Over 75 Alzheimer's disease (AD) and dementia-associated variants have been identified through genome-wide association studies, but the utility of polygenic risk scores (PRS) for predicting AD and dementia in diverse and admixed populations remains unclear. We compared how PRS approaches differing in p-value thresholds, variant weights, and source ancestry perform in predicting dementia in 6338 African American, Chinese, Hispanic, and White individuals from the Multi-Ethnic Study of Atherosclerosis. We tested clumping and thresholding (C+T) methods with varying parameters against Bayesian approaches (PRS-CS, PRS-CSx). We compared the ability of each method to predict incident dementia in all participants and in groups stratified by self-reported race/ethnicity. We additionally analyzed performance across groups stratified by estimated proportion of non-Finnish European (NFE)-like ancestry. Including more variants does not improve performance. We found comparable associations between dementia and PRS when comparing a C+T method with only 15 SNPs and PRS derived from Bayesian models that include > 800,000 SNPs (HR5e-08 = 1.18, 95% CI: 1.08-1.28; HRCSx = 1.17, 95% CI: 1.07-1.27). The p < 5e-08 C+T method was more strongly associated with incident dementia in populations genetically dissimilar from the source data (HRlowNFE_5e-08 = 1.27, 95% CI: 1.08-1.50; HRlowNFE_CSx = 1.12, 95% CI: 0.94-1.33). More selective PRS models using genome-wide significant SNPs may be preferable for dementia prediction in diverse populations.