Package: mixedsubjects 1.0.0

Overview | Simulation Engine | Scenario 1: Poor Predictions — DiM Wins | Scenario 2: Negatively Correlated Predictions | Scenario 3: Heterogeneous Prediction Quality — Double-Tuned Wins | Scenario 4: High Common-Mode Error | Scenario 5: Common-Mode Error + Heterogeneous Quality | Scenario 6: Near-Perfect Predictions | Summary Table

Overview | Why Mixed-Subjects Designs? | What This Package Does | Installation | A Quick Example | Simulating Experimental Data | Creating an MSD Data Object | Estimating the Treatment Effect | Understanding the Data Structure | What Data Do You Need? | Two Types of Predictions | Creating Your Data Object | Flexible Column Names | Formula Interface | The Seven Estimators | 1. Difference-in-Means (DiM) | 2. GREG (Calibration Estimator) | 3. PPI++ (Power-Tuned) | 4. D-T (Doubly-Tuned) | 5. DiP (Difference-in-Predictions) | 6. DiP++ (Power-Tuned DiP) | 7. D-T DiP (Doubly-Tuned DiP) | Comparing All Estimators | Choosing the Right Estimator | Decision Tree | When to Use DiP-Type Estimators | Variance Estimation | Delta-Method (Default) | Bootstrap (Optional) | Optimal Experimental Design | The Design Problem | Using optimal_design() | Interpreting the Results | Comparing Designs | Practical Workflow | Step 1: Pilot Study | Step 2: Estimate Prediction Quality | Step 3: Plan Your Main Study | Step 4: Run the Main Study | Step 5: Analyze and Report | Common Questions | Q: What if my LLM predictions are biased? | Q: How much do predictions need to correlate with outcomes? | Q: Should I use cross-fitting? | Q: What if predictions are worse in one treatment arm? | Q: How many folds should I use for cross-fitting? | Q: Can I use predictions from multiple LLMs? | Technical Details | Assumptions | Variance Formulas | Citation | Session Info