Model Selection for High-Dimensional Temporal Disaggregation in Official Statistics

Traditional methods for producing economics statistics, for instance GDP, rely on data gathered through surveys of a population. Whilst such methods are accurate, and well calibrated, they are very expensive to run, and take a long time to feed-back information. As such, National Statistics Institutes such as the UK's Office for National Statistics (ONS) are looking to integrate so-called administrative data, and alternative data-streams such as web-scrapped data into their estimation of economic statistics. Using such data can potentially increase both the frequency and the accuracy at which economic statistics are produced. However, it is often unclear how these alternative data-sources (of which there can be many) relate to the traditional survey results, and how we can produce high-frequency series which are consistent with the survey data.

Given that we could measure many different aspects of the population, only a few of these might actually be relevant to producing a particular statistic of interest. From a methodological viewpoint, this mandates that we choose between several competing statistical models, a problem known as model selection. Traditional model selection methods assume that the number of data-points is much larger than the number of data-streams, however, when linking administrative, and alternative data-sources, that assumption will no longer hold and one has to consider the so-called high-dimensional statistical setting. This project proposes to adapt recent advances in high-dimensional methodology to the analysis and production of bench-marked economic statistics. The project aims to examine both the empirical behaviour of these methods via simulation, and work with practitioners at the ONS to implement and test these methods through the development of a easy to use software package.