Observational-Interventional Priors for Dose-Response Learning

Authors: Ricardo Silva

NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we introduce a hierarchical Gaussian process prior that constructs a distribution over the doseresponse curve by learning from observational data, and reshapes the distribution with a nonparametric affine transform learned from controlled interventions. This function composition from different sources is shown to speed-up learning, which we demonstrate with a thorough sensitivity analysis and an application to modeling the effect of therapy on cognitive skills of premature infants. Section 4 provides a thorough set of experiments assessing our approach, including sensitivity to model misspecification.
Researcher Affiliation Academia Ricardo Silva Department of Statistical Science and Centre for Computational Statistics and Machine Learning University College London ricardo@stats.ucl.ac.uk
Pseudocode No The paper describes the inference process as Gibbs sampling and slice sampling, but it does not present pseudocode or an algorithm block.
Open Source Code Yes Code is also provided to regenerate our data and re-run all of these experiments.
Open Datasets Yes We consider an adaptation of the study analyzed by [7]. Targeted at premature infants with low birth weight, the Infant Health and Development Program (IHDP) was a study of the efficacy of educational and family support services and pediatric follow-up offered during the first 3 years of life [3].
Dataset Splits No The paper describes the generation of observational and interventional data with specified sizes (N=1000, M=40, 100, 200, |X|=20) and the use of the interventional data for evaluation. However, it does not explicitly define traditional training, validation, and test splits with percentages or sample counts for hyperparameter tuning or final evaluation sets.
Hardware Specification No The paper does not provide any specific details about the hardware used for running experiments.
Software Dependencies No The paper mentions 'R/MATLAB scripts' for data preprocessing, but it does not specify version numbers for these or any other software dependencies, libraries, or solvers.
Experiment Setup Yes We generate studies where the observational sample has N = 1000 data points and |Z| = 25 confounders. Interventional data is generated at three different levels of sample size, M = 40, 100 and 200 where the intervention space X is evenly distributed within the range shown by the observational data, with |X| = 20. Hyper-hyperpriors on λa and σa are set as log(λa) N(0, 0.5), log(σa) N(0, 0.1). We infer posterior distributions by Gibbs sampling, alternating the sampling of latent variables f(X), a(X), b(X) and hyperparameters λa, σa, λb, σb, σ2 int, using slice sampling [15] for the hyperparameters.