Orthogonalization

As explained in Causal Inference for Data Science

Orthogonalization is a causal estimation technique that reduces the influence of nuisance-model estimation errors on a target parameter estimate.

Orthogonalization reduces the impact of nuisance-model errors on a causal parameter by enhancing existing estimators with machine learning. It expresses the causal effect as the solution to an equation defined by a score function, analogous to a loss function. DML is one important case of this broader technique. section 8.1.3 section 8.1.3

The practical picture has three parts:

• A target causal parameter is estimated through a score function specific to the causal model and parameter. section 8.1.3
• Nuisance functions, such as propensity scores, support estimation but are not themselves the primary causal target. section 8.1.3
• DML combines orthogonalization with cross-validation and cross-fitting to address confounding, overfitting, and same-data prediction bias. section 8.1.3

Two-fold cross-fitting applied to the DML algorithm. Once the process is finished, you must switch roles between datasets 1 and 2. This process can be generalized to k-fold cross-fitting by breaking the data into k datasets instead of two.

Two-fold cross-fitting illustrates how datasets switch roles during DML, with the process generalizing to k-fold cross-fitting. section 8.1.3

How Causal Inference for Data Science covers it

How deep it goes

Defines score-based estimation and nuisance functions, but defers intricate technical details.

Examples

Uses PLM with DoubleML and points to IRM, IV, and DiD extensions.

What you can do after

Operate DoubleML by recognizing score and nuisance functions without mastering their technical details.

How it connects

Connects DML, machine learning, score functions, nuisance functions, and cross-fitting.

liveBookmanning.com

Read Causal Inference for Data Science on liveBook Get the book at manning.com

Related concepts

• double machine learningincludes · Double machine learning is a causal-effect estimation method that uses machine-learning nuisance models with orthogonalized estimation to reduce bias from nuisance-function errors.

This explainer was generated by Manning's knowledge graph from the text of Causal Inference for Data Science. The section links point to the parts of the book it draws on.