We use mean field theory to study the response properties of a simple randomly-connected model cortical network of leaky integrate-and-fire neurons with balanced excitation and inhibition. The formulation permits arbitrary temporal variation of the input to the network and takes exact account of temporal firing correlations. We find that neuronal firing statistics depend sensitively on the firing threshold. In particular, spike count variances can be either significantly greater than or significantly less than those for a Poisson process. These findings may help in understanding the variability observed experimentally in cortical neuronal responses.