There are many types of dependencies within operational risk, such as loss frequency dependence, loss severity dependence and aggregate loss dependence. However, most studies simply choose one dependence type for their models, without solid justifications. This paper therefore proposes a method to choose the most appropriate dependence type for banks. First, we model different types of dependencies under the loss distribution approach to verify whether they indeed have distinct effects on operational risk measurement results. Then, we put forward three innovative criteria – strength, sensitivity and simplicity – to comprehensively evaluate the dependence types. The empirical analysis is based on the Chinese Operational Loss Database, the largest operational risk data set developed for China. The results show that the different dependence types do indeed have distinct effects on the estimated operational risk values for the Chinese banking industry. Further, loss frequency dependence and aggregate loss dependence are good choices for banks in practice, because they are very easy to model as well as very important.