Indirect testing of analog and RF integrated circuits is a widely studied approach, which has the benefits of relaxing requirements on test equipment and reducing industrial test cost. It is based on machine-learning algorithms to train a regression model that maps an indirect and low-cost measurement space to the performance parameter space. In this work, we explore the benefit of using ensemble learning. Rather than using one single model to estimate targeted parameters, ensemble learning consists of training multiple individual regression models and combining their outputs in order to improve the predictive power. Different ensemble methods based on bagging, boosting or stacking are investigated and compared to classical individual models. Results are illustrated and discussed on three RF performances of a LNA for which we have production test data.