A Bayesian semiparametric model for estimating time-varying Case Fatality Rate (CFR) with intervention effects and counterfactual inference.

• Bayesian sCFR Model: Separates smooth baseline trends from intervention effects
• Counterfactual Inference: Estimates what would have happened without interventions
• Benchmark Methods: Includes cCFR, aCFR, and frequentist fsCFR for comparison
• Simulation Framework: Monte Carlo study with parallel execution and checkpointing

pip install -r requirements.txt

Requirements: JAX, NumPyro, NumPy, SciPy, pandas, matplotlib, joblib, tqdm, numba

# Demo mode (5 runs per scenario)
python simulation.py --simulate --demo

# Full simulation
python simulation.py --simulate --full

# Analysis only
python simulation.py --analyze

# Reset and restart
python simulation.py --simulate --full --reset

# sCFR only
python real_data_analysis.py

# Include all benchmarks
python real_data_analysis.py --full

├── config.py              # Configuration parameters
├── data_generation.py     # Synthetic data generation
├── methods.py             # CFR estimation methods (sCFR, cCFR, aCFR)
├── fsCFR_python.py        # Frequentist sCFR implementation
├── evaluation.py          # Metrics and visualization
├── simulation.py          # Monte Carlo simulation runner
├── real_data_analysis.py  # UK COVID-19 analysis
└── requirements.txt       # Dependencies

The sCFR model decomposes the logit-CFR as:

logit(r_t) = B(t)α + δ_t + β·Z(t) + γ·H(t)

• B(t)α: B-spline baseline with RW2 penalty
• δ_t: i.i.d. random effects ~ N(0, σ_δ)
• β·Z(t): Step intervention effects
• γ·H(t): Hinge (slope) intervention effects

Priors:

• σ_δ ~ HalfCauchy(0.1)
• |β|, |γ| ~ LogNormal(log(0.5), 0.5)
• Likelihood: Poisson with delay convolution

Simulation outputs are saved to ./simulation_outputs/:

• plots/: Visualization figures
• results_csv/: Aggregated metrics
• posterior_samples_raw/: MCMC samples
• benchmarks_results/: Benchmark estimates

Real data outputs are saved to ./real_data_outputs/.

• Hengtao Zhang: [email protected]
• Yuanke Qu: [email protected]