Causal Prediction Algorithms

Dataiku offers two different methodologies for estimating causal effects:

Meta-learning

Meta-learning leverages machine learning algorithms to learn the causal effect of the treatment (Conditional Average Treatment Effect or CATE). These ML models are called base learners. The specific way they are trained and combined is referred to as the meta-learner. Dataiku supports three meta-learners:

• S-learner: a single model is trained with the treatment variable as an input feature. The predicted CATE is the difference between the prediction with the treatment variable set as treated and the prediction with the treatment variable set as control.

• T-learner: two models are trained, one on the treated group, the other on the control group. The predicted CATE is the difference between the predictions from the two models.

• X-learner:

  • Two models are trained, one on the treated group, the other on the control group (same as T-learner).

  • They are used to individually predict the counterfactual outcome (i.e. control for treated group and treated for control group), which is combined with the observed outcome to estimate the individual treatment effect.

  • Then, two other models (again, one for treated group and one for control group) are trained on the individual effects previously predicted. The CATE prediction is the average prediction of these last two models weighted by a predicted propensity (individual predicted probability of getting the treatment).

You can combine any meta-learner with any available Python-based ML algorithm as a base learner.

Causal forest

Causal Forests are tree-based ensemble models similar to Random Forests. The main differences are:

• the value of nodes and leaves are based on both the treatment and the outcome variables - they are an estimation of the Conditional Average Treatment Effect (CATE)

• the splitting criterion is also based on both the treatment and the outcome variables

• the honest framework (whenever enabled) allows using two separate subsets of the train set to compute the optimal split and the value of a node.

Parameters:

• Number of trees: Number of trees in the causal forest. Increasing the number of trees in a causal forest does not result in overfitting. This parameter can be optimized.

• Feature sampling strategy: Adjusts the number of features to sample at each split.

  • Automatic will select 30% of the features.

  • Square root and Logarithm will select the square root or base 2 logarithm of the number of features respectively

  • Fixed number will select the given number of features

  • Fixed proportion will select the given proportion of features

• Maximum depth of tree: Maximum depth of each tree in the causal forest. Higher values generally increase the quality of the prediction but can lead to overfitting. Higher values also increase the training and prediction time. Use 0 for unlimited depth (i.e., keep splitting the tree until each node contains the minimum number of samples per leaf),

• Minimum samples per leaf: Minimum number of samples required in a single tree node to split this node. Lower values increase the quality of the prediction (by splitting the tree mode), but can lead to overfitting and increased training and prediction time.

• Honest: Whether or not the honest framework is used to build trees. If set to true, the learning algorithm will use two separate subsets of the train set to compute the optimal split and the value of a node.

• Parallelism: Number of cores used for parallel training. Using more cores leads to faster training but at the expense of more memory consumption, especially for large training datasets.