A previous work estimated the cosmic microwave background (CMB) variance from the 3-yr Wilkinson Microwave Anisotropy Probe (WMAP) data, finding a lower value than expected from Gaussian simulations using the WMAP best-fitting cosmological model. We repeat the analysis on the 5-yr WMAP data using a new estimator with lower bias and variance. Our results confirm this anomaly at higher significance, namely with a p-value of 0.31 per cent. We perform the analysis using different exclusion masks, showing that a particular region of the sky near the Galactic plane shows a higher variance than 95.58 per cent of the simulations, whereas the rest of the sky has a lower variance than 99.96 per cent of the simulations. The relative difference in variance between both regions is bigger than in 99.64 per cent of the simulations. This anisotropic distribution of power seems to be causing the anomaly since the model assumes isotropy. Furthermore, this region has a clear frequency dependence between 41 and 61 GHz or 94 GHz suggesting that Galactic foreground residuals could be responsible for the anomaly. Moreover, removing the quadrupole and the octopole from data and simulations the anomaly disappears. The variance anomaly and the previously reported quadrupole and octopole alignment seem therefore to be related and could have a common origin. We discuss different possible causes and Galactic foreground residuals seem to be the most likely one. These residuals would affect the estimation of the angular power spectrum from the WMAP data, which is used to generate Gaussian simulations, giving rise to an inconsistency between the estimated and expected CMB variance. If the presence of residuals is confirmed, the estimation of the cosmological parameters could be affected.