A0974
Title: Estimation in network-linked high-dimensional multinomial probit models
Authors: Gourab Mukherjee - University of Southern California (United States) [presenting]
Abstract: The multinomial probit model (MNP) is widely used for analyzing unordered categorical data. In a host of contemporary applications, cross-sectional datasets are encountered with categorical responses and high-dimensional covariates. In the absence of repeated observations from the respondents, MNP models in these applications are often equipped with additional network structures that leverage contiguity in the responses of similar units. Pooling information across similar units through these network structures can provide significantly better inference in these data scarcity problems. However, estimating the effects of sparse high-dimensional covariates in the presence of network linkages among the responses is challenging. A penalized composite likelihood-based algorithm is developed that efficiently estimates the covariate effects. Decision theoretic guarantees are provided on the operational characteristic of the proposed algorithm. The application of the proposed method is demonstrated in spatial autocorrelation network structured MNP models. The performance of the proposed algorithm is compared on a wide range of simulation experiments, and encouraging performance is obtained.
