Hi,
I have a question about the options for covariance matrices in Jamovi.
I'm running a LMM with three timepoints. Pre-treatment, mid-treatment, post-treatment.
I'm wanting to model the interaction between a baseline random effect and Time (fixed categorical) for the dependent variables.
I've been told that Jamovi's default covariance matrix does not allow correlations to vary between timepoints, and given that these are likely to vary and change will not be linear I cannot use Jamovi's default covariance matrix to model data with three timepoints.
Is this correct?
If so, is there an option to select an unstructured covariance matrix in Jamovi so that the correlations between the different time points are estimated/can vary?
Thanks!
Covariance matrix for LMM
-
mcfanda@gmail.com
- Posts: 462
- Joined: Thu Mar 23, 2017 9:24 pm
Re: Covariance matrix for LMM
Hi
there are a number of issues here. First, the "unstructured covariance matrix" is the default choice in jamovi LMM. This is equivalent of choosing "unstructured" in SPSS LMM or fitting the same model in R glmer(). However, "unstructured covariance matrix" (in any software) does not pertain the correlations among time points, but the correlation among random coefficients. The correlation among the three points measures regards the residual covariance matrix.
Second, you can look at the correlations among time points instead of assuming that they are likely to be different. Consider that the issue of different correlations arises when correlations are dramatically different, otherwise it has very little impact on the results.
Third, no issue arises if changes are not linear, because time is a categorical variable and thus the estimated effects can be of any shape. If the effects of baseline (continuous variable) are expected to be non-linear, you can add these non -linear effects in the model (here an example with GLM but the same logic applies to LMM: https://gamlj.github.io/glm_example2.html)
hope it helps
there are a number of issues here. First, the "unstructured covariance matrix" is the default choice in jamovi LMM. This is equivalent of choosing "unstructured" in SPSS LMM or fitting the same model in R glmer(). However, "unstructured covariance matrix" (in any software) does not pertain the correlations among time points, but the correlation among random coefficients. The correlation among the three points measures regards the residual covariance matrix.
Second, you can look at the correlations among time points instead of assuming that they are likely to be different. Consider that the issue of different correlations arises when correlations are dramatically different, otherwise it has very little impact on the results.
Third, no issue arises if changes are not linear, because time is a categorical variable and thus the estimated effects can be of any shape. If the effects of baseline (continuous variable) are expected to be non-linear, you can add these non -linear effects in the model (here an example with GLM but the same logic applies to LMM: https://gamlj.github.io/glm_example2.html)
hope it helps