Steve Fienberg has presented a wide and interesting range of applications of Bayesian methods in public policy and government settings (including election night forecasting which I might prefer to classify as fleeting public entertainment!). The examples exhibit the common feature that they all involve highly complex problems that are difficult to handle in a non-Bayesian framework. Sedransk ( 2008) has provided some other examples of Bayesian methods in such settings which also share this feature. I am sympathetic to the use of Bayesian methods in such special circumstances, as illustrated below.My initial comments focus on the choice of modes of inference for large-scale government surveys, particularly surveys of households and persons, that are the backbone for satisfying policy and government data needs. An important feature of these surveys, in common with most surveys, is that they typically collect data on many variables and these data are then used to produce very large numbers of estimates. In this area, I generally favor the frequentist repeated sampling mode of inference, commonly termed design-based inference (Kalton, 2002), and I believe that my views are in line with most other survey statisticians (see, e.g., Rao, 2011, in this issue). However, there are situations in which designbased inference cannot satisfy analytic objectives. Also, limitations in the practical application of design-based inference are becoming increasingly troublesome. To the extent possible, I prefer to minimize the dependence of survey estimates on statisti-
