Multiscale permutation entropy (MPE) has recently been proposed to evaluate complexity of time series. MPE has numerous advantages over other multiscale complexity measures, such as its simplicity, robustness to noise and its low computational cost. However, MPE may loose statistical reliability as the scale factor increases, because the coarse-graining procedure used in the MPE algorithm reduces the length of the time series as the scale factor grows. To overcome this drawback, we introduce the refined composite MPE (RCMPE). Through applications on both synthetic and real data, we show that RCMPE is much less dependent on the signal length than MPE. In this sense, RCMPE is more reliable than MPE. RCMPE could therefore replace MPE for short times series or at large scale factors.