T-learner

As explained in Causal Inference for Data Science

A T-learner is a causal estimation algorithm that fits separate outcome models for each treatment value and compares their predictions.

A T-learner estimates causal effects by fitting two separate outcome models: one for treatment X = 1 and another for X = 0. Applying both models to the same covariates produces two potential-outcome predictions; their difference estimates each observation’s treatment effect, and averaging those differences estimates the ATE. section 4.2.4 section 5.4.1

A practical T-learner workflow is:

• Split historical observations by treatment value and train one outcome model on each group, using the available confounders. section 4.2.4
• For every observation, predict the outcome twice—once under X = 1 and once under X = 0—and subtract the predictions. section 5.4.1
• Average the resulting treatment-effect vector to estimate the ATE, while checking whether model fitting and prediction reuse the same data. section 4.2.4 section 4.2.5

Two-fold cross-fitting

Split the data into D1 and D2. Train treatment-specific models on D1, predict both treatment outcomes for D2, and calculate its treatment-effect vector. Reverse the roles of D1 and D2, then combine the estimates. In an exercise, treatment is RHC versus non-RHC and the outcome is death. section 5.4.1 section 4.2.5

Figure 4.7 Two-fold cross-fitting applied to the T-learner. We split the dataset into two and use one part to train two models with X = 0 and X = 1, respectively. In the other part, we make two interventional predictions (do(X = 0) and do(X = 1)) using both previous models and calculate the ATE. Finally, we switch the roles of the two datasets.

Compare the training and prediction halves of two-fold cross-fitting for a T-learner. section 4.2.5

How Causal Inference for Data Science covers it

How deep it goes

Defines separate models and prediction differences, then adds cross-fitting to address same-data overfitting.

Examples

Uses binary treatment predictions, RHC versus non-RHC treatment, death outcomes, two-fold splits, and treatment-effect vectors.

What it assumes

Assumes adjustment-formula reasoning, treatment and outcome variables, confounders, predictive modeling, and average treatment effects.

What you can do after

Builds treatment-specific predictive models and estimates ATEs with cross-fitting while checking for overfitting.

How it connects

Connects the adjustment formula with S-learners, cross-fitting, overfitting, propensity scores, and average treatment effects.

liveBookmanning.com

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

Related concepts

• average treatment effect estimationproduces · Average treatment effect estimation quantifies the population-average difference in outcomes that would result from two alternative treatments or interventions.
• cross-fittinguses · Cross-fitting is a procedure that partitions data, trains models on complementary subsets, and evaluates predictions on held-out subsets to reduce overfitting in causal estimation.
• s-learneraddresses · An S-learner is a causal estimation algorithm that fits one outcome model containing the treatment and confounders and predicts outcomes under alternative treatment values.
• root-n consistencycontrasts with · Root-n consistency is convergence of an estimator to its target at a rate proportional to the inverse square root of the sample size.
• double machine learningcontrasts with · 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.