Random effects regression imputation has been recommended for multiple imputation in cluster randomized trials since it is congenial to analyses that use random effects regression. However, this method relies heavily on model assumptions and may not be the most powerful or most appropriate analysis, particularly when there are few clusters. We propose three new multiple imputation procedures based on predictive mean matching (PMM) that are more robust to misspecification of the imputation model. Instead of using a single model to define “closeness” of predictive means, we use two models in combination: one that ignores clustering (PMM-IGN) and one that uses fixed effects for clusters (PMM-FE). On their own, these imputation models result in underestimation (PMM-IGN) or overestimation (PMM-FE) of variance estimates. To leverage the ambidirectional bias, our proposed PMM procedures combine these two models (1) using a weighted distance metric; (2) using a weighted average of the responses selected for imputation; or (3) using a weighted draw from the selected responses. Our methods effectively reduce the bias in the variance estimates relative to established methods.