DETERMINATION OF SAMPLE SIZE TO ACHIEVE AGREEMENT BETWEEN TRUE AND REPORTED SIGNIFICANCE LEVELS FOR ADJUSTED WALD STATISTICS (CATEGORICAL DATA)
Little has been known about the small sample properties of the Wald statistic used in the weighted least squares approach for categorical data analysis. This paper explored the bias in the estimation of variance using this procedure and identified situations in which there would be a large bias. An enumeration and simulation study was conducted to compare several adjustment procedures in terms of their small sample properties and then provided guidelines for the sample size required for the test statistics to be well behaved. The logarithmic transformation is found to have a problem when extreme cell proportions are encountered and the sample sizes are small. Both the subpopulation total (n(,i.)) and the minimum cell expectation (MCE) must be considered in establishing sample size requirements. The procedure of adding 1/2 to all cell frequencies for the additive model and Bhapkar's 1984 procedure for the log-linear model are found to be the best adjustments for samples as small as n(,i.) (GREATERTHEQ) 15 and MCE (GREATERTHEQ) 2. Other procedures require an n(,i.) (GREATERTHEQ) 25 and an MCE (GREATERTHEQ) 5. Maximum likelihood procedures are also evaluated in some situations.
LIN, TSAI-LIEN, "DETERMINATION OF SAMPLE SIZE TO ACHIEVE AGREEMENT BETWEEN TRUE AND REPORTED SIGNIFICANCE LEVELS FOR ADJUSTED WALD STATISTICS (CATEGORICAL DATA)" (1984). Texas Medical Center Dissertations (via ProQuest). AAI8601786.