Statistical Models for Disaster Relief Classification

The Problem

After natural disasters, relief agencies must quickly identify survivors most likely to need immediate assistance. Shelter conditions (type, damage level, access to resources) are among the first observable factors. Can statistical models predict which individuals are at highest risk based on shelter characteristics alone?

What I Built

A comparative analysis of multiple statistical classification models predicting survivor outcomes based on shelter conditions. The project emphasizes proper model selection methodology and interpretability over raw accuracy.

Models Compared

• Logistic Regression: Baseline linear classifier with coefficient interpretation
• Linear Discriminant Analysis (LDA): Assumes normally distributed features, finds optimal linear boundary
• Quadratic Discriminant Analysis (QDA): Relaxes LDA's equal covariance assumption
• Decision Trees: Non-parametric, captures feature interactions
• Random Forest: Ensemble method reducing variance of individual trees

Methodology

• Train/test split with cross-validation for honest performance estimation
• ROC curves and AUC for threshold-independent model comparison
• Confusion matrix analysis highlighting false negative rates (missing at-risk survivors)
• Feature importance ranking to identify which shelter characteristics matter most

Key Takeaway

In disaster relief contexts, false negatives (classifying at-risk people as safe) carry far greater cost than false positives. The analysis weights model selection toward recall over precision, which changes which model is "best" compared to accuracy-only evaluation.

Links

Full Analysis (R Markdown) | GitHub Repository