A Bayesian non parametric approach for bias correction for underreported data.

Data quality is emerging as an essential characteristics of all data driven processes. The implications that data quality issues in computing health or vital statistics may have on government intervention policies and distribution of financial resources highlight the relevance of the problem. In this paper, we deal with the issue of underreporting, paying particular attention to its effects on the estimation of the prevalence of a phenomenon. We propose a non parametric compound Poisson model that allows for the estimation of the reporting probabilities. The pro-posed model will be applied to a data set concerning early neonatal mortality in Minas Gerais, Brazil. Comparisons of the estimates obtained under several alternative models reveal that the proposed approach is accurate and particularly suitable when there is no prior information about the reporting probability.