Finite Sample Behavior of MLE in Network Autocorrelation Models
This work evaluates the finite sample behavior of ML estimators in network autocorrelation models, a class of auto-regressive models studying the network effect on a variable of interest. Through an extensive simulation study, we examine the conditions under which these estimators are normally distributed in the case of finite samples. The ML estimators of the autocorrelation parameter have a negative bias and a strongly asymmetric sampling distribution, especially for high values of the network effect size and the network density. In contrast, the estimator of the intercept is positively biased but with an asymmetric sampling distribution. Estimators of the other regression parameters are unbiased, with heavy tails in presence of non-normal errors. This occurs not only in randomly generated networks but also in well-established network structures.
KeywordsNetwork effect model Density Network topology Non-normal distribution
- 13.Zhang, B., Thomas, A.C., Doreian, P., Krackhardt, D., Krishnan, R.: Contrasting multiple social network autocorrelations for binary outcomes with applications to technology adoption. ACM T. Man. Inf. Syst. 3, 1–21 (2013)Google Scholar