Online communities are composed of individuals sharing similar opinions or behavior in the virtual world. Facilitated by the fast development of social media platforms, the expansion of online communities have raised many attentions among the researchers, business analysts, and decision makers, leading to a growing list of literature studying the interactions especially conflicts in the online communities. A conflict is often initiated by one community which then attacks the other, leading to an adversarial relationship and worse social impacts. Many studies have examined the origins and process of online community conflict while failing to address the possible spatial effects in their models. In this paper, we explore the prediction of online community conflict by decomposing and analyzing its prediction error taking geography into accounts. Grounding on the previous natural language processing based model, we introduce pseudo spatial permutation to test the model expressiveness with geographical factors. Pseudo spatial permutation employs different geographical distributions to sample from and perturbs the model using the pseudo geographical information to examine the relationship between online community conflict and spatial distribution. Our analysis shows that the pseudo spatial permutation is an efficient approach to robustly test the conflict relation learned by the prediction model, and also reveals the necessity to incorporate geographical information into the prediction. In conclusion, this work provides a different aspect of analyzing the community conflict that does not solely rely on the textual communication.