Small area estimation with R
We use a form of small area estimation in which survey data is carefully modelled, and the parameter estimates applied to national, small area administrative data. As a result we can produce a detailed map for the survey target variable which is of great value in marketing.
Clearly we cannot create data where it doesn't exist, and note that the mapped results are model estimates rather than actual data. However the maps have great interpretative value. Even relatively simple models fitted to demographic predictors can be difficult to interpret, and when that interpretation is complicated by the actual demographics of actual locations mentally relating the two can be error prone. Conversely, model based small area estimates combine the complexity of the survey model with the complexity of real world demographics to produce an intuitive result of immediate business use.
We use Bayesian methods, and illustrate our estimation on the 6,781 middle level census output areas in England. With 8 predictor variables, geographic random effects we have over 800,000 person types to model. Whilst the statistics are well established, it is clear that considerable computational effort is needed in order to visualise these results.