Brms with uncertain phylogenetic signal

General
,
1

Hi,

I am trying to use brms to run models that include a phylogenetic signal. However, I am unsure if my response variable actually contains a phylogenetic signal. The tests for phylogenetic signal in the response variable, Blomberg et al.'s (2003) K and Pagel’s \lambda, give different results. I recall reading that we should include the phylogenetic signal when working with species data regardless. The reasoning is that if a phylogenetic signal exists, the model will account for it; if not, it won’t cause any harm. Is this true for phylogenetic models in brms?

One more question: one of my predictors definitely contains a phylogenetic signal (in the formula below, the variable ‘logGS’). Could this cause any issues?

logCover ~ logGS + Fertilization + logGS * Fertilization + 
        (1|gr(phylo, cov = A)) + (1|PlotID) + (1|species))

Thanks in advance,
Emma

2

Hi Emma,

In principle, yes, just like any random effect that captures residual clustered variation. So I would be inclined to estimate it, even if the phylogenetic signal is low.

I wouldn’t say it “causes issues”, but you’re right to think about this because if logGS has a strong phylogenetic signal and logGS → logCover, then adjusting for logGS in your model should result in a weaker phylogenetic signal because it “mediates” the variance captured by the phylogenetic random effects. So if you want a more traditional estimate of phylogenetic signal, you might first fit a model without the fixed effects (or at least without any fixed effects where the predictors are phylogenetically correlated).

2 Likes
3

Hi Erik,

Thank you very much for the answer! But I am not sure if I get your idea. Please spare me if I am asking some stupid questions.

How exactly does including the phylogenetic signal adjust for logGS? My understanding was that phylogenetic adjustment affects the response variable rather than the predictor?

Unfortunately, my real target is the effect of logGS. I have tried two different models. In model 1 below (without phylogenetic signal), the logGS has a significant effect, while in model 2 (with phylogenetic signal), it is no longer significant. Does this imply that the effect of logGS observed in Model 1 might actually be due to phylogenetic relationship?

# Model 1:
model_withoutPhylo <- brm(bf(
    logCover ~ logGS + Fertilization + logGS:Fertilization + 
        (1|PlotID) + (1|species)),
    data = AEG_GS_fertilization,
    warmup= warmup,
    family = gaussian(),
    iter=iter,
    thin=thin,
    init="random",
    chains=chains,
    cores= parallel::detectCores(),
    sample_prior = TRUE,
    control=list(adapt_delta = 0.9))

# Model 2:
model_withPhylo <- brm(bf(
    logCover ~ logGS + Fertilization + logGS:Fertilization + 
        (1|gr(phylo, cov = A)) + 
        (1|PlotID) + (1|species)),
    data = AEG_GS_fertilization,
    data2 = list(A = A),                   
    warmup= warmup,
    family = gaussian(),
    iter=iter,
    thin=thin,
    init="random",
    chains=chains,
    cores= parallel::detectCores(),
    sample_prior = TRUE,
    control=list(adapt_delta = 0.9))

Thank you again for your insight!

Emma

4

The idea is that there is a “total effect” of phylogeny, which is all the variation that can be predicted using the expected phylogenetic covariance matrix (A). If more closely related species are more similar on the response variable (logCover) in part because they have similar values of logGS, then logGS partially mediates the influence of phylogeny on logCover. Or maybe think of it this way: the (unadjusted) phylogenetic signal tells us what percentage of the variance can be accounted for by phylogeny. Some percentage of that percentage is accounted for by logGS, which presumably also has a phylogenetic history. So we would expected the phylogenetic signal to go down once adjusting for logGS.

It is consistent with that interpretation, yes.

1 Like
5

Thank you very much for the clarification!

1 Like