Networks with nonidentical nodes and global coupling may display a large variety of dynamic behaviors, such as phase clustered solutions, synchrony, and oscillator death. The network dynamics is a function of the parameter dispersion and may be captured by conventional mean field approaches if it is close to the completely synchronous state. In this Letter we introduce a novel method based on a mode decomposition in the parameter space, which provides a low-dimensional network description for more complex dynamic behaviors and captures the mean field approach as a special case. The example of globally coupled Fitzhugh-Nagumo neurons is discussed.,