Sensitivity analysis (SA) can be applied to building energy models (BEM) to identify which input parameters that drive the majority of the model output variation. The screening-based Morris method is often applied for this purpose; however, consideration regarding the effect of the user-defined number of levels (p) and trajectories (r) on the obtained results are rare. This paper investigates how the choice of p and r affects the outcome of a SA using the Morris method on a high fidelity BEM. The results indicates that the Morris method was not able to replicate the ranking from the variance-based Sobol' method no matter the choice of r and p. It was, however, able to identify groups of input parameters (parameter clusters) most sensitive to the model output variability, but it required significantly more r than usually applied in studies featuring the Morris method. The reason is that marginal differences in absolute values of elementary effects (the sensitivity indices of the Morris method) for some input parameters may lead to a change in ranking position several times as the number of r increases. Users of the Morris method must therefore not be predetermined on the size of the parameter cluster; instead, they must make a visual assessment of the convergence of the parameter ranking to qualitatively determine the appropriate size of parameter cluster. The final recommendation for future studies deploying the Morris method for SA applied to a high fidelity BEM is to choose p >= 4 as it seems to lead the analysis towards a more truthful ranking, and then run simulations in steps of r = 100 when making the visual assessment to determine convergence and the size of parameter cluster. The identified need for more r questions the general notion that the Morris method is a computationally efficient screening method in terms of absolute time use. However, the Morris method is still much more computational efficient than a Sobol'-based analysis if the purpose of the SA is to identify a cluster of input parameters most sensitive to the model output variability. (C) 2018 Elsevier B.V. All rights reserved.