Benchmarking methods for repeated business surveys

Benchmarking corresponds to a combination of two sources of information on a given variable. In many situations, the problem consists of combining a series of frequent data with a series of less frequent but more accurate data for producing more accurate estimates of the former series. In what follows we focus in the problem of benchmarking monthly data with annual estimates; then, the benchmarking consists of forcing the sum of the monthly signals to equal the signal of the benchmark. Alternative estimators have been proposed in the literature for benchmarking. When the adjusted series agrees exactly with these benchmarks, the benchmarking is called binding. The binding process is implemented by setting the variance of the annual survey errors to zero. However, it is necessary to account for the variance of the annual survey errors when computing the variances of the benchmarked estimators. In this thesis, we develop the theoretical expression of the correct variance as well as an expression for the excess in the variance due to the binding process. The results are extended to the most known benchmarking methods proposed in the literature. An application to business surveys used for official statistics in the UK is presented, illustrating some particular issues regarding the state space modelling. Finally, the problem of how to prepare tabular data classified by attributes as columns and points in time as rows in analyzed. This multivariate extension of the benchmarking problem distinguishes two basic type of problems: when only marginal totals are available (contemporaneous disaggregation) and when the aggregates do not correspond with the sum of the disaggregated values by year and/or by attributes (reconciliation). The scope of this thesis is based basically in a state space model approach.