deep-inference
Deep Learning for Individual Heterogeneity
References:
- Farrell, Liang, Misra (2021) "Deep Neural Networks for Estimation and Inference" Econometrica
- Farrell, Liang, Misra (2025) "Deep Learning for Individual Heterogeneity"
deep-inference is a Python package for enriching structural economic models with deep learning. It implements the framework developed by Farrell, Liang, and Misra (2021, 2025) to recover rich, non-linear parameter heterogeneity ($\theta(X)$) while maintaining the interpretability and validity of structural economics.
Standard deep learning minimizes prediction error, which leads to biased parameter estimates ("The Inference Trap"). This package implements Influence Function-based Debiasing to provide valid confidence intervals and p-values for economic targets.
Why deep-inference?
Economic structure and Machine Learning are complements, not substitutes.
- Deep Learning provides the capacity to learn complex, high-dimensional heterogeneity.
- Structural Models provide the constraints necessary for causal interpretation and counterfactuals.
deep-inference enforces the structural loss (e.g., Tobit likelihood) on the output of a neural network, ensuring that the estimated parameters $\alpha(X), \beta(X)$ respect the economic theory.
Installation
pip install deep-inference
Quickstart
Estimate a linear demand model where Price Elasticity $\beta$ varies non-linearly with customer characteristics $X$.
import numpy as np
from deep_inference import structural_dml
# 1. Generate synthetic data (Linear Demand with Heterogeneity)
np.random.seed(42)
n = 2000
X = np.random.randn(n, 10) # Customer characteristics
T = np.random.randn(n) # Treatment (e.g., price)
beta_true = np.cos(np.pi * X[:, 0]) * (X[:, 1] > 0) + 0.5 * X[:, 2]
Y = X[:, 0] + beta_true * T + np.random.randn(n)
mu_true = beta_true.mean()
# 2. Run Inference
# The package trains the structural network, computes the influence function,
# and aggregates results via cross-fitting.
result = structural_dml(
Y=Y, T=T, X=X,
family='linear',
hidden_dims=[128, 64, 32],
epochs=50,
lr=0.01,
n_folds=50,
)
print(f"True mu* = {mu_true:.4f}")
print(result.summary())
New inference() API
The new inference() API provides additional flexibility:
- Flexible targets: Average Marginal Effect (AME), custom target functions
- Randomization mode: For RCTs, compute Λ directly instead of estimating
- Regime auto-detection: Automatically chooses optimal Lambda strategy
from deep_inference import inference
from deep_inference.lambda_.compute import Normal
# Average Marginal Effect (probability scale, not log-odds)
result = inference(Y, T, X, model='logit', target='ame', t_tilde=0.0)
# Custom target with autodiff Jacobian
import torch
def my_target(x, theta, t_tilde):
return torch.sigmoid(theta[0] + theta[1] * t_tilde)
result = inference(Y, T, X, model='logit', target_fn=my_target)
# Randomized experiment (Regime A)
result = inference(Y, T, X, model='logit', target='beta',
is_randomized=True, treatment_dist=Normal(0, 1))
Three Estimation Regimes
| Regime | Condition | Lambda Method | Cross-Fitting |
|---|---|---|---|
| A | RCT with known treatment distribution | Compute via MC | 2-way |
| B | Linear structural model | Closed-form analytic | 2-way |
| C | Observational, nonlinear | Neural network estimation | 3-way |
Supported Structural Families
The package abstracts the math of influence functions. You simply select the family that matches your outcome variable.
| Family | Model Structure | Use Case |
|---|---|---|
| Linear | $Y = \alpha(X) + \beta(X)T + \varepsilon$ | Wages, Test Scores, Consumption |
| Logit | $P(Y=1) = \sigma(\alpha(X) + \beta(X)T)$ | Binary Choice, Market Entry |
| Poisson | $Y \sim \text{Pois}(\exp(\alpha(X) + \beta(X)T))$ | Patent Counts, Doctor Visits |
| Gamma | $Y \sim \text{Gamma}(k, \exp(\alpha(X) + \beta(X)T))$ | Healthcare Costs, Insurance Claims |
| Gumbel | $Y \sim \text{Gumbel}(\alpha(X) + \beta(X)T, s)$ | Extreme Value Analysis |
| Tobit | $Y = \max(0, \alpha(X) + \beta(X)T + \sigma\varepsilon)$ | Labor Supply, Censored Demand |
| NegBin | $Y \sim \text{NegBin}(\exp(\alpha(X) + \beta(X)T), r)$ | Count Data with Overdispersion |
| Weibull | $Y \sim \text{Weibull}(k, \exp(\alpha(X) + \beta(X)T))$ | Duration Analysis, Survival |
| Multinomial Logit | $P(Y=j) = \text{softmax}(\alpha_j(X) + X'_j \beta(X))$ | Transportation, Brand Choice |
Note: For complex models like Tobit, deep-inference automatically handles the joint estimation of structural variance $\sigma(X)$ required for consistent inference.
Methodological Details
1. The Enriched Model
We replace fixed parameters $\theta$ with neural networks $\theta(X)$:
\hat{\theta}(\cdot) = \arg \min_{\theta \in \mathcal{F}_{DNN}} \sum \ell(y_i, t_i, \theta(x_i))
2. The Influence Function Correction
Naive averaging of $\hat{\theta}(X)$ yields biased inference. We construct a Neyman-Orthogonal score $\psi$ using the Influence Function:
\psi(z) = H(\hat{\theta}) + \nabla_\theta H \cdot \Lambda(x)^{-1} \cdot \nabla_\theta \ell(z, \hat{\theta})
Where $\Lambda(x) = \mathbb{E}[\nabla^2 \ell \mid X=x]$ is the conditional Hessian.
- Automatic Differentiation:
deep-inferenceuses PyTorch Autograd to compute exact Jacobians and Hessians for any model family. - Stability: Includes Tikhonov regularization for inverting Hessians in non-linear models (e.g., Logit/Tobit).
Validation
Comprehensive eval suite validates every mathematical component. Full results →
| Eval | Component | Result |
|---|---|---|
| 01 | Parameter Recovery | 12/12 families PASS |
| 02 | Autodiff Accuracy | 31/31 PASS |
| 03 | Lambda Estimation | 9/9 PASS |
| 04 | Target Jacobian | 92/92 PASS |
| 05 | Influence Functions | Coverage 88% |
| 06 | Frequentist Coverage | PASS |
| 07 | End-to-End | 7/7 PASS |
| 09 | Multinomial Logit | 98% coverage PASS |
Citation
@article{farrell2021deep,
title={Deep Neural Networks for Estimation and Inference},
author={Farrell, Max H. and Liang, Tengyuan and Misra, Sanjog},
journal={Econometrica},
volume={89},
number={1},
pages={181--213},
year={2021}
}
@article{farrell2025heterogeneity,
title={Deep Learning for Individual Heterogeneity},
author={Farrell, Max H. and Liang, Tengyuan and Misra, Sanjog},
journal={Working Paper},
year={2025}
}
License
MIT
Metadata
Release files for deep-inference 0.2.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| deep_inference-0.2.0.tar.gz | 82.9 kB | Without provenance | Details for deep_inference-0.2.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| deep_inference-0.2.0-py3-none-any.whl | Python 3 | none | any | Details for deep_inference-0.2.0-py3-none-any.whl |
Total release size: 195.6 kB
Release files / deep_inference-0.2.0.tar.gz
| Download URL | deep_inference-0.2.0.tar.gz |
|---|---|
| Size | 82.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
b11d6eafcb69ad389d44e5e0f636655a39342c1d9a9e21811f1bf34f66503942
|
|
BLAKE2b-256 checksum How to use checksums |
7806ccdf69ea5befb59a5f80b4ac16aa1e2b08c2e67915295b6e976b85adb88e
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.14
|
Release files / deep_inference-0.2.0-py3-none-any.whl
| Download URL | deep_inference-0.2.0-py3-none-any.whl |
|---|---|
| Size | 112.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
eb1276d3b577cbae96939411a8e8243ec37dd4cc9e5550474c9d6850bc834055
|
|
BLAKE2b-256 checksum How to use checksums |
9c655b8b085f00b7f2c26d1118ac8014087378148ace83a5e97384406941cfa7
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.11.14
|
Security for deep-inference 0.2.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.