One possibility is discussed in Cameron and Trevedi's "Regression Analysis of Count Data" which describes the deviance as the GLM generalization of the sum of squares. They refer to two papers by Cameron and Windmeijer where a pseudo-R-squared is proposed based on a decomposition of the deviance. In essence, the proposal is to use R2 = 1-(D(intercept-only-model)/D(full-model)). I would recommend reviewing the original papers to verify that the interpretation is appropriate for your particular model before using the result.
... View more
