topic Re: missing output in mixed effects models for some variables only in Discussions

missing output in mixed effects models for some variables only

Jake_b — Fri, 09 Jun 2023 00:42:43 GMT

hello,

I am trying to model the effects of metal ions on different metabolites in earthworms - i.e. it is an environmental pollution question. I have both concentration values for a number of different metal ions, and for a number of different metabolites, both on a per-worm (individual) basis.

The analysis is complicated by the fact that the worms are collected from a number of different sites, and there is clear clustering within sites - both in terms of the metabolite similarity, and also the metal ion concentrations (this second not surprising, as some of the sites were sampled specifically because they were polluted).

I am trying to model the effect on metabolites as a mixed effects model, using the Fit Model menu, with "metal" as a fixed effect, and "site" and "site x metal" as random effects ("metal" and "metabolite" are continuous variables, "site" is categorical).

Now, here is my question. For some metabolites, the JMP output gives me what I am looking for, a significance test ("fixed effect test") for the effect of metal ion on metabolite concentration, BUT for some other metabolites, it does not calculate a value, and just gives "." Any suggestions as to what is going on? There are no missing values, so the number of samples and variables is identical - it is just that it gives me an answer for some dependent variables and not others.

thank you...

Re: missing output in mixed effects models for some variables only

Mark_Bailey — Fri, 03 Dec 2021 16:59:42 GMT

Can you share the data table, either as an attached file, or as a screen capture?

Can you also share the JMP platform (as a picture)? It might be that you specified the model in a way that confounds some effects.

Are the sites acting as a blocking factor, with a random effect? Is that interpretation what you mean by "clear clustering within sites?"

Re: missing output in mixed effects models for some variables only

Jake_b — Tue, 07 Dec 2021 10:31:04 GMT

Can you share the data table, either as an attached file, or as a screen capture?

I'll upload a cut-down version - it's the same as the original file, but I have just deleted almost all of the variables, just leaving enough to represent two comparisons.

(NB that two of the samples, rows 21 and 22, are excluded from the analysis.)

Can you also share the JMP platform (as a picture)? It might be that you specified the model in a way that confounds some effects.

Yes, I'll give that a go, thanks - hopefully this is what you meant, a screenshot of the relevant options chosen?

NB that for me, the two Y variables shown have different outcomes, in that JMP successfully calculates a P value for one but not the other.

Are the sites acting as a blocking factor, with a random effect? Is that interpretation what you mean by "clear clustering within sites?"

This is not a designed field trial - these are worms from indigenous populations, where we have gone and dug them up. (For fuller reference: we do classify the sites as "control" and "predicted", and I would also be keen to try and model this classification, but for the current question, I am trying to model the relationship for potential pollutants (metal ions) on a per-worm basis, rather than a per-site basis.) The sites are highly distinct from each other, whether you make the classification on the basis of metal ions, or of the worm metabolite concentrations - e.g. clearly visualised by multivariate approaches such as PCA. I definitely don't consider myself expert at all in using mixed effects models - my approach here is based on discussion with colleagues, and I would not be shocked to find that I am trying to do something that is not justified.. NB one of their comments was that assigning random effects is only really meaningful for categorical variables, but JMP does allow one to assign random effects to continuous variables: should I just assume that this is not a good idea?

Summarising, I can get to an answer that works, using PCA and PLS modelling (with cross-validation hold-out sets based on site, using a different package than JMP - maybe I should be looking at some kind of resampling approach for the univariate analyses too??), but I would really like to be able to get the answer to the univariate case as well. There are three reasons:

I would like to be able to present a univariate analysis alongside the multivariate one.
Even if I primarily present the multivariate analyses, it is really useful to be able to pull out key variables as potential biomarkers, and analyse these in more detail.
I want to be able to understand what is going on!

many thanks ...