Using satellite ice water path (IWP) retrieval as an example, two types of systematic errors when using a Bayesian algorithm are documented. These errors arise from the discontinuous and heavily skewed data distributions of IWP in the supporting database. Discontinuity in the data distribution occurs at IWP ¼ 0 since there are no negative IWPs, which causes a positive bias in the IWP retrievals near icefree conditions. On the other hand, since clouds with small IWPs occur more often than clouds with large IWPs in nature, data points in the database are heavily populated in the low IWP range if we build the supporting database based on observed data. This skewed data distribution leads to a negative bias. A number of remedies, while some are ad hoc, to reduce these errors are also suggested.