Double ML: Causal Inference based on ML

.title[
# Double ML: Causal Inference based on ML
]
.subtitle[
## Part II: Double Machine Learning in Practice
]
.author[
### uai2022, August 1, 2022, Eindhoven
]
.date[
### Philipp Bach, Martin Spindler (UHH & Economic AI)
Collaborators: Victor Chernozhukov (MIT), Malte Kurz (TUM)
]

---

# Introduction to DoubleML

---

## Overview

* `DoubleML` provides a general implementation of the Double Machine Learning approach by Chernozhukov et al. (2018) in Python and R

* There are also other open source libraries available for causal machine learning
    + `CausalML` (uber, https://github.com/uber/causalml, Chen et al., 2020) - variety of causal ML learners, i.a. with focus on uplift modeling, CATEs and IATEs
    
    + `EconML` (microsoft research, https://github.com/microsoft/EconML, Battocchi et al., 2021) - various causal estimators based on machine learning, among others based on double machine learning approach
    
  + ...

`CausalML` and `EconML` have a focus on heterogeneity of treatment effects from their start on

`DoubleML` focuses on implementing the DML approach and its extensions (example: heterogeneity)

`$\rightarrow$` Object-orientated implementation based on orthogonal score

`$\rightarrow$` Extendibility and flexibility

---

## Building Principles

### Key ingredient and Implementation

- **Orthogonal Score**
  - Object-oriented implementation
  - Exploit common structure being centered around a (linear) score function `$\psi(\cdot)$`

- **High-quality ML**
  - State-of-the-art ML prediction and tuning methods
  - Provided by `scikit-learn` and `sckit-learn`-like learners

- **Sample Splitting**
  - General implementation of sample splitting

---

## DoubleML Package Dependencies

`scikit-learn`

`pandas`

`numpy`

]

`SciPy`

`statsmodels`

`joblib`

]

---

## Why an Object-Orientated Implementation?

*  Given the components `$\psi^a(\cdot)$` & `$\psi^b(\cdot)$` of a linear Neyman orthogonal score function `$\psi(\cdot)$`, a **general implementation** is possible for

- The estimation of the **orthogonal parameters**
  - The computation of the **score** `$\psi(W; \theta, \eta)$`
  - The estimation of **standard errors**
  - The computation of **confidence intervals**
  - A **multiplier bootstrap** procedure for simultaneous inference

*  The **sample splitting** can be implemented in general as well

`$\quad$` `$\rightarrow$` Implemented in the **abstract base class** `DoubleML`

* The **score components** and the estimation of the **nuisance models** have to be implemented **model-specifically**

`$\quad$` `$\rightarrow$` Implemented in **model-specific classes** inherited from `DoubleML`

---

## Class Structure and Causal Models

---

## Advantages of the Object-Orientation

* `DoubleML` gives the user a **high flexibility** with regard to the specification of DML models:

- Choice of ML methods for approximating the nuisance functions
  - Different resampling schemes (repeated cross-fitting)
  - DML algorithms DML1 and DML2
  - Different Neyman orthogonal score functions

* `DoubleML` can be **easily extended**

- New model classes with appropriate Neyman orthogonal score function can be inherited from `DoubleML`
  - The package features `callables` as score functions which makes it easy to extend existing model classes
  - The resampling schemes are customizable in a flexible way

---

# Getting started with DoubleML!

---

## Installation

Install the latest release via pip or conda, see [**installation guide**](https://docs.doubleml.org/stable/intro/install.html)

```python
pip install -U DoubleML
```

```python
conda install -c conda-forge doubleml
```

Install development version from GitHub **https://github.com/DoubleML/doubleml-for-py**

See the [**Getting Started**](https://docs.doubleml.org/tutorial/stable/getstarted_py/) page of the tutorial website for more information on prerequisites.

---

## Data Example: Demand Estimation

#### Data Source

* Data example based on a [**blogpost by Lars Roemheld (Roemheld, 2021)**](https://towardsdatascience.com/causal-inference-example-elasticity-de4a3e2e621b)

* Original real data set publicly available via [**kaggle**](https://www.kaggle.com/vijayuv/onlineretail), [**preprocessing notebook available online**](https://github.com/DoubleML/doubleml-docs/blob/master/doc/examples/py_elasticity_preprocessing.ipynb)

]

#### Causal Problem

* **Price elasticity of demand:** What is the **effect** of a **price change**, `$dLnP$`, on **demanded quantity**, `$dLnQ$`?

* **Observational study**: Flexibly adjust for confounding variables `$X$`, e.g. product characteristics

#### Causal Diagram (DAG)

]

---

# Hands On Notebook

---

## Data Example: A/B Testing

#### Data Source

* Data example based on a randomly chosen DGP created for the [**2019 ACIC Data Challenge**](https://sites.google.com/view/acic2019datachallenge/data-challenge).

]

#### Causal Problem

* **Online shop:** What is the **effect** of a **new ad design** `$A$` on **sales** `$Y$` (in $100 )?

* **Observational study**: Necessary to adjust for confounding variables `$V$`

#### Causal Diagram (DAG)

]

---

## Online Resources

* The notebook is organized according to the [**DoubleML Workflow**](https://docs.doubleml.org/stable/workflow/workflow.html)

* Extensive [**User Guide**](https://docs.doubleml.org/stable/guide/guide.html) available via [**docs.doubleml.org**](https://docs.doubleml.org)

* [**Documentation for the Python API**](https://docs.doubleml.org/stable/api/api.html) available via [**https://docs.doubleml.org/stable/api/api.html**](https://docs.doubleml.org/stable/api/api.html)

* Paper for the Python package available from [**JMLR**](http://jmlr.org/papers/v23/21-0862.html) or [**arxiv**](https://arxiv.org/abs/2104.03220)

---

# References

---

## References

#### Open-Source Libraries for Causal Machine Learning

* Battocchi, K, Dillon, E., Hei, M., Lewis, G., Oka, P., Oprescu, M. and Syrgkanis, V. (2021), EconML: A Python package for ML-based heterogeneous treatment effects estimation, https://github.com/microsoft/EconML. Version 0.11.1

* Chen, H., Harinen, T., Lee, J.-Y., Yung, M. and Zhao, Z. (2020), CausalML: Python package for
causal machine learning, [**arXiv:2002.11631**](https://arxiv.org/abs/2002.11631) [cs.CY]

#### DoubleML Package for Python and R

* Bach, P., Chernozhukov, V., Kurz, M. S., and Spindler, M. (2021), DoubleML - An Object-Oriented Implementation of Double Machine Learning in R, [arXiv:2103.09603](https://arxiv.org/abs/2103.09603).

* Bach, P., Chernozhukov, V., Kurz, M. S., and Spindler, M. (2022), DoubleML - An Object-Oriented Implementation of Double Machine Learning in Python, Journal of Machine Learning Research, 23(53): 1-6, https://www.jmlr.org/papers/v23/21-0862.html.

---

## References

#### Double Machine Learning Approach

* Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W. and Robins, J. (2018), Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 21: C1-C68, doi:10.1111/ectj.12097.

* Chernozhukov, V., Hansen, C., Kallus, N., Spindler, M., and Syrgkanis, V. (forthcoming), Applied Causal Inference Powered by ML and AI.

---

# Appendix

---

## DoubleML Class Structure