Search for: All records

Summary Estimation of crossed random effects models commonly incurs computational costs that grow faster than linearly in the sample size $ N $, often as fast as $$ \Omega(N^{3/2}) $$, making them unsuitable for large datasets. For non-Gaussian responses, integrating out the random effects to obtain a marginal likelihood poses significant challenges, especially for high-dimensional integrals for which the Laplace approximation may not be accurate. In this article we develop a composite likelihood approach to probit models that replaces the crossed random effects model with some hierarchical models that require only one-dimensional integrals. We show how to consistently estimate the crossed effects model parameters from the hierarchical model fits. We find that the computation scales linearly in the sample size. The method is illustrated by applying it to approximately five million observations from Stitch Fix, where the crossed effects formulation would require an integral of dimension larger than $$ 700\,000 $$. 

Abstract Let $$f:[0,1]^{d}\to{\mathbb{R}}$$ be a completely monotone integrand as defined by Aistleitner and Dick (2015, Acta Arithmetica, 167, 143–171) and let points $$\boldsymbol{x}_{0},\dots ,\boldsymbol{x}_{n-1}\in [0,1]^{d}$$ have a non-negative local discrepancy (NNLD) everywhere in $$[0,1]^{d}$$. We show how to use these properties to get a non-asymptotic and computable upper bound for the integral of $$f$$ over $$[0,1]^{d}$$. An analogous non-positive local discrepancy property provides a computable lower bound. It has been known since Gabai (1967, Illinois J. Math., 11, 1–12) that the two-dimensional Hammersley points in any base $$b\geqslant 2$$ have NNLD. Using the probabilistic notion of associated random variables, we generalize Gabai’s finding to digital nets in any base $$b\geqslant 2$$ and any dimension $$d\geqslant 1$$ when the generator matrices are permutation matrices. We show that permutation matrices cannot attain the best values of the digital net quality parameter when $$d\geqslant 3$$. As a consequence the computable absolutely sure bounds we provide come with less accurate estimates than the usual digital net estimates do in high dimensions. We are also able to construct high-dimensional rank one lattice rules that are NNLD. We show that those lattices do not have good discrepancy properties: any lattice rule with the NNLD property in dimension $$d\geqslant 2$$ either fails to be projection regular or has all its points on the main diagonal. Complete monotonicity is a very strict requirement that for some integrands can be mitigated via a control variate. 

Langevin Monte Carlo (LMC) and its stochastic gradient versions are powerful algorithms for sampling from complex high-dimensional distributions. To sample from a distribution with density π(θ) ∝ exp(−U(θ)), LMC iteratively generates the next sample by taking a step in the gradient direction ∇U with added Gaus- sian perturbations. Expectations w.r.t. the target distribution π are estimated by averaging over LMC samples. In ordinary Monte Carlo, it is well known that the estimation error can be substantially reduced by replacing independent random samples by quasi-random samples like low-discrepancy sequences. In this work, we show that the estimation error of LMC can also be reduced by using quasi- random samples. Specifically, we propose to use completely uniformly distributed (CUD) sequences with certain low-discrepancy property to generate the Gaussian perturbations. Under smoothness and convexity conditions, we prove that LMC with a low-discrepancy CUD sequence achieves smaller error than standard LMC. The theoretical analysis is supported by compelling numerical experiments, which demonstrate the effectiveness of our approach.