“…For example, Copas and Eguchi (2005) discuss a similar issue that they term as incomplete-data bias , in which the maximum likelihood estimators can be (sometimes severely) biased when incomplete data are present, and an incorrect model is being fit, and yet still appears to give a good fit to the available data. Jiang et al (2011a) showed that if one derives the parameter estimators by evaluating the best predictor (BP) under the assumed model, say, M , using the distribution also under M , the resulting predictor is not robust in the sense that it may perform poorly when M is not the true model. Here, the failure of the BP is due to a similar double-dipping strategy, that is, (1) the measure of lack-of-fit (sum of squared prediction errors), is for the BP under M ; and (2) the distribution under which the measure of lack-of-fit is evaluated is also under on M .…”