An Efficient Doubly-Robust Test for the Kernel Treatment Effect
Authors: Diego Martinez Taboada, Aaditya Ramdas, Edward Kennedy
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we explore the empirical calibration and power of the proposed test AIPW-x KTE. For this, we assume that we observe (xi, ai, yi)n i=1 (X, A, Y ) and that (causal inference assumptions) consistency, no unmeasured confounding, and overlap hold. Both synthetic data and real data are evaluated. All the tests are considered at a 0.05 level. |
| Researcher Affiliation | Academia | Diego Martinez-Taboada Department of Statistics and Data Science Carnegie Mellon University Pittsburgh, PA 15213 diegomar@andrew.cmu.edu Aaditya Ramdas Department of Statistics and Data Science Machine Learning Department Carnegie Mellon University Pittsburgh, PA 15213 aramdas@stat.cmu.edu Edward H. Kennedy Department of Statistics and Data Science Carnegie Mellon University Pittsburgh, PA 15213 edward@stat.cmu.edu |
| Pseudocode | Yes | Algorithm 1 AIPW-x KTE |
| Open Source Code | No | The paper does not provide explicit statements or links to open-source code for the described methodology. |
| Open Datasets | Yes | We use data obtained from the Infant Health and Development Program (IHDP) and compiled by Hill (2011) |
| Dataset Splits | No | The paper describes data splitting for the test statistic derivation (D1 and D2 for cross U-statistics) and training of nuisance parameters, but does not provide explicit train/validation splits for model tuning or evaluation in the traditional sense. For example, it states 'train them on D. Otherwise, train them on D1 r.' but no specific validation split or methodology for hyperparameter tuning using a validation set is detailed. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models, memory, or cloud instance types. |
| Software Dependencies | No | The paper mentions various methods and models like 'L2 regularized logistic regression', 'conditional mean embeddings', 'Causal Forests', 'BART', and 'RBF kernel', but does not specify any software names with version numbers or library dependencies used for implementation. |
| Experiment Setup | Yes | We define Y 0 = βT X + ϵ0, Y 1 = βT X + b + ϵ1, such that ϵ0, ϵ1 N(0, 0.5) are independent noises, X N(0, I5) and β = [0.1, 0.2, 0.3, 0.4, 0.5]T . We set b = 0 and b = 2 for Scenario I and Scenario II respectively. For Scenario III, we set b = 2Z 1, where Z is an independent Bernoulli random variable Z Bernoulli(0.5). In Scenario IV, b Uniform( 2, 2). In the observational setting, we define π(X) = s(αT X + α0), where s(z) = 1/(1 + exp( z)) (sigmoid function), α0 = 0.05 and α = [0.05, 0.04, 0.03, 0.02, 0.01]T . In such setting, we estimate ˆπ using an L2 regularized logistic regression with the regularization term set to 1e-6. ... All the kernels considered on Y for AIPW-x KTE, IPW-x KTE, and KTE are RBF i.e. ky(y1, y2) = exp(νy|y1 y2|2), with scale parameter νy chosen by the median heuristic. ... The kernel considered for such conditional mean embeddings on X is also RBF kx(x1, x2) = exp(νx|x1 x2|2) with scale parameter νx chosen by the median heuristic as well. The regularization parameter λ was taken to be equal to νx. |