We consider a set of potentially misspecified structural models, geometrically combine their likelihood functions, and estimate the parameters using composite methods. In a Monte Carlo study, composite estimators dominate likelihood‐based estimators in mean squared error and composite models are superior to individual models in the Kullback–Leibler sense. We describe Bayesian quasi‐posterior computations and compare our approach to Bayesian model averaging, finite mixture, and robust control procedures. We robustify inference using the composite posterior distribution of the parameters and the pool of models. We provide estimates of the marginal propensity to consume and evaluate the role of technology shocks for output fluctuations.

Model misspecification composite likelihood Bayesian model averaging finite mixture C13 C51 E17