With the advance of sequencing technologies, it has become a routine practice to test for association between a quantitative trait and a set of rare variants (RVs). While a number of RV association tests have been proposed, there is a dearth of studies on the robustness of RV association testing for nonnormal distributed traits, e.g., due to skewness, which is ubiquitous in cohort studies. By extensive simulations, we demonstrate that commonly used RV tests, including sequence kernel association test (SKAT) and optimal unified SKAT (SKAT-O), are not robust to heavy-tailed or right-skewed trait distributions with inflated type I error rates; in contrast, the adaptive sum of powered score (aSPU) test is much more robust. Here we further propose a robust version of the aSPU test, called aSPUr. We conduct extensive simulations to evaluate the power of the tests, finding that for a larger number of RVs, aSPU is often more powerful than SKAT and SKAT-O, owing to its high data-adaptivity. We also compare different tests by conducting association analysis of triglyceride levels using the NHLBI ESP whole-exome sequencing data. The QQ plots for SKAT and SKAT-O were severely inflated (λ = 1.89 and 1.78, respectively), while those for aSPU and aSPUr behaved normally. Due to its relatively high robustness to outliers and high power of the aSPU test, we recommend its use complementary to SKAT and SKAT-O. If there is evidence of inflated type I error rate from the aSPU test, we would recommend the use of the more robust, but less powerful, aSPUr test.