Date of Award
Campus Access Dissertation
Educational Psychology / Research
Bethany A Bell
Sample size guidelines for binary outcome multilevel models are rare, despite well-established rules-of-thumb for continuous outcome multilevel models. When considering requisite sample sizes for designing studies utilizing such models, applied practitioners would be remiss in failing to take into account other aspects of their models including estimation method, size of variance components, or number of predictors. This Monte Carlo study examined the performance of subject-specific binary outcome multilevel models under varying methods of estimation, level-1 and level-2 sample size, outcome prevalence, variance component sizes and number of predictors. All simulations and results were generated and analyzed using SAS├é┬« commercial software. Although sample sizes at both levels were typically the most influential factors impacting mean estimates of statistical power, confidence interval coverage and width, and likelihood of non-positive definite G-matrices, other key features of multilevel models including the level of random intercept and slope variance and estimation method also impact results. Greater levels of intercept and slope variance tended to decrease mean estimates of fixed effect statistical power, increase mean estimates of variance component statistical power, increase the rate of non-positive definite G-matrices and decrease the rate of variance component coverage. The RSPL estimation method tended to yield more conservative fixed and variance component effect coverage estimates and less variance component bias, while Laplace tended to yield fixed and variance component effect coverage rates below the nominal 95% value and increased levels of variance component bias.
Schoeneberger, J.(2012). The Impact of Sample Size, Prevalence, Estimation Method and Other Factors When Estimating Multilevel Logistic Models. (Doctoral dissertation). Retrieved from http://scholarcommons.sc.edu/etd/1036