Model uncertainty is prevalent in probabilistic seismic hazard analysis (psha) because the underlying statistical signatures for hazard are unknown. Although methods for incorporating parameter uncertainty of a particular model in psha are well understood, methods for incorporating model uncertainty are more difficult to implement because of the high degree of dependence between different earthquake-recurrence models. We show that the method used by the 2002 Working Group on California Earthquake Probabilities (wgcep-2002) to combine the probability distributions given by multiple earthquake-recurrence models has several adverse effects on their results. In particular, wgcep-2002 uses a linear combination of the models that ignores model dependence and leads to large uncertainty in the final hazard estimate. Furthermore, model weights were chosen based on data, which has the potential to systematically bias the final probability distribution. The weighting scheme used in the Working Group report also produces results that depend on an arbitrary ordering of models. In addition to analyzing current statistical problems, we present alternative methods for rigorously incorporating model uncertainty into psha.