustness and Calibration

Análysis	Result
AUC-ROC CV (k=10)	0.8413 ± 0.0532
Brier Score	0.1650 (referencia nula: 0.2277)
Hosmer-Lemeshow χ²(df=8)	12.08, p = 0.1475 ✅ Buen ajuste
Influential Observations (Cook)	36 / 537 (6.7%)

Comparación con LASSO (glmnet, α=1)

Model	AUC-ROC	Features
GLM (MI m=5)	0.8244	8 predictores
LASSO (λ 1SE)	0.8140	5 predictors (Pregnancies, Glucose, Insulin, BMI, Age)

LASSO automatically drops BloodPressure, SkinThickness, and DiabetesPedigreeFunction, reducing model complexity without any significant drop in AUC.

Multicollinearity (VIF) and Significance

VIF: All predictors < 2 (max. 1.958 for BMI). No problematic multicollinearity detected.

Significant variables in pooled coefficients (p < 0.05): Glucose (***), BMI (***), Age (**), Pregnancies (*), DiabetesPedigreeFunction (*).

The detailed execution history is saved in results/logs/model_training.log.

Quality Control (QA)

At the end of the pipeline, 31 automated assertions are executed via testthat, distributed across 5 blocks:

1. Preprocessed Dataset:: correct dimensions (768×9), binary outcome, zeros removed, NAs present.
2. Trained Model: confirms it is a list of 5 binomial GLM models, each with 9 coefficients.
3. Training Medians: all critical variables present, values are positive
4. Visual Artifacts and Logs: both logs are non-empty, both PNG plots are generated.
5. Log Content: checks for the presence of "Accuracy", "Sensibilidad", "AUC-ROC", "VIF", "Brier Score", "Hosmer-Lemeshow", "Youden", and "LASSO".

Output: [ FAIL 0 | WARN 0 | SKIP 0 | PASS 31 ]

The QA report is stored in results/logs/pipeline_qa.log.

Inference on New Data

To generate predictions on a new dataset, you can invoke the script from your interactive environment or directly from the system terminal:

# Desde la terminal usando Rscript
Rscript scripts/04_predict.R

# O dentro de R
source("scripts/04_predict.R")

he script loads the list of 5 models (results/modelo_logistico.rds) and the training medians (results/medianas_train.rds), applies the same preprocessing steps (zeros → NA, training median imputation), averages the probabilities across the 5 models, and saves the predictions to results/predicciones.csv.

The input file path can be configured via CONFIG$archivo_nuevos_datos inside main.R.

Visual Analysis

The pipeline automatically generates plots in results/plots/:

• boxplot_glucosa_vs_outcome.png — Plasma glucose distribution grouped by diagnosis outcome.
• roc_curve.png — ROC curve for the GLM model (m=5), complete with annotated AUC and the optimal Youden threshold marker.

Note: If you have just cloned the repository, the results/ folder will be created automatically after running source("main.R") for the first time.