Because the assumption of normality is crucial for many types of statistical inference, there are hundreds of test for normality in the literature. Comparison of tests via simulations and other methods has not proven very fruitful or informative, because each test has its area of strengths and weaknesses. The comparisons depend critically on the alternatives, which cannot be specified. We argue that the only way to get unambiguous comparisons between tests is to evaluate them with respect to a specific purpose. One important goal of tests for normality is to ensure validity of statistics used in inference in regression models. In this paper, we evaluate tests for how well they do in terms of ensuring validity of the t-statistics used for assessing significance of regressors. For this, we have explored 40 distributions to find the most damaging ones for the t-statistic and then explored for the best test against these distributions. Power study results yields that Anderson-Darling statistic is the best option among the four tests, Jarque-Bera, D’Agostino and Pearson, Anderson-Darling & Lilliefors.