Unbalanced phase currents, which flow in transformer windings and distribution wires, cause a significant increase (~33%) of phase energy losses in low voltage (LV, 415 V) networks. However, these additional phase energy losses (APELs) are hard to calculate for most LV networks. A key challenge is that these LV networks are data-scarce, with only yearly average and maximum phase currents. To estimate the APEL for data-scarce LV networks, this study proposes a statistical approach that effectively overcomes the above challenge. Firstly, the approach calculates APEL for a sample set of data-rich networks with year-round time-series phase current data. Secondly, features are extracted from these networks by considering: (i) whether the features are strongly correlated to APELs; and (ii) whether the features can be derived from available data (e.g. yearly average and maximum phase currents) from data-scarce networks. Thirdly, to approximate mappings from the features (derived in stage 2) to the APEL (derived in stage 1), a kernel-based regression model is developed, using the above customised features. Given any data-scarce network, its APEL is then estimated by applying the regression model. Cross-validation shows that the statistical approach incurs an average error of 13% for 90% of the data-scarce LV networks, excluding the networks with very low APEL values.