Testing Inference in Accelerated Failure Time Models

We address the issue of performing hypothesis testing in accelerated failure time models for non-censored and censored samples. The performances of the likelihood ratio test and a recently proposed test, the gradient test, are compared through simulation. The gradient test features the same asymptotic properties as the classical large sample tests, namely, the likelihood ratio, Wald and score tests. Additionally, it is as simple to compute as the likelihood ratio test. Unlike the score and Wald tests, the gradient test does require the computation of the information matrix, neither observed nor expected. Our study suggests that the gradient test is more reliable than the other classical tests when the sample is of small or moderate size.