Mixed models for association genetics of outcrossing plant species such as maize have been developed recently, but validation of selected markers associated with agronomic traits in different populations has not been extensively studied. Moreover, the mixed models developed for outcrossing may not be appropriate for inbred plants such as rice. Our first research objective was to evaluate the mixed model for selection of markers that explain phenotypic variance associated with four agronomic traits in two genetically diverse and narrow germplasm collections of inbreds and validate results in separate lines not involved in the original analysis. The results showed that kinship estimates incorporated into the mixed model for the narrow germplasm did reduce Type I errors, but this improvement was not sufficient for high phenotypic prediction rates in a separate validation sample. Our second research objective was to create and evaluate a mixed model-regression procedure that identified main and epistatic effects by standard hypothesis testing and Bayesian information criteria in a multivariate format. The new procedure resulted in increased power and enhanced prediction ability of markers evaluated in validation samples from both diverse and narrow germplasm. Addition of an epistatic component to the procedure improved validation results by ~ 18%. The results indicated that a sequential mixed model-regression approach with epistatic effects coupled with a validation step should be considered for association genetic studies in rice.