Previous studies recommend the empirical best linear unbiased predictor (EBLUP) for small-area estimation. However, EBLUP estimation requires at least one observation from each small area, while most of the areas may be non-sampled. One approach to overcome this problem is to predict the area-effects for the non-sampled areas with a model developed using the estimated area-effects from the sampled areas. Another approach is to cluster the small areas to larger groups and introduce a cluster-effect into the prediction model. We tested these approaches in a set of simulated small areas (domains). When observations from all or most domains were available, EBLUP with a domain-effect or combined cluster- and domain-effect were the most reliable calibration methods. When the sampling fraction and the size of the domains were smaller, calibrating with the cluster-effect only was the most reliable method. Without any calibration, the model-based estimates for the domains with the highest volumes were severely underestimated. When observations were available, the EBLUP calibration improved the results in the high-end of the distribution. With the smallest sampling fractions and domains, also the predicted area-effects reduced the underestimation. However, the modelled area-effects were estimated from the population data, rather than from a sample.
Kangas et al. (Thu,) studied this question.