Applied researchers are often interested in obtaining confidence intervals for key nonlinear model parameters so as to answer important research questions, and the usual "plus and minus 2 SE's" confidence interval leads easily into the usual Wald hypothesis test covered in most introductory statistics courses. However, since information about a specific parameter is often asymmetric, a skewed confidence interval is often more appropriate and reasonable in practice. This leads to the use of likelihood-based tests, typically introduced in intermediate undergraduate and basic graduate course. This paper argues that the superiority (in terms of for example increased power) of likelihood-based and score hypothesis tests over the Wald test is most easily conveyed and appreciated by first providing a reasonable motivation (as well as examples) using confidence intervals, and then exploiting the confidence interval-hypothesis test equivalence.
The CAUSE Research Group is supported in part by a member initiative grant from the American Statistical Association’s Section on Statistics and Data Science Education