CentER

Jan van Tongeren

Date of PhD defense:	14 October 2011
Title of thesis:	From national accounting to the design, compilation, and use of Bayesian policy analysis frameworks
ISBN:	978 90 5668 295 8
Promotor:	Prof.dr. Jan Magnus

Abstract:

The thesis is a selection of articles dealing with two interrelated topics, i.e. the design of data frameworks and the compilation of estimates within those frameworks. The articles included are a selection from a larger number of essays, referred to in the Introduction - the results of forty years of research. The frameworks typically include variables related to the System of National Accounts (SNA) and the compilation consists of a Bayesian formalization of conventional compilation techniques developed in SNA related frameworks. While focusing mainly on SNA related frameworks, an example in the Introduction shows how such frameworks could be developed for any data framework, also including non-SNA variables, as quantitative support to any policy analysis study.

The first article deals with the analytical features of SNA frameworks, and the role of classifications therein. The second article extends the framework to environmental satellite accounts, in which the variables reflect the interaction between economic development and environmental impacts. The last three articles develop a Bayesian approach of compilation estimates in frameworks, in which conventional compilation techniques used in SNA related systems are formalized. The approach is based on the use of available basic data, definitions and values of ratios between variables of the framework, definitional and other identities that should hold between the variables and prior reliabilities assigned to basic data and ratio values; the posterior estimates include estimates of variables and ratios, and also posterior values of reliabilities. The first of the three articles presents a linear programming (LP) approach, in which the prior reliabilities are defined as intervals in which the basic data and ratio values should lie. In the second article on compilation, the fixed prior intervals of prior reliabilities in the LP approach are replaced by a normal distribution that govern the possible values of basic data and ratios. It is based on a simple example of approximately 40 variables, in which the method has been developed. In a third and last article included in the thesis, the number of variables has been increased to a realistic framework of approximately 2500 variables.

Full text