Reduction of sample sizes in network sampling models

Abstract

This paper develops and analyzes some statistical sampling methods suitable in a class of telecommunications applications including network cost calculations. The main objective is to obtain the smallest possible sample from a population in such a way that a prespecified desired level of accuracy of the estimated cost is achieved. The cost is assumed to be a function of a number of variables such as CCS and Message. Principal component analysis is used first to reduce the dimensionality of original variables. On the basis of the first few principal components, a graphical analysis is performed to determine the number of clusters, and then cluster analysis is employed to decompose the population into nine nearly homogeneous groups. We have also studied the performance of two sampling designs. Compared to the simple random sampling without any preliminary analysis, the stratified random sampling method with the proposed preliminary analysis provided a considerable reduction in sample size. In a typical application that we tested, the sample sizes required to guarantee the specified accuracy were 1 692 and 359 for simple random sampling and the proposed stratified random sampling, respectively.