Gaussian Mixture Counterfactual Generator
Authors: Jong-Hoon Ahn, Akshay Vashist
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate GMCG on synthetic crossover trial data and simulated datasets, demonstrating its superior performance over existing methods, particularly in scenarios with limited control data. GMCG derives its effectiveness from modeling the joint distribution of covariates, treatments, and outcomes using a latent state vector while employing a conditional distribution of the state vector to suppress confounding and isolate treatment-outcome relationships. |
| Researcher Affiliation | Industry | Jong-Hoon Ahn and Akshay Vashist Otsuka Pharmaceutical Development & Commercialization, Inc. 508 Carnegie Center Dr, Princeton, NJ 08540 EMAIL |
| Pseudocode | Yes | Algorithm 1 Gaussian Mixture Counterfactual Generator (GMCG) Require: (1) Factual training data set {xn,f}; (2) Factual test data xf and alternative treatment acf Ensure: (1) GMCG parameters Q = {{πk}, {µk}, {Σk}, W , Ψ}; (2) Counterfactual data xcf |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. It only mentions that "The appendices present the GMCG EM algorithm and pseudocode, among other details, to facilitate a deeper understanding of its implementation." This refers to conceptual details, not runnable code. |
| Open Datasets | No | For quantitative analysis, LDL cholesterol simulated data was utilized. Data for the treatment and control groups were created using the PK/PD model (Faltaos et al., 2006; Yokote et al., 2008; Kim et al., 2011)... and "Additionally, in this section, we applied the algorithm to tumor growth data (Geng et al., 2017), which is commonly used for validating general causal inference algorithms." The paper describes using simulated data and references models or previous work, but does not provide concrete access information (link, DOI, repository) for the specific datasets used in their experiments. |
| Dataset Splits | Yes | In Table 1, N0 = 200 control data and N1 = 200 treatment data were used. In addition, N0 = 1000 control data and N1 = 200 treatment data were also used as training data. But now, we aim to evaluate a much more challenging task: using no control data at all... After training the model using random doses data, we tested it on the 10mg test data used in Table 1... |
| Hardware Specification | No | The paper does not provide specific hardware details used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers. It refers to CRN model hyperparameters but not specific software versions for its own implementation or general experimental setup. |
| Experiment Setup | Yes | In the GMCG algorithm for crossover trials, the choice of initial values affected performance... The results in Table 2 used two mixture components for dose data between 0mg and 3mg, one for 3mg to 7mg, and one for 7mg to 10mg. Thus, a total of K = 4 mixture components were used. Also, Appendix F.1 provides hyperparameters for the CRN model: "num epochs = 100 { rnn hidden units : 24, br size : 12, fc hidden units : 36, learning rate : 0.01, batch size : 128, rnn keep prob : 0.9 }" |