Uplift modeling measures the impact of an action (marketing, medical treatment) on a person's behavior. Uplift prediction is based on groups of people who have received different treatments. These groups are assumed to be equivalent. However, in practice, we observe that there are biases between these groups. In this paper, we propose a protocol to evaluate and study the impact of non-random assignment bias (NRA) on the performance of the main uplift methods. Then we present a weighting technique to reduce the effect of NRA bias. Experimental results show an improvement in the performance of the uplift methods.