Discussion on the Paper From Start to Finish: a Framework for the Production of small area official statistics” by Tzavidis et al.,

I strongly welcome this paper which sets out a framework to produce small area official statistics. Tzavidis and his colleagues have proposed an iterative process on three broadly defined stages which are specification, analysis and adaptation, and evaluation. Keeping in view the fast developments in the area of machine learning I strongly suggest using these tools at the specification stage. Whereas machine learning tools have been successfully used in many areas I am a little surprised that the application of these tools in small area estimation is still at its infancy. Machine learning methods like classification and regression trees, random forests and stochastic gradient boosting (‘TreeNet’) have several advantages over ordinary least squares such as that they can handle outliers, missing values and model non‐linear relationships and local effects which are common problems in small area estimation. These methods are quite efficient in selecting variables and modelling variable interactions. They are also particularly useful for unbalanced data sets. They can be used to select optimal predictors of the target population under the model and can yield parsimonious models.