When dealing with small populations of elusive species, capture–recapture methods suffer from sampling and analytical limitations, making abundance assessment particularly challenging. We present an empirical and theoretical evaluation of multiple data source sampling as a flexible and effective way to improve the performance of capture–recapture models for abundance estimation of small populations. We integrated three data sources to estimate the size of the relict Apennine brown bear (Ursus arctos marsicanus) population in central Italy, and supported our results with simulations to assess the robust-ness of multiple data source capture–recapture models to violations of main assumptions. During May–August 2008, we non-invasively sampled bears using systematic hair traps on a grid of 41 5 Â 5 km cells, moving trap locations between five sampling sessions. We also live-trapped, ear-tagged, and genotyped 17 bears (2004–2008), and integrated resights of marked bears and family units (July– September 2008) into a multiple data source capture–recapture dataset. Population size was estimated at 40 (95% CI = 37–52) bears, with a corresponding closure-corrected density of 32 bears/1000 km 2 (95% CI = 28–36). Given the average capture probability we obtained with all data sources combined (^ p ¼ 0:311), simulations suggested that the expected degree of correlation among data sources did not seriously affect model performance, with expected level of bias <5%. Our results refine previous simula-tion work on larger populations, cautioning on the combined effect of lack of independence and low cap-ture probability in application of multiple data source sampling to very small populations (N < 100).