Longitudinal Targeted Minimum Loss-based Estimation with Temporal-Difference Heterogeneous Transformer
Authors: Toru Shirakawa, Yi Li, Yulun Wu, Sky Qiu, Yuxuan Li, Mingduo Zhao, Hiroyasu Iso, Mark J. Van Der Laan
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Simulation results demonstrate our method s superior performance over existing approaches, particularly in complex, long time-horizon scenarios. It remains effective in small-sample, short-duration contexts, matching the performance of asymptotically efficient estimators. To demonstrate our method in practice, we applied our method to estimate counterfactual mean outcomes for standard versus intensive blood pressure management strategies in a realworld cardiovascular epidemiology cohort study. |
| Researcher Affiliation | Academia | 1Osaka University Graduate School of Medicine, Suita, Japan 2Univerity of California, Berkeley, United States 3National Center for Global Health and Medicine, Tokyo, Japan. |
| Pseudocode | Yes | Algorithm 1 Temporal Difference Learning of Conditional Counterfactual Mean Outcomes, Algorithm 2 Temporal-Difference Targeting, Algorithm 3 Sequential Targeting. |
| Open Source Code | Yes | Code for experimens with synthetic data is available at https://github.com/shirakawatoru/dltmle-icml-2024. |
| Open Datasets | Yes | For this experiment, we used covariates from the Circulatory Risk in Communities Study (CIRCS) (Yamagishi et al., 2019) |
| Dataset Splits | Yes | We selected hyperparameters shown in Table 4 which optimized the empiricall loss LQ + Le in the validation set which is the 30% of the entire dataset. |
| Hardware Specification | Yes | Deep ACE and Deep LTMLE were run on a GPU (Tesla T4) with 16 GB memory and LTMLE on CPU (Intel Xeon Skylake 6230 @ 2.1 GHz) with 40 cores and 96 GB memory. |
| Software Dependencies | No | We used the R package ltmle with GLM and a super learner (SL) library consisting of GLM, maltivariate adaptive regression spline with earth package, and xgboost for the simple synthetic data and the real world data (Lendle et al., 2017; Polley et al., 2021; Milborrow, 2023; Chen et al., 2022). |
| Experiment Setup | Yes | We selected hyperparameters shown in Table 4 which optimized the empiricall loss LQ + Le in the validation set which is the 30% of the entire dataset. Table 4 provides specific values for 'Embedding dimension', 'Dropout rate', 'Hidden size', 'Number of Layers', 'Number of heads', 'Learning rate', 'α', 'β', and 'Number of epochs' for different models and τ values. |