Source-data free domain adaptation (SFDA) is from unsupervised domain adaptation (UDA) and applied to the special situations in reality that the source domain data is not accessible. To this end, a common method is self-supervision with marking pseudo labels on the data in target domain. While the method is simple and effective, error hard labels degenerate the final model seriously. In this paper, we introduce the latent dirichlet allocation (LDA) algorithm to enhance the predictive ability of the target model. We consider all the sample features within a neighborhood follow the same distribution. The prediction from the target model is viewed as a prior distribution of the sample. And we improve the prediction with pseudo-labels of samples in its neighborhood with the help of LDA. Compared with the baseline SHOT-IM, the experiment results show that our method is improved 1.3% and 1.6% on datasets Office-31 and Office-Home respectively.