Causal Modeling of Policy Interventions From Treatment-Outcome Sequences
Authors: Çağlar Hızlı, S. T. John, Anne Tuulikki Juuti, Tuure Tapani Saarinen, Kirsi Hannele Pietiläinen, Pekka Marttinen
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show with real-world and semi-synthetic data on blood glucose progression that our method can answer causal queries more accurately than existing alternatives. and 7. Experiments In this section, we validate that our model can estimate interventional and counterfactual outcome trajectories under policy interventions. The empirical validation is composed of two parts. First, on a real-world observational dataset, we show the proposed model can learn clinically meaningful treatment response curves and treatment intensities that can handle time-varying confounding. Second, we evaluate our model on two causal inference tasks: (i) the policy intervention and (ii) the policy counterfactual. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Aalto University, Helsinki, Finland 2Helsinki University Hospital and University of Helsinki, Helsinki, Finland. |
| Pseudocode | Yes | Algorithm 1 Ogata s Thinning algorithm and Algorithm 2 Counterfactual Sampling Algorithm for TPPs (Based on Ogata s) |
| Open Source Code | Yes | The study is reproducible and our implementation in GPflow (van der Wilk et al., 2020) can be found at https://github.com/caglar-hizli/ modeling-policy-interventions. |
| Open Datasets | Yes | We show with real-world and semi-synthetic data on blood glucose progression that our method can answer causal queries more accurately than existing alternatives. and We first demonstrate that our model can learn clinically meaningful treatment intensities and response curves from a real-world observational dataset on physiological dynamics of blood glucose (Zhang et al., 2020; Wyatt et al., 2021). |
| Dataset Splits | No | We train each model on the first two days of observations and use the third day as the test set. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU, memory) used for running the experiments. |
| Software Dependencies | No | The study is reproducible and our implementation in GPflow (van der Wilk et al., 2020) can be found at https://github.com/caglar-hizli/ modeling-policy-interventions. |
| Experiment Setup | Yes | We choose the constant baseline intensity β0 to be equal to 0.1. For the treatment-dependent function g a(τ; a), we use a relative-time kernel kt only... We choose the dimensionality parameters {Qa, Qo} of the kernels as Qa = 1, Qo = 1... We choose kernel hyperparameters θa,h = {γq, ℓq}Qa+Qo q=0 by inspection... The constant kernel function has an intercept parameter {b(v) b }, initialized to {1.0}. Each periodic kernel has parameters {α(v) b , ℓ(v) b }, initialized to {1.0, 1.0}. |