Hi -
I am running an analysis on a complete randomized block design. We have 3 treatment groups (between subjects; Treatment A, Treatment B, and Treatment C) and 5 time points per animal (repeated measure). Because our design is not perfectly balanced, I can’t use the ANOVA option and I have to use the Linear Model option. My primary metadata is treatment. I am modeling time as a continuous covariate, subject as the blocking factor to account for repeated measures, and then I have block and 2 other covariates to account for variation as covariates. Please let me know if you have any thoughts on how I set this up.
How do I conduct pairwise comparisons between my 3 treatment groups using a linear model? OR do I need to run the analysis 3 separate times, one time for each comparison (Treatment A v. B, A v. C, and B v. C)? I see a P-value and an adjusted P-value (presumably for the main effect of treatment), but I don’t see pairwise comparisons.
Currently the linear model tool is not set-up to return p-values from individual contrasts when the categorical primary metadata has more than two groups. For a categorical primary metadata, you must set a reference treatment (ie. Treatment A). Then, if there are more than two groups, MetaboAnalyst will return the fold-change for each other treatment vs. reference (ie. Treatment B/Treatment A; Treatment C/Treatment A) - see the first columns in the results table. For categorical variables with two groups, the results table p-value is for the one contrast and derived from an t-stat. For those with more than two, it’s an ANOVA-style p-value for the set of contrasts, derived from an F-stat.
I am currently implementing an interface that will allow you to get statistics for a specific contrast when the categorical primary metadata has more than two groups. It should be done in a few days, and I’ll update this with how to use it to address your situation.
In this example, I am looking for metabolites that are significantly different between subjects exposed to “High” and “Low” concentrations of TCE (the primary metadata variable), while accounting for age and sex as covariates (fixed effects) and batch as a blocking factor (random effect).
To get pairwise results, I would save this, then perform two more analyses as you outlined, in my case: “Medium” vs. “Low” and “High” vs. “Medium”.
