That’s a great question Megan and I know I often give the “before we had computers” argument and it’s not that satisfying.  And I agree that we shouldn’t just see them as an alternative when theory-based methods don’t work, but as a reasonable approach on their own with large and small sample sizes. But you can say they are still what is predominantly done in practice and you can still talk about test statistics as giving additional information beyond the p-value.  Because I still tend to do fairly traditional confidence intervals, then the theory-based p-value and test statistic give me a nice transition there (though I often talk more in terms of 2SEs in both cases and don’t worry too much about the exact multiplier/critical value).  I think I would be pretty happy teaching a course (like the CATALST folks do) that does very little with the theory-based methods, but I don’t think my client departments would be very happy with me yet.  Mostly I want students to understand what the p-value and confidence interval mean and then turn things over to a computer (and to be honest maybe not always know the exact details of what the computer is doing) to get the output and focus on the conclusions we can draw…

It is nice to show them two different reasonable approaches to help contrast them. In fact, maybe we should be doing more with exact methods in place of both of these?

Beth

From: sbi-bounces@causeweb.org [mailto:sbi-bounces@causeweb.org] On Behalf Of Olson Hunt, Megan
Sent: Friday, January 02, 2015 11:13 AM
To: 'sbi@causeweb.org'
Subject: [SBI] Explaining motivation for theory-based models

I’m working on transitioning our undergrad introductory statistics course to one that involves simulation-based ideas alongside parametric, theory-based models. At the risk of sounding naïve, I would like your opinion on the following: With the advent of computers that make simulation-based p-value calculations fast and easy, and for those of you that teach this idea in parallel with theory-based p-values, how do you explain the motivation for even using theory-based p-values to your students?

In other words, parametric models require assumptions that may not be met. If we can use simulation to obtain a reliable p-value without worrying about these assumptions, then why bother with a t-test (e.g.) for a p-value at all? It seems like a historical argument (“It’s been done and probably will continue to be done for quite some time”) is one, but… I’m left feeling like I have to tell my students to conduct these theory-based tests “just because.”

Thanks for your thoughts.

Megan