Global sensitivity analysis based on polynomial chaos expansion (PCE) shows interesting characteristics, including reduced simulation runs for computer models and high interpretability of sensitivity results. This paper explores these features of the PCE-based sensitivity analysis using an office building as a case study with the EnergyPlus simulation program. The results indicate that the predictive performance of PCE models is closely correlated with the stability of the sensitivity index, depending on sample number and expansion degree. Therefore, it is necessary to carefully assess model accuracy of PCE models and evaluate convergence of the sensitivity index when using PCE-based sensitivity analysis. It is also found that more simulation runs of building energy models are required for a higher expansion degree of the PCE model to obtain a reliable sensitivity index. A bootstrap technique with a random sample can be used to construct confidence intervals for sensitivity indicators in building energy assessment to provide robust sensitivity rankings.