A Comparison of RAS and Entropy Methods in Updating IO Tables

Since the first half of the 20th century, the input-output (IO) table has been the backbone of much empirical work used to support policy analysis and develop economy-wide models. The need for accurate, up-to-date IO tables is thus essential for establishing the validity of the empirical work that follows from them. However, the construction of an IO table for any given country is an expensive and time-consuming endeavor. Current and accurate IO tables for many countries are thus often difficult to obtain on a regular basis. Once an initial IO table has been constructed, a common workaround is to collect partial information for subsequent periods, such as final demands for commodities within the economy, and then employ a Bayesian parameter estimation technique to determine values for a new IO matrix using the previous period IO table as a prior. Two such techniques to achieve this are RAS and Minimum Cross Entropy (CE). The literature has largely ignored the question of the relative merits of these two methods. This paper uses the actual IO tables for South Korea from two distinct time periods to compare the accuracy of the RAS and CE methods. The 1995 IO table for Korea is updated to 2000 using column and row totals from the true 2000 IO table using both RAS and CE methods. The estimated IO tables are then compared to the actual 2000 IO table in order to make some observations on the relative accuracy of the methods. The sums of squared deviations of the estimates tables from the true tables are used as the main instrument to measure deviations of the updated matrices from the true year 2000 IO matrix. It is found that the CE approach is more accurate than the RAS approach, based on the lower summed squared deviations of the elements of the CE estimated 2000 matrix from the elements of the true 2000. The maximum absolute differences between the true and estimated tables were also calculated. It was found that the maximum absolute difference between CE-estimated table and the true posterior table was smaller than the difference between the RAS-estimated table and the true posterior.