How do I analyse paired comparisons with an internal control

[#1]

Hello,

I am relatively new to JMP (currently using JMP 15) and I am struggling to identify how to analyse paired comparisons where each test subject has a control measure and an intervention measure at every timepoint.

The analysis being performed is looking at the change in continuous variable measured on the face over a defined time period, we know from experience that this variable fluctuates over time regardless of intervention therefore the face is split by left and right and the intervention is applied to one side of the face and the other side kept untreated as an internal control. Therefore each subject has 2 measures at each timepoint "intervention" and "control" and then is measured at baseline and T1 after a predefined length of treatment.

I have included a dummy set of data as an attachment that shows how the raw data is presented.

I want to perform analysis that will tell us if there is a significant difference between control and baseline measures within treatments and then a significant difference in change over time between treatments.

When I use the matched pairs analysis method I get the comparison for each treatment back to baseline but I cannot see a way to also get the intevention vs control analysis as the data we are working on is confidential.

Should I be using the fit model analysis for this comparison instead, if so how should I set up my model to get a paired comparison rather than an unpaired analysis?

@Emmanuel_Romeu 

[#2]

First of all, this is the JMP Community, so please provide examples in a JMP data table.

Please see this technical note about such analyses. Of course, come back with more questions if it does not suffice.

[#3]

Hello Mark,

Sorry I am working from home today and therefore do not have access to JMP which is why my dummy data is in excel format, I will update the data table as JMP format when I have access to the software again.

Can you just confirm that my understanding of the technical note is correct that for my example I should attempt to use a multivariate approach where the T0 and T1 measures are in separate columns but aligned by row for each subject and treatment type and repeated measures is selected? Therefore my understanding is there is no way to perform the analysis I want when all the Time points are in the same column and assigned to a specific time point by a variable in another column, is that correct?

[#4]

Matched pairs tests are generally based on the difference, not the individual responses. Select the control measure data column and the intervention measure data column, right click at the top of one of them, and select New Formula Column > Combine > Difference or Reverse Difference, depending on the order of these two columns in your data table. You want to subtract the control measure from intervention measure. Now use the difference data column in the Y role.

Does that approach make sense? Is it appropriate for your purposes?

[#5]

Hello Mark,

Sorry for not replying sooner this is the first day I have had access to JMP since your response. I appreciate that matched pairs are based on the differences, seeing as my raw data has all of the responses in a single column and then a column defining which time point a row is and a second column defining whether the response is from treatment or control I am trying to get my head around how much pre processing of the data I will need to do before running the statistical comparisons.

I now have 3 different JMP examples of my data which will hopefully demonstrate the issue I am having a bit better.

Dummy 1 - Response data in single column and therefore matched pairs or manova are not possible. I have saved a fit least squares full factorial comparison of timepoint and treatment which gives me p values for T1Control v T0control and T1invervention v T0Intervention. However these are unpaired comparisons and not appropriate for my data set.

Dummy 2 - Response data split into 2 columns T0 and T1 with a 3rd column defining treatment type. I have saved a manova repeated measures which when I look at the within-subjects section I think I understand demonstrates a significant change from baseline across the whole data (treated and control together) and a significant difference between time effect for treated and control. But the data does not give me separate p values for control T1vsT0 and treated T1vsT0.

Dummy 3 - Response data split into 4 columns T0control, T1control, T0intervention, T1intervention. I have saved a matched pairs analysis using all 4 columns which gives me the separate paired analysis p values for control T1vsT0 and treated T1vsT0 but does not give me the anlysis of whether the mean difference T1-T0 is significantly different between intervention and control.

I am sure I am missing something blindingly obvious but what I am looking for is a way to get the difference in time effect between treated and control from the Dummy 2 manova and and the difference between timepoints for each treatment from Dummy 3 all in one set of analysis.

Hope this makes things a little clearer.

[#6]

I think I have found my solution from Julian's first post in response to this question from 2017 Solved: Interpreting MANOVA repeated measure - JMP User Community

I have run a new fit least squares with random subject effects nested in the treatment, using a minimal report with REML  as the method and this seems to give me my difference between treatments in the Fixed Effect tests and also my T1vsT0 for each treatment in the timepoint*treatment section.

I have reattached the file Dummy.jmp with this second analysis saved. Have I interpreted it all correctly?

[#7]

See Tables > Split. This command will create two new columns of response values based on the levels in your column indicating control or treatment.

[#8]

Hello Malcolm,

for repeated measure, Mixed Model (JMP Pro) is the appropriate analysis. I've attached 2 analysis, Residual and Unstructured which test various covariance structure, the lower AICc should help decide on which is the most appropriate (as Dummy variable here I did not conclude). Emmanuel

[#9]

Hello Emmanuel, thank you for the additional information. It looks like I have some more evidence for my stakeholders as to why JMPpro would be a valuable purchase as most of our test data comes in this sort of format.

[#10]

Hello Emmanuel,

I have had a look at the data and it looks to me like the analysis you have shared is doing unpaired comparisons between T0 and T1 have I interpreted that correctly? For my 8 intervention measures I made sure that every one was an improvement but that there was a wide variation in the baseline starting point as that is what we often see in our data sets. My 8 subjects had improvements of -11, -5, -5, -8, -6, -7 and -3 but the timepoint*treatment intervention T1 vs intervention T0 give a p value = 0.71. I would have expected for a paired comparison to see the p value much closer to the 0.05 threshold with the data as it has been created.

Kind Regards

Malcolm