Random Forests Approach for Causal Inference with Clustered Observational Data.

Area of Science:

• Causal inference
• Machine learning
• Educational statistics

Background:

• Machine learning (ML) methods are increasingly used for causal inference.
• Existing ML causal inference methods primarily focus on independent, single-level data.
• Clustered or nested data, common in education, present challenges for standard ML causal inference.

Purpose of the Study:

• To investigate the efficacy of Causal Forests for estimating treatment effects in multilevel observational data.
• To adapt ML causal inference methods for hierarchical data structures found in educational research.
• To evaluate the performance of modified Causal Forests against existing methods.

Main Methods:

• Utilized Causal Forests, an ML method based on random forests.
• Conducted simulation studies on two-level, three-level, and cross-classified multilevel data.
• Proposed and tested a modified ML approach integrating multilevel model propensity scores.

Main Results:

• The modified ML method, incorporating multilevel propensity scores, outperformed preexisting methods.
• The enhanced approach demonstrated effectiveness across various multilevel data complexities.
• Significant improvements in treatment effect estimation were observed in simulation studies.

Conclusions:

• ML causal inference methods can be effectively adapted for multilevel observational data.
• Supplementing Causal Forests with multilevel model propensity scores is a viable strategy for hierarchical data.
• The study provides a robust method for analyzing complex educational datasets, exemplified by estimating private math lesson effects.