Marginal Causal Flows for Validation and Inference

Authors: Daniel de Vassimon Manela, Laura Battaglia, Robin Evans

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the above with experiments on both simulated and real-world datasets.
Researcher Affiliation Academia Daniel de Vassimon Manela University of Oxford manela@stats.ox.ac.uk Laura Battaglia University of Oxford battaglia@stats.ox.ac.uk Robin J. Evans University of Oxford evans@stats.ox.ac.uk
Pseudocode No The paper describes the architecture and training process in prose and uses diagrams but does not include any pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes The FF software used for this paper can be found in the Git Hub repository https://github.com/llaurabatt/frugal-flows.git.
Open Datasets Yes The first is the Lalonde data, taken from a randomised control trial to study the effect of a temporary employment program in the US on post intervention income level (La Londe, 1986). The second is an observational dataset used to quantify the effect of individuals 401(k) eligibility on their accumulated net assets, in the presence of several relevant covariates (Abadie, 2003).
Dataset Splits Yes For both datasets, a random hyperparameter search was conducted by choosing the hyperparameter set which minimised the validation loss, given a train/test data split of 9/1.
Hardware Specification Yes All experiments were run on a Mac Book (16-inch, 2021) with an M1 Max chip and 32GB memory using the CPU.
Software Dependencies Yes FF software builds upon Flow Jax (Ward, 2024), a Python package implementing normalising flows in JAX (Bradbury et al., 2018). [...] For MLE optimisation, we take advantage of JAX automatic differentiation capabilities and use the Adam optimiser (Kingma and Ba, 2015). [...] The simulated data generated for the inference experiments was generated using the causl package written in R, which was called in Python via the rpy2 package (Evans, 2021; Evans and Didelez, 2024).
Experiment Setup Yes Tunable hyperparameters to the Frugal Flow component are the number of subflows of the multivariate NSF, the width and depth of the MLPs and the number of spline knots, together with the specific hyperparameters of the chosen FY |T. [...] Table 2: Runtime and hyperparameters for fitting 25 different runs of each model, with a datasize of 15,000. [...] Table 4: Runtime and hyperparameters for fitting both a Frugal and Propensity Flow to the Lalonde and e401(k) data.