Generalized Optimal Matching Methods for Causal Inference
Authors: Nathan Kallus
JMLR 2020 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate this in examples with both synthetic and real data. ... We study the practical usefulness of KOM by applying the new methods developed to a semi-simulated case study using real data and find that KOM offers significant benefits in estimation error and also in robustness to practical issues like limited overlap and lack of model specification. |
| Researcher Affiliation | Academia | Nathan Kallus EMAIL Department of Operations Research and Information Engineering and Cornell Tech Cornell University New York, NY 10044, USA |
| Pseudocode | Yes | Algorithm listings are given in Appendix A. ALGORITHM 1: Cross-Validation Estimation for NNM++ ALGORITHM 2: BVE-CEM (solves eq. (12)) ALGORITHM 3: Affine Invariance by Studentization ALGORITHM 4: K-Affine Invariance by K-Studentization |
| Open Source Code | No | The paper does not provide explicit links to source code repositories or statements about code being released for the methodology described. It mentions using 'off-the-shelf solvers like Gurobi' but this refers to a third-party tool, not the authors' own implementation code. |
| Open Datasets | Yes | We next consider evaluating the practical usefulness of KOM++ by studying data from the Infant Health and Development Program (IHDP). IHDP was a randomized experiment intended to measure the effect of a program consisting of child care and home visits from a trained provider on early child development (Brooks-Gunn et al., 1992)... Following Hill (2012), mother s age and race are omitted as covariates. |
| Dataset Splits | Yes | Example 1 ...Fix a draw of X1:n, T1:n with n = 200. We plot the resulting draw, which has n0 = 130, n1 = 70... In sum, the treatment group (n1 = 94) consists only of children with older white mothers and the control group (n0 = 279) consists only of children with younger nonwhite mothers... For CEM, we select half of the covariates (13) ... leaving 69 units (only for CEM). |
| Hardware Specification | No | easily computed with off-the-shelf solvers like Gurobi (www.gurobi.com), which we use in all numerics in this paper to solve such problems in tens to hundreds of milliseconds on a personal laptop computer. The paper mentions 'a personal laptop computer' but does not provide any specific model, CPU, GPU, or memory details. |
| Software Dependencies | No | easily computed with off-the-shelf solvers like Gurobi (www.gurobi.com), which we use in all numerics in this paper to solve such problems... as implemented in the popular R package optmatch. The paper mentions 'Gurobi' and the 'R package optmatch' but does not specify their version numbers. |
| Experiment Setup | Yes | Example 1 Let X Unif[ 1, 1]2, P (T = 1 | X) = 0.95/(1 + 3 2 X 2). Fix a draw of X1:n, T1:n with n = 200. We plot the resulting draw, which has n0 = 130, n1 = 70, in Fig. 1a. For a range of λ we compute the resulting BVE-NNM weights using the Mahalanobis distance δ(x, x ) = (x x )ˆΣ 1 0 (x x ) where ˆΣ0 is the sample covariance of X | T = 0. ... Let Y (0) | X N( X 2 2 e T X/2, 3). In Fig. 1c, varying λ, we plot the resulting CMSE of ˆτW (solid) and ˆτW, ˆf0 (dashed) for ˆf0 given by ordinary least squares (OLS). |