Sampling error is one of several types of error in econometric modeling. The relationship between sampling error and sample size is well known for both estimation and prediction. The objective of this paper is to provide an empirical foundation for using these relationships to guide researchers and planners in the determination of sample size for model development. Analytic relationships are formulated for sample size, precision of parameter estimates, replication of parent population, and replication of an alternative (transfer) population. Application of these relationships to an empirical case indicates that the sample sizes required to obtain reasonably precise parameter estimates are substantially larger than the sample sizes generally considered to be needed for disaggregate model estimation. Nevertheless, these sample sizes appear to be adequate for obtaining reasonably accurate replication of observed choice behavior in the parent population. The corresponding results for prediction to a different population are complicated by the issue of intrapopulation transferability. Although the results reported in this paper should be validated in other contexts, it appears that accurate estimation requires the use of samples that are substantially larger than formerly believed. Samples on the order of 1,000 to 2,000 observations may be needed for estimation of relatively simple disaggregate choice models. Although some reduction in this requirement may be obtained by improved sample design, it is unlikely that the final sample requirements can be reduced to less than 1,000 observations.
Abstract