High-dimensional propensity score (HDPS)

Automated covariate selection for observational studies using the High-Dimensional Propensity Score (HDPS) algorithm. Features modular workflow, parallel processing, interactive visualizations, and comprehensive documentation.

✨ Key Features

• 🚀 Modular workflow: 3-step process (identify → assess → prioritize)
• ⚡ High Performance: 20-30x speed improvement with optimized algorithms
• 🔄 Parallel Processing: Multi-core support with batch processing for large datasets
• 💾 Memory Efficient: Minimal memory overhead with smart data handling
• 📊 Visualizations: Interactive plots for bias analysis
• 🔄 Data flexibility: Multiple input formats (long, wide, matrix)
• 📱 Interactive app: Shiny interface for covariate selection
• 📚 Documentation: Comprehensive vignettes and examples

📦 Installation

From GitHub (Development Version)

# Install from GitHub
devtools::install_github("Cainefm/hdps")

From Local Package File

# Install from .tar.gz file
install.packages("hdps_0.9.4.tar.gz", repos = NULL, type = "source")

Dependencies

Required: data.table, pbapply
Suggested: ggplot2, shiny, DT, testthat, knitr

🚀 Quick Start

1-Step Workflow

library(hdps)

# Load your data
data(dx)  # Example diagnosis data
data(master)  # Example master table with exposure/outcome

# Complete HDPS workflow in one function
results <- hdps(
  data = dx,
  id_col = "pid",
  code_col = "icd9code", 
  exposure_col = "exposure",
  outcome_col = "outcome",
  master_data = master,  # Pass master dataset separately
  n_candidates = 200,
  min_patients = 10  # Minimum patients required per covariate
)

# View results
head(results$prioritization)

Sample Output:

                   code    e1    e0    d1    d0    c1    c0  e1c1  e0c1  e1c0  e0c0  d1c1  d0c1  d1c0  d0c0   pc1   pc0      rrCE       rrCD       bias absLogBias ce_strength cd_strength
                 <char> <int> <int> <int> <int> <int> <int> <int> <int> <int> <int> <num> <num> <num> <num> <num> <num>     <num>      <num>      <num>      <num>       <num>       <num>
 1: cov_cov_250.00_freq   295   705   139   861    34   966    15    19   280   686   4.0  30.0 135.0 831.0 0.034 0.966 1.5220588 0.84183007  1.1740003  0.1604169   0.5220588  0.15816993
 2: cov_cov_250.00_once   295   705   139   861    33   967    18    15   277   690   5.0  28.0 134.0 833.0 0.033 0.967 1.9041680 1.09339665  0.9199933  0.0833889   0.9041680  0.09339665
 3: cov_cov_250.00_spor   295   705   139   861    34   966    17    17   278   688   0.1  34.1 139.1 827.1 0.034 0.966 1.7374101 0.02042542 17.9913866  2.8898931   0.7374101  0.97957458

The prioritization data.table shows covariates ranked by bias, with columns for bias estimates, prevalence, and strength measures.

3-Step Workflow

# Step 1: Identify candidate covariates
candidates <- identify_candidates(dx, "pid", "icd9code", "dx", n = 200, min_patients = 10)

# Step 2: Assess recurrence patterns
recurrence <- assess_recurrence(candidates$data, "pid", "code", "dx")

# Step 3: Prioritize covariates
cohort_data <- merge(recurrence, master, by = "pid", all.x = TRUE)
prioritization <- prioritize(cohort_data, "pid", "exposure", "outcome")

Advanced Options: Parallel Processing

# Use parallel processing for large datasets
results <- hdps(data = dx, 
                id_col = "pid", 
                code_col = "code", 
                exposure_col = "exposure", 
                outcome_col = "outcome",
                n_cores = 4,        # Number of CPU cores
                batch_size = 50,    # Batch size for parallel processing
                progress = TRUE)     # Show progress bar

Performance Parameters:

• n_cores: Number of CPU cores for parallel processing (NULL for auto-detection)
• batch_size: Batch size for parallel processing (default: 50)
• progress: Show progress bar during processing (default: TRUE)

Interactive Analysis

# Launch interactive Shiny app
hdps_interactive()

Interactive Features: The Shiny app provides real-time covariate selection with interactive plots for bias distribution, covariate strength relationships, and bias vs prevalence analysis.

📊 Visualizations

Bias Distribution Plot

# Plot top covariates by bias
plot_bias_distribution(prioritization, top_n = 20)

Covariate Strength Relationships

# Plot CE vs CD strength relationships
plot_covariate_strength(prioritization)

Bias vs Prevalence

# Create additional visualizations
plot_bias_vs_prevalence(prioritization)

📚 Documentation

• Tutorial: vignette("hdps-tutorial")
• Function Help: help(hdps)
• Interactive App: hdps_interactive()

🧪 Testing

library(testthat)
test_package("hdps")

📖 Citation

citation("hdps")

🤝 Contributing

Contributions welcome! See Contributing Guidelines.

🙏 Acknowledgments

• HDPS algorithm by Schneeweiss et al. (2009)