MaAsLin3 analysis for longitudinal clinical analysis

Downstream analysis and statistics MaAsLin
1

I was wondering if I could get some clarification for how to properly analyze longitudinal clinical data in MaAsLin3. Essentially, I have clinical data that is comparing the change in microbial taxon relative abundance over time between a baseline and treatment group. Saliva was sampled for sequencing at week 0 before treatment began (baseline), then another 4 times during the treatment period at weeks 1-4, totaling 5 sampling points per subject.

I am mainly interested in the difference in how microbial taxon relative abundances change over time between the control and treatment groups. I figured a valid model would be structured like:

~ Week*Treatment + [other covariates] + (1|Subject)

However, there are a few details I am unsure about. First, I have noticed that in other statistical approaches that are popular for longitudinal data such as mixed models for repeated measures (mmrm), baseline is often treated as a separate covariate rather than being included in the response, implying a model like:

~ Week(1-4)*Treatment + Baseline(Week0) + [other covariates] + (1|Subject)

From what I have read, MaAsLin3 typically just treats baseline values as part of the overall Week term as the baseline level. For analysis of clinical data such as mine in MaAsLin3, would the first approach still be more appropriate?

My other question involves the use of Week as a single continuous effect vs discrete factors. To my understanding, treating Week as a single continuous effect would still work, but it would assume a linear effect between weeks, as opposed to treating Week as 5 separate levels. I am worried I may not have enough subjects to account for 1 Treatment parameter, 4 Week parameters, and 4 Week*Treatment parameters. I have seen time treated discretely in most examples for MaAsLin3, so would treating it as continuous still be viable given the assumption of linear change over the whole period?

2

Hi,

First, whether to use Week0 interacted with treatment or not depends on whether you expect the treatment to have an effect in week 0. If week 0 is before treatment begins and you have randomized treatments, knowing the treatment status in week 0 should be irrelevant, so week 0 should be included on its own without the interaction term.

For continuous vs. discrete, which models you can actually use depends a lot on how many data points you have. With enough data points relative to treatment groups (i.e. >20 subjects, 1 treatment vs. control), the categorical weeks is a strictly more general model and therefore is more likely to properly describe the data. However, if you have fewer than 10 subjects or more than 1 treatment group, using the fully interacted model might leave you with too few data points per (week, treatment) pair, and you might need to use the continuous effect. If you think the effect of the treatment consistently increases or consistently decreases over time, this is probably fine. If your number of subjects or treatments is somewhere in the middle, it’s probably worth fitting both models and seeing if the conclusions you care about are consistent. If so, you can probably report either, but if not, it’s worth figuring out why.

Will

3

Ok, I believe that makes sense. To make sure I understand, if the non-interacting Week0 is input as a separate covariate, and Week is input as a factor, then Week1 would become the dummy encoded week that weeks 2-4 are compared to, or if Week is continuous, it would model the linear rate of change just from weeks 1-4, correct?

Thank you for the clarification.

4

Yes - what you said is correct for both cases.

Will