Logistic regression on longitudinal data

Question

[Aiswarya]

Hai,

I am at the starting stage of trying to analyze differential RNA editing (from a pool of already identified editing sites) among my samples. I have a longitudinal dataset from three different stages of viral infection ( pre, mid and post) from same patients. Although, it is a longitudinal dataset I am starting with doing two group comparison ( pre vs mid, mid vs post). in normal case I could have done something like a paired t test, but my response variable ( RNA editing level - a measure to quantify RNA editing) is between 0 and 1(ie not normal distribution) and I want to take into account the read coverage at each site (a covariate).

Therefore, I though of using logistic regression analysis (I am skeptical here since my dependent variable is not binary but continuous numbers ), but later I realized that one of the assumptions of logistic regression is that the data should not be paired.

Now, I have a very basic understanding of statistics , therefore if someone can comment if my though process is in fact correct, if not, suggest a more appropriate method that would be greatly appreciated.

Thanks in advance,

Answer 1

[seta]

As your response variable is between 0-1, it is not binary, so you cannot use logistic regression. instead, you can use linear regression, and specially mixed linear models to account for longitudinal data. Also, you can use non-parametric tests like Wilcoxon signed rank test to compare two paired groups.

[Aiswarya]

Thank you for your response!

If I understand correctly Wilcoxon signed rank test compare means and cannot account for any covariates (in my case read depth of my samples and sex of the patients)

I came across GLMM, that I believe is suited for my purpose here.

But I want to understand the nuances of the model before I can use it, If you can point me to some resources for the same, that would be awesome.

Thanks again!

[seta]

Wilcoxon signed-rank test compares the distribution of your variable rather than the mean between two groups. it cannot account for the covariate as you mentioned. Just google GLMM to find too many resources. In GLMM, the key point is that you have fixed-and random-effect variables. Fixed-effect variables can be your main variable (I guess viral infection, here) that you would like to assess its effect on the response variable and also covariates. Random effect is related to variability in your data that is specific to your individuals and account for repeated measurements within the same individuals.

Hope it would be helpful.