I would be very grateful for the opinion of someone, with more experience of mixed-effects models than myself, on whether the following sounds reasonable:
I would like to look at whether several variables, such as blood test results, significantly change among diabetic cats as they enter diabetic remission. I have the variable results from a selection of diabetic cats who successfully entered remission, and also from a group of diabetic cats who never achieved remission. To analyse my data, I was planning to split the results from the remission cats into several groups according to their "stage" of remission. These groups are:
1) more than 45 days before stopping remission 2) less than 30 days before stopping insulin 3) less than 28 days after stopping insulin 4) more than 28 days after stopping insulin (in long-standing remission) There is also a group 5 made up of the cats who never entered remission.
I was then planning on using a mixed effects model to study the effect of Group (my fixed effect) on my various outcome variables. I was going to add each cat's identity as a random effect. All the cats in group 5 are unique, but cats who entered remission might contribute samples to any of groups 1 to 4. Also, some cats might have several samples that are eligible for one, or several, of groups 1 to 4 e.g. a cat might have 2 samples that could technically be included in group 1. Unfortunately, not all remission cats have a sample in every one of groups 1 to 4.
I would be very grateful if anyone could advise me on whether the above seems suitable, and whether it would be possible to include multiple samples from an individual cat in any of groups 1 to 4? I would also be grateful to know which covariance structure might be most suitable.
Thank you very much in advance!
I think this should be fine for mixed effects; supposedly one of the advantages of mixed effects is that it's good for handling missing cases, unbalanced designs, stuff like that.
Not sure about the covariance structure issue; I'm lazy and do my mixed models with lme4 in R, which pretty much leaves all that stuff under the hood.
Your idea of using the cat as a random effect works fine. Your linear mixed model will be: y = mu + Group + cat + e, where Group is fixed effect and cat random effect, and you can have a single error term. Here, the random cat effect allows you to consider the fact that you have more than one measurement of a given cat, and in fact will 'correlate' the observations that belong to the same cat. This is a typical analysis with several more measurements of a subject and is the correct analysis. And also will take care of your lack of balanced (particularly in group 5 where cats are measured a single time). Now, this corresponds to a compound symmetry structure and that is fine.
If you really want to get more complex on your analysis, you can add heterogeneity of variances, in this case, you can make your errors different for each group. This can be reasonable if you expect that each group will have different error variance. Something that makes sense given that some groups will respond differently than others.
Hoep this helps.
Dear Stephen and Salvador,
Thank you very much for your replies. Salvador, I am using SPSS, which offers linear mixed models, but not non-linear mixed models. I was concerned that I couldn't carry out a linear mixed model on my data because the cats in group 5 do not have repeated measures. When I create my model, is "group" still a repeated measure, despite this?
Best wishes,
Ruth
- Badges
- Science topic
- Similar topics
- Medicine
- Disease
- Metabolic Diseases
- Glucose Metabolism Disorders
- Diabetes