Network interdependence of sample units in contingency tables

Abstract

Analytical, empirical, and simulation evidence suggest that chi‐square and likelihood‐ratio tests reject at rates substantially lower than the nominal rate whenever the data have been generated using a sampling scheme that does not ensure independence of sample units. A number of deflation factors have been proposed as adjustments to these test statistics in order to raise rejection rates. Most, however, appear to overadjust and result in very conservative tests. In this paper, we propose a new deflation factor in the general case of network interrelated sample units. An empirical example suggests that this new procedure results in less conservative inference than previous approaches.