Non-probability sampling involves selecting samples from a population in which the probability of selection is unknown and some population units may have zero selection probabilities. This differentiates it from probability sampling where selection is governed by a probability model and every population unit has a non-zero chance of being selected. Nonprobability samples usually suffer from selection bias and hence may not represent the target population accurately. An important problem that arises in this context is the prediction of responses corresponding to non-sampled units, which should ideally have been sampled. In this article, we propose three modeling frameworks to address this issue. We use propensity scores to balance the sampled and non-sampled units and a Bayesian estimation scheme for parameter inference and prediction. We incorporate a spatial poststratification scheme to assess the predictive ability of our models on a simulated dataset. In addition, we perform model selection routines to identify the optimal model having the best predictive ability.