Tutorials

Step-by-step tutorials for each supported model family.

Tutorials

• Gallery
  • At a Glance
  • 1. Linear: Wage Returns to Experience
  • 2. Logit: Discount Effectiveness
  • 3. Poisson: Open Access Citation Advantage
  • 4. Tobit: Charitable Donations
  • 5. Gamma: Insurance Claims
  • 6. Weibull: Customer Churn
  • 7. Multinomial Logit: Transportation Mode Choice
  • 8. Gaussian: Manufacturing Quality Control
  • Run It Yourself
• Linear Model Tutorial
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Expected Results
  • Real-World Applications
  • Key Takeaways
• Logit Model Tutorial
  • Safe by Default
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Expected Results
  • Alternative Targets
  • Real-World Applications
  • Numerical Considerations
  • Key Takeaways
• Poisson Model Tutorial
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Interpreting Coefficients
  • Real-World Applications
  • Poisson vs Negative Binomial
  • Key Takeaways
• Tobit Model Tutorial
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Alternative Targets
  • Parameter Structure
  • Real-World Applications
  • Handling Different Censoring
  • Key Takeaways
• Negative Binomial Model Tutorial
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Poisson vs Negative Binomial
  • Real-World Applications
  • Overdispersion Parameter
  • Key Takeaways
• Gamma Model Tutorial
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Expected Results
  • Real-World Applications
  • Key Takeaways
• Gumbel Model Tutorial
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Expected Results
  • Real-World Applications
  • Key Takeaways
• Weibull Model Tutorial
  • When to Use
  • Mathematical Setup
  • Complete Example
  • Expected Results
  • Real-World Applications
  • Key Takeaways
• Multinomial Logit (Conditional Logit) Tutorial
  • Training Configuration
  • When to Use
  • Mathematical Setup
  • Data Encoding
  • Complete Example
  • Expected Results
  • Alternative Targets
  • Real-World Applications
  • Key Takeaways
  • Reference
• Multimodal Tutorial: Text & Image Embeddings
  • Why Embeddings?
  • Gallery of Examples
  • Example 1: Linear — Wages with Job Embeddings
  • Example 2: Logit — Purchases with Image Embeddings
  • Example 3: Poisson — Citations with Abstract Embeddings
  • Using Real Embeddings
  • Key Takeaways
  • Run the Full Gallery
• Personalized Pricing Tutorial
  • Motivation
  • When to Use
  • Mathematical Setup
  • Example: Heterogeneous Price Sensitivity
  • Optimal Pricing
  • Uniform vs. Personalized Pricing
  • References
• Continuous Treatment Tutorial
  • Motivation
  • When to Use
  • Connection to Colangelo & Lee (2026)
  • Example: Advertising and Sales
  • Multiple Evaluation Points
  • Supported Model Families
  • References

Overview

Each tutorial covers:

1. When to use - What types of data and research questions

2. Mathematical setup - The DGP, loss function, and estimand

3. Code example - Complete working code

4. Results interpretation - Understanding coverage and SE ratios

5. Real-world applications - Practical use cases

Model Selection Guide

Your Outcome	Distribution	Model	Tutorial
Continuous (unbounded)	Normal	Linear	Linear
Binary (0/1)	Bernoulli	Logit	Logit
Count (0, 1, 2, …)	Poisson	Poisson	Poisson
Censored at boundary	Tobit	Tobit	Tobit
Overdispersed counts	Negative Binomial	NegBin	NegBin
Positive right-skewed	Gamma	Gamma	Gamma
Extreme values (maxima)	Gumbel	Gumbel	Gumbel
Time-to-event	Weibull	Weibull	Weibull
Discrete choice (J>=3)	Categorical	Multinomial Logit	Multinomial
High-dim embeddings	Any	Any	Multimodal
Pricing / Elasticity	Bernoulli	Logit	Pricing
Dose-response (continuous T)	Any	Any	Continuous Treatment

Common Workflow

All tutorials follow the same basic pattern:

from deep_inference import structural_dml
import numpy as np

# 1. Prepare your data
# Y: outcome variable
# T: treatment variable
# X: covariates (must be 2D array)

# 2. Run inference
result = structural_dml(
    Y=Y, T=T, X=X,
    family='linear',  # or 'logit', 'poisson', etc.
    hidden_dims=[64, 32],
    epochs=100,
    n_folds=50
)

# 3. Check results
print(f"Estimate: {result.mu_hat:.4f}")
print(f"SE: {result.se:.4f}")
print(f"95% CI: [{result.ci_lower:.4f}, {result.ci_upper:.4f}]")

# Coverage should be ~95%, SE ratio should be ~1.0