turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

- JMP User Community
- :
- Discussions
- :
- Discussions
- :
- Pairwise comparison on interaction term

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Oct 12, 2018 4:24 AM
(323 views)

I would Like to know if Is it possibile and how to perform pairwise comparisons on an interaction term when this one Is between a categorica and a continuous variable. In Jmp I found only multiple comparison for the mai effect. Thank you.

5 REPLIES

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Oct 15, 2018 10:48 AM
(279 views)
| Posted in reply to message from Sarettathepooh 10/12/2018 07:24 AM

Hi @Sarettathepooh,

Can you explain a little bit more about what kind of comparison you are interested in making? When you have one continuous and one categorical factor, the interaction term is a test of whether the differences among the levels of your categorical factor differ at different levels of your continuous variable (or alternatively and equivalently, whether the slope between Y and your continuous X variable varies at different levels of your categorical factor). For instance, in a model predicting Weight from age, height, and age*height (from Big Class.jmp), we can see the interaction both ways with the prediction profiler.

Now, to your point about pairwise comparisons, if you wish to compare levels of your categorical factor, you would do so under the main effect Red Triangle. Those tests would be of the mean response for each level of the categorical variable at the mean level of the continuous variable. In a model with an interaction, it becomes important to ask at what level of the continuous variable those comparisons are made at. The mean is the typical default (and this is what JMP will do unless told to do something else). If what you're looking for is a comparison of the mean responses for the different levels of your factor, at something other than the mean level of the continuous variable, you won't find that test under the interaction term; rather, you find that under the same main effect red triangle, but using the LSMeans Contrast command. In the gif below, you can see me comparing the 12-year-olds and 13-year-olds, but at 70" for height.

I hope this helps!

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Hi @julian,

This is very instructive thank you for providing this and for your explanation! I did not understand how specified contrasts could be used in this way - that is, specifically in the context of one continuous and one categorical predictor versus a continuous response.

I've watched some of your previous online content (https://youtu.be/3skUZZF-CuE) where you discussed the general utility of running specified contrasts. Could you attempt to summarize the practical advantages (and disadvantages) here in this forum?

I'm intersted in understanding better how I can use contrasts within my LS regression models (run in Fit Model) to test for specific comparisons, and why - if at all - making these comparisons gives me anything different or more valuable than simply interpreting the results of my standard multiple comparisons outputs? (Tukey HSD detailed pair-wise comparisons and/or connecting letters reports) in Fit Model?

Thanks, @PatrickGiuliano

If you are at a university other than UCSD and have found this or any of my other videos to be useful, please do me a favor and send me a note at ProfessorParris@gmail.com indicating your university affiliation and which videos you've found useful. Thank you! - Dr. Julian Parris --- Tutorial on ...

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Dec 6, 2018 10:31 AM
(110 views)
| Posted in reply to message from PatrickGiuliano 12/05/2018 11:50 PM

Hi @PatrickGiuliano,

Absolutely! I keep intending to write a blog about this, so what I did is recorded a quick tutorial introducing some of the ways I use contrasts (which I can do more quickly), and I'll carve out some time to start working on a blog that covers these topics in some more detail. Hope this is a helpful start!

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Hi @julian Yes this is great thanks, definitely a great start and I look forward to the Blog post. I hope you don't mind if I follow-up with some further questions as I study and reflect on your commentary in this video. :-)

For starters: I'm curious about about the US Demographics example you showed in the context of comparing S and W vs. MW and NE regions using a specified contrast, vs. simply re-coding by lumping S/W and M/NE into Group 1 and Group 2 respectively and then running an ANOVA. A couple things I observed were a little puzzling to me: why is it that the estimate of the mean difference is 7.2929 in the Contrasts Panel, but in the ANOVA it is 7.310? I know the difference is very very small, but is the mean being calculated slightly differently here?

I know you mentioned the mean is NOT changing, its the Variances that are being calculated differently! And to that point, in this example we observe an F Ratio from the Contrasts panel fo 16.5530 vs an F Ratio of 16.6194 using the ANOVA on the recoded (lumped) n of 2 groups. I see the MSE in the ANOVA output, it is 40.126. I don't see the MSE for the contrasts test in Fit-Model, but anyway my question here is: Shouldn't the F Ratio for the contrasts test be *larger* than for the ANOVA because of more precision on the estimate of MSE in the denominator (and therefore a *smaller* MSE)? In this case it is the opposite of what I thought it would always be: smaller for the contrasts test as compared to the ANOVA (I know the difference numerically is practically negligible but it's still making me question my understanding of what you're fundamentally asserting). Thanks!

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Dec 12, 2018 7:26 AM
(51 views)
| Posted in reply to message from PatrickGiuliano 12/07/2018 12:09 AM

Hi @PatrickGiuliano,

Great questions!

Estimate of mean difference: you're right, the estimate of the mean differences there is a little different, and I glossed over a complexity here. When you do not have equally sized groups, as is the case here (there are different numbers of states in each of the regions), the least squares means in an ANOVA model will differ from the means that you would calculate by summing up the observations in the group and dividing by the number of individuals in the group. You can read about the estimations of LSMeans in the JMP Docs, or elsewhere, but most briefly, they're predicted population margins, estimates of the marginal means in a balanced population. So, that's the origin of this difference there - when we recode the groups and calculate a difference based on those LSmeans, that will be different from first estimating the LSmeans and then averaging those. If the group sizes were all equal, there would be no difference.

This is a regrettable inconsistency you've noticed in my poorly chosen example :) Even though the MSE is smaller in the example using the contrasts, as it should be based on what I was saying (39.954 vs 40.126, and you can find this MSE in the regular ANOVA table -- a contrast uses the model's MSE) because the mean difference is slightly larger in the example where we recoded (for the reason above), the F Ratio ends up being larger in that example. Not the best example, I really should have used a dataset where the groups are equally sizes, or one where the benefit conferred by using the contrasts rather than recoding was more substantial (which is usually the case). I'll certainly choose a different example for the upcoming blog. Thanks for so carefully watching this video!