A Monte Carlo investigation of the accuracy of two models for validity generalization

A Multiplicative Model for the generalization of validity developed by Callender and Osburn (1978) was tested by Monte Carlo methods. The model provides a method for estimating the mean and variance of distributions of unattenuated unrestricted population validities by removing the effects of range restriction, criterion unreliability and chance sampling error. The model generally builds on concepts previously outlined by Schmidt and Hunter (1977). The accuracy of the model was compared with the model originally proposed by Schmidt and Hunter on hypothetical infinite sample size cases and on small sample size cases (N= 30, 68, 200) by means of computer simulation of sample restricted attenuated data sets. The correctness of the computer simulation was verified analytically and empirically. Both models were than applied to estimate the known mean and variance of the population distributions of unrestricted unattenuated validities. The Multiplicative Model was found to give reasonably accurate estimates of the mean and variance of the population correlations and of lower credibility values. The errors made by the model were generally conservative, tending to overestimate the variance of the population correlations. The Schmidt-Hunter model was less accurate and in some conditions made substantial nonconservative errors. It was concluded that the Multiplicative Model could be recommended for future use in validity generalization analysis. A Monte Carlo study of the accuracy for a formula for the standard error of a correlation was also conducted. Results indicated that it was nearly as accurate as the more familiar Fisher's z formula. The Multiplicative Model was applied to four distributions of validities summarized by Ghiselli (1966). It was found that the variation in true validities was greater in each case than that previously reported by Schmidt and Hunter (1977). The analysis indicated that validity generalization was supported for two of the four distributions.