Comparison of Benchmarking Methods with and without a Survey Error Model

Abstract

For a target socio-economic variable, two sources of data with different precisions and collecting frequencies may be available. Typically, the less frequent data (e.g., annual report or census) are more reliable and are considered as benchmarks. The process of using them to adjust the more frequent and less reliable data (e.g., repeated monthly surveys) is called benchmarking.

In this paper, we show the relationship among three types of benchmarking methods in the literature, namely the Denton (original and modified), the regression, and the signal-extraction methods. A new method called ''quasi-linear regression'' is proposed under the multiplicative assumption. The numerical Denton method is currently widely used. The aim of this paper is to promote the other two methods which are statistically model-based; the model for the survey error is assumed to be known. Assuming the survey-error series follows an autoregressive model of order 1, by simulation, we investigate the impact of misspecification of the model on the benchmarking prediction according to the criterion of minimizing the root-mean-squared error of prediction. It is concluded that both statistical methods have great advantages over the Denton method and they are robust to misspecification of the survey-error model. The problem of how to obtain a survey-error model is also mentioned.