Counterfactual Analysis in Dynamic Latent State Models

Authors: Martin B Haugh, Raghav Singal

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply it on a breast cancer case study. We now apply our approach to the breast cancer application we described in 1. The results for path 1 are displayed in Figure 3 (and for path 2 in Figure 10 ( E.5)), where we show the PN bounds as we vary T.
Researcher Affiliation Academia Martin Haugh 1 Raghav Singal 2 1Imperial College 2Dartmouth College. Correspondence to: MH <m.haugh@imperial.ac.uk>, RS <singal@dartmouth.edu>.
Pseudocode Yes Algorithm 1 Counterfactual analysis via optimization and Algorithm 2 Counterfactual simulations under the independence copula and Algorithm 3 Counterfactual simulations under the comonotonic copula
Open Source Code No The paper does not provide an explicit statement about releasing its source code or a link to a code repository.
Open Datasets Yes The primitives (p, E, Q) are calibrated to real-data using a mix of sources, which we discuss in E.1. NIH (2020). URL https://seer.cancer.gov/ archive/csr/1975_2017/results_merged/ sect_04_breast.pdf. UWBCS. University of Wisconsin Breast Cancer Simulation Model. 2013. URL https://resources.cisnet.cancer.gov/ registry/packages/uwbcs-wisconsin/.
Dataset Splits No The paper describes using "real-data" from external sources to calibrate model primitives, but does not specify train/validation/test dataset splits for its experiments. The mention of "B = 100 samples" refers to Monte Carlo samples, not dataset splits.
Hardware Specification Yes It solved each of our problem instances to global optimality within minutes / hours (depending on T), with an absolute termination tolerance of 0.01 (on an Intel Xeon E5 processor with 16 GB RAM).
Software Dependencies Yes We implemented in MATLAB (MATLAB, 2021). The feasibility set F over (θ, π) corresponds to (10), (11), and (12). To solve the polynomial optimizations, we use the MATLAB-BARON interface (Sahinidis, 2023) with CPLEX (IBM, 2017) as the LP / MIP solver. MATLAB. Version 9.10.0 (R2021b). The Math Works Inc., Natick, Massachusetts, 2021. IBM. ILOG CPLEX Optimizer Version 12.8. 2017. Sahinidis, N. V. BARON 2023.1.5: Global Optimization of Mixed-Integer Nonlinear Programs, User s Manual, 2023.
Experiment Setup Yes We vary T {4, . . . , 10}, with a larger value of T suggesting the cancer may have progressed more slowly. We generated B = 100 samples using our sampling method in B. We ensured this stability by computing our results for 20 seeds (for each (path, T) pair) and verifying the standard deviations to be small. It solved each of our problem instances to global optimality within minutes / hours (depending on T), with an absolute termination tolerance of 0.01