We propose new model-based methods for unit non-response in two-stage survey samples. A commonly used design-based adjustment weights respondents by the inverse of the estimated response rate in each cluster (method WT). This approach is consistent if the response probabilities are constant...