Allow method = "lmm" in mia::getCrossAssociation with parameter for random effects

[xec-cm]

Description:
Currently, mia::getCrossAssociation does not support using linear mixed models (method = "lmm"). It would be very useful to extend the function so that users can specify method = "lmm" and provide an additional parameter to define the variable(s) used for random effects.

Use case:
In many microbiome studies, repeated measures or clustered data (e.g., multiple samples from the same subject) are common. Being able to include random effects would allow a more appropriate modeling of within-subject correlation, avoiding inflated type I error rates and improving inference.

Proposed solution:

• Add support for method = "lmm" in mia::getCrossAssociation.
• Introduce a new argument (e.g., random_effects or random) where the user can specify the variable(s) to be used as random effects.

Example (hypothetical):

res <- mia::getCrossAssociation(
  x = data,
  y = outcome,
  method = "lmm",
  random_effects = "subject_id"
)

Benefits:

• Enables analysis of longitudinal or clustered microbiome data.
• Increases flexibility and applicability of mia::getCrossAssociation to real-world study designs.

[TuomasBorman]

Thanks @xec-cm for the suggestion! You are absolute right that it would make the analysis more robust.

The initial design of getCrossAssociation() was to enable simple correlation analysis. Because of the design of it, this might not be too straight-forward. The function has association.fun argument, but it is not suitable in this case as it does not allow passing covariates or random effects.

That said, you highlight clear limitation and area to improve. I would approach this with different view. We are currently planning to add better support for linear modeling (for differential abundance analysis) to the ecosystem. It would probably be easier to support this kind of correlation analysis with random effects in that framework.

Any thoughts @antagomir ?

[antagomir]

Yes I assume so, the new developments could support screening across multiple features, effectively yielding the same.