The recent published literature on linear network autocorrelation models of actor behavior or mutable attributes has revealed a curious finding. Likelihood-based (e.g., maximum likelihood and Bayesian posterior mean, median and mode) estimators of a peer effect parameter ( \(\rho \) ) tend to be negatively biased and become increasingly biased as the density of the network increases irrespective of network size and the status of other network features. A notable limitation of the current literature is that bias and other operating characteristics of estimators of \(\rho \) for several common generative models of network relationships have not been studied. In this paper we investigate the pattern of bias of estimators of \(\rho \) when analyzing multiple mutually exclusive sub-networks. In addition to binary-valued networks, we investigate the pattern of bias for networks with weighted edges. We perform simulation studies that reveal that bias diminishes rapidly as the number of sub-networks increases and as the variance of the edge-weights across the network increases.