Interacting biases, non-normal return distributions and the performance of tests for long-horizon event studies