Assistance in use of mixed-model analysis

[#1]

Hi there,

I'd like some assistance/comments on how to perform a mixed model analysis in JMP, as it seems to be the most suitable for my data.

I'm looking to compare within-group differences in pre- and post-sleep test scores between two groups (patient and healthy controls).

The steps I did are as follows:

-Added test scores (both pre and post) under Role Variables (Y)

-Added the group variable under fixed effects (0 and 1 for controls and patient group respectively) 

-Added the individual ID under random effect.

Is the above correct? Also, which covariance structure is most appropriate? The use of this modelling is new to me, so I'd appreciate any assistance.

[#2]

Hi @OddsMouse312534,

Welcome in the Community !

If I understand your topic well enough, you're interested to see if treatment group benefit from higher/better post-sleep test score than the control group, taking into account pre-sleep test scores (that may be different across individuals and groups). Without indications on how data was collected (which would guide the appropriate modeling), it is very difficult to come up with a solution.

There may be several ways to analyze your data. Here are some ideas, some options may be irrelevant to your topic, so you can choose which one(s) may seem the most appropriate based on your domain expertise and how the data was collected (random/DoE ?) :

1. You can analyze the difference between your post and pre sleep test scores, so that the initial value (pre score) won't affect the results, you would only analyze the difference in sleep score depending on the group in which the individuals are.
2. What may also be possible is to use the pre-test sleep score as a random effect, group variable as fixed effect, and post-sleep score as the response Y. In this way, you account for variability in the pre-sleep test score across individuals, and you are interested to see how the final post-sleep test score may be different depending on the group (control vs. treatment).

If there are no individuals both in the control and treatment group, I don't see the point of using ID as a random effect (as the option 2 may already take into account variability of pre-sleep test score across individuals). Option 1 may not require random effects, as only the variation between scores is analyzed (so random influence of the individuals may already be taken into account in the difference of the scores).

You can also check the datasets provided by JMP in the "Mixed Models" section (go to "Help", then "Sample Index" and look for section "Mixed Models"). Depending on how much information you may have on the individuals, there are some datasets that can help you figure out how to deal with the variables : Cholesterol Stacked.jmp for example to deal with time, patients and groups, with a nesting of Treatment in Patient as random effect. Growth Measurement.jmp is a good example to deal with patients evaluated at different age (repeated structure).

These are some quick suggestions and ideas to start the discussion, if you can provide more context about data generation, objective of the study, number of participants, etc... and possibly a short anonymized dataset, that would be helpful.
I am sure other members of the Community may jump into the discussion with more infos provided.

Hope this first discussion starter may help you, 

[#3]

Hi Victor,

Thank you very much for your reply. If I may give elaborate further: 

The study involves a patient group with a specific disorder, matched with a group of healthy controls. The purpose of the study is to evaluate the effect of sleep on the results of a cognitive test, where the testing happens two timepoints: at baseline (pre-sleep) and the morning after (post-sleep). There is no specific intervention, the purpose is to evaluate if there is a "between group" difference in change in scores, as to measure the effects of sleep on the test results and see if there is a significant difference between groups. Total group size is 48 for patients, 45 for control group. I've compared the means of delta (difference between scores pre- and post sleep) between groups with an independent t test, but I purport this is insufficiënt to capture the whole picture. This is why I looked at a mixed model: as there are individual patients and controls that were measured, I wanted to use ID as a random effect.

[#4]

If I may I add: I was wondering if a mixed between-within ANOVA was applicable in this situation. The problem is that I don't know how to include the time-points into the model in JMP (for the repeated effect), i.e. how to include this in the database. All I have in the database at the moment is ID of the participant, the condition (0 for control, 1 for patient) and the pre and post test result.

[#5 · Accepted solution]

Hi @OddsMouse312534,

You may find this JMP note interesting to compare the different modeling options available : Analyzing Repeated Measures in JMP® Software - JMP User Community

I think you may have a very similar data structure to what is presented in the Cholesterol Stacked.jmp datatable (available in the sample index or in copy of this message), so this section of JMP Help may help you to structure your datatable and run the appropriate modeling : Example of Repeated Measures

I did a simple toy dataset on your example with the possible Mixed analysis available if you want to have a look on a more concrete and similar use case : Test_MixedModel.jmp

Again, choosing the most appropriate model is a question of "best fit", but also how the data has been structured and collected. Different repeated measurements structures are possible in Mixed Models, and should be evaluated based on the statistical results and the way the study has been conducted.

Hope this answer may help you,