Getting Started

This section will help you get up and running with deep-inference.

Getting Started

• Installation
  • Requirements
  • Install from PyPI
  • Install from Source
  • Optional Dependencies
  • Verify Installation
  • Dependencies
• Quickstart
  • Basic Workflow
  • Configuration Options
  • Comparing Naive vs Debiased
  • Supported Families
  • New inference() API
  • Next Steps
• Usage Guide
  • The Fundamental Workflow
  • Step 1: Specify Your Economic Model
  • Step 2: Define Your Target
  • Step 3: Fit the Model
  • Step 4: Validate the Results
  • Step 5: Interpret and Use
  • Complete Example
  • Common Mistakes to Avoid
  • When to Use What
  • The Mindset

Overview

deep-inference provides valid statistical inference for neural network estimators. The core workflow is:

1. Generate or load data with covariates \(X\), treatment \(T\), and outcome \(Y\)

2. Select a statistical family that matches your outcome distribution

3. Run inference using the influence function method

4. Interpret results with valid confidence intervals

Why Influence Functions?

Neural networks are powerful function approximators, but naive inference (just averaging predictions) severely underestimates uncertainty. The influence function approach:

• Corrects for regularization bias

• Provides Neyman-orthogonal scores

• Yields valid confidence intervals with proper coverage

Supported Models

Family	Use Case	Example
Linear	Continuous outcomes	Wages, test scores
Logit	Binary outcomes	Purchase decisions
Poisson	Count data	Patent counts
Tobit	Censored data	Labor supply
NegBin	Overdispersed counts	Doctor visits
Multinomial Logit	Discrete choice (J>=3)	Transportation mode