Generative Learner for Distributional Causal Effects
Organizations: Graduate School of Business, Stanford University · University of Glasgow, Adam Smith Business School · Booth School of Business, University of Chicago · Department of Systems Engineering and Operations Research, George Mason University
Abstract
We propose a generative learner for estimating conditional average treatment effects and characterizing the full distribution of these effects. The learner takes the form of a multi-head feed-forward neural network with three jointly estimated subnetworks: the propensity score, the conditional quantile function of the baseline outcome, and the conditional quantile treatment effect. Each quantile-indexed subnetwork combines a covariate representation with cosine quantile embeddings through element-wise multiplication. We then recover the conditional average treatment effect as an integral over conditional quantile treatment effects. Under the classical causal assumptions within the Neyman--Rubin potential outcomes framework, we find that the proposed generative learner reduces out-of-sample mean squared error relative to the generalized random forest, double machine learning, and generative adversarial networks, with gains ranging from 5.4% to 93.5% on average across experimental designs. In an empirical application, we formalize the Stefan--Boltzmann law within a unidirectional causal model and apply the method to publicly available stellar data. The estimated effects satisfy the restrictions the law implies.
Figures & tables
| GRF | DML-XGB | GANITE | GenTE | Avg. % gain (GenTE) | |||||||||||
| Sample size | vs GRF | vs DML | vs GANITE | ||||||||||||
| Linear | |||||||||||||||
| 200 | 0.800 | 0.861 | 0.994 | 0.465 | 0.610 | 0.808 | 1.029 | 0.874 | 1.123 | 0.107 | 0.345 | 0.712 | 58.3% | 44.1% | 62.3% |
| (0.188) | (0.188) | (0.222) | (0.148) | (0.166) | (0.215) | (0.362) | (0.294) | (0.381) | (0.067) | (0.165) | (0.380) | ||||
| 500 | 0.298 | 0.354 | 0.500 | 0.225 | 0.281 | 0.419 | 0.987 | 0.889 | 1.118 | 0.064 | 0.222 | 0.409 | 44.7% | 31.7% | 77.3% |
| (0.063) | (0.067) | (0.130) | (0.070) | (0.067) | (0.116) | (0.287) | (0.330) | (0.344) | (0.046) | (0.138) | (0.232) | ||||
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| GenTE (separate) | GenTE (joint, grid) | GenTE (joint, MC) | |||||||
| Sample size | |||||||||
| Linear | |||||||||
| 200 | 0.722 | 1.124 | 1.151 | 0.107 | 0.345 | 0.712 | 0.106 | 0.346 | 0.715 |
| (0.163) | (0.335) | (0.330) | (0.067) | (0.165) | (0.380) | (0.067) | (0.165) | (0.381) | |
| 500 | 0.676 | 1.068 | 1.124 | 0.064 | 0.222 | 0.409 | 0.063 | 0.223 | 0.410 |
| (0.122) | (0.200) | (0.145) | (0.046) | (0.138) | (0.232) | (0.046) | (0.138) | (0.233) | |
| GenTE (quantile, grid) | GenTE (quantile, MC) | GenTE (direct, squared loss) | |||||||
| Sample size | |||||||||
| Gaussian errors | |||||||||
| 200 | 0.107 | 0.345 | 0.712 | 0.106 | 0.346 | 0.715 | 0.321 | 0.337 | 0.366 |
| (0.067) | (0.165) | (0.380) | (0.067) | (0.165) | (0.381) | (0.218) | (0.123) | (0.140) | |
| 500 | 0.064 | 0.222 | 0.409 | 0.063 | 0.223 | 0.410 | 0.100 | 0.221 | 0.245 |
| (0.046) | (0.138) | (0.232) | (0.046) | (0.138) | (0.233) | (0.050) | (0.071) | (0.062) | |
Causal Effect Estimation
Causal effect estimation measures how outcomes would change under interventions, using experiments or observational data and explicit assumptions.
See all 287 papers