We used a network of 135 NO2 passive diffusion tube sites to develop land use regression (LUR) models in a UK conurbation. Network sites were divided into four groups (32-35 sites per group) and models developed using combinations of 1-3 groups of "training" sites to evaluate how the number of training sites influenced model performance and residential NO2 exposure estimates for a cohort of 13 679 participants. All models explained moderate to high variance in training and independent "hold-out" data (Training adj. R2: 62-89%; Hold-out R2: 44-85%). Average hold-out R2 increased by 9.5%, while average training adj. R2 decreased by 7.2% when the number of training groups was increased from 1 to 3. Exposure estimate precision improved with increasing number of training sites (median intralocation relative standard deviations of 19.2, 10.3, and 7.7% for 1-group, 2-group and 3-group models respectively). Independent 1-group models gave highly variable exposure estimates suggesting that variations in LUR sampling networks with relatively low numbers of sites (≤35) may substantially alter exposure estimates. Collectively, our analyses suggest that use of more than 60 training sites has quantifiable benefits in epidemiological application of LUR models.