A Significance Test for Graph-Constrained Estimation

Abstract

Graph-constrained estimation methods encourage similarities among neighboring covariates presented as nodes on a graph, which can result in more accurate estimations, especially in high dimensional settings. Variable selection approaches can then be utilized to select a subset of variables that are associated with the response. However, existing procedures do not provide measures of uncertainty of the estimates. Moreover, the vast majority of existing approaches assume that available graphs accurately capture the association among covariates; violating this assumption could severely hurt the reliability of the resulting estimates. In this paper, we present an inference framework, called the Grace test, which simultaneously produces coefficient estimates and corresponding $p$-values while incorporating the external graph information. We show, both theoretically and via numerical studies, that the proposed method asymptotically controls the type-I error rate regardless of the choice of the graph. When the underlying graph is informative, the Grace test is asymptotically more powerful than similar tests that ignore external information. We further propose a more general Grace-ridge test that results in a higher power than the Grace test when the choice of the graph is not fully informative. Our numerical studies show that as long as the graph is reasonably informative, the proposed testing methods deliver improved statistical power over existing inference procedures that ignore external information.