Problem with pglmm Bayes=TRUE model syntax

[Flaiba]

hello phyr team.
I am interested in replicating the analysis carried out by Vincze et al. 2022 (https://doi.org/10.1038/s41586-021-04224-5), employing other explicatory variables using the library phyr since this package might allow better analysis. For my question, I used the data frame that these authors make available (https://github.com/OrsolyaVincze/VinczeEtal2021Nature/blob/main/SupplementaryData.xls).
The work carried out by Vincze et al. (2022) studied a simple measure of cancer mortality risk (CMR) in mammals in the relationship with variables like body size or lifespan. Addionatilly, it was necessary to incorporate phylogenetic information due to the lack of independence between the species analyzed.
The data frame contains:
Variable Response (CMR) = the ratio between the number of cancer-related deaths (Neoplasia) and the total number of individuals whose postmortem pathological records were entered in the database (knownDeaths).

Like a first approach, we think of the following model:
(cbind(Neoplasia, knownDeaths-Neoplasia) ~ log(BodyMass)+log(lifeexp), data = data, family = "binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE)

However, when trying to make this model, we find some inconveniences.
First of all, I would like to ask to confirm my suspicions: the phyr library does not allow models to be made without any random variable,? Since I did not find any library that allows me to run a binomial model for proportions considering the phylogenetic information and not including any random variable. I ask this because it is possible that a binomial model (for proportions) does not have overdispersion, so in this case for these data I would not need to use mixed models.
However, it is known that binomial distribution models tend to suffer from overdispersion, so we decided to incorporate a random effects variable (OLRE) to take into account overdispersion. In this way, we can use the library without any inconvenience, using the following model:

M1=pglmm(cbind(Neoplasia, knownDeaths-Neoplasia) ~ log(BodyMass)+log(lifeexp)+ (1|OLRE), data = data, family = "binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE)

However, this model shows a bad fit and overdispersion caused by an excess of 0.

So, as a next step, we decided to implement zero-inflated models that this library would allow us to incorporate, using the following model:
M2=pglmm(cbind(Neoplasia, knownDeaths-Neoplasia) ~ log(BodyMass)+log(lifeexp)+ (1|OLRE), data = data, family = "zeroinflated.binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE,bayes=TRUE,verbose=TRUE), but.....

Consider entering the response variable as the proportion (CMR), but according to inla I have to enter it as an interger, but my problem is that I have no way to define the weights or the Ntrial for that proportion.

I admit that I have problems with the syntax of my model in relation to the response variable.
Could someone help me with the syntax to be able to run the model I need?

Thank you very much for your time and help.
Nicolas, thanks, thanks, thankssssssss........
Attached the database and the script used

`library(ggplot2)
library(ape)
library(car)
library(phytools)
library(phylolm) # phyloglm
library(phyr)
library(DHARMa)
library(dplyr)

str(data)
data$OLRE=as.factor(data$OLRE)
data$Species=as.factor(data$Species)
data$order=as.factor(data$order)

Species-specific body mass

data$BodyMass <- (data$MaleMeanMass + data$FemaleMeanMass)/2

phylogeny

phy <- read.nexus("consensus_phylogeny.tre") # consensus vertlife tree
phy <- bind.tip(phy, "Cervus_canadensis", where = which(phy$tip.label=="Cervus_elaphus"),
edge.length=0.5, position = 0.5)
phy <- bind.tip(phy, "Gazella_marica", where = which(phy$tip.label=="Gazella_subgutturosa"),
edge.length=0.5, position = 0.5)

#firts model

M0=pglmm(cbind(Neoplasia, Nodeadbycancer) ~ log(BodyMass), data = data, family = 'binomial',cov_ranef = list(sp = phy))

No random terms specified, use lm or glm instead, There is a possibility that running a model without random effects?

M1=pglmm(cbind(Neoplasia, knownDeaths-Neoplasia) ~ log(BodyMass)+log(lifeexp)+ (1|OLRE), data = data, family = "binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE)
summary(M1)

simulationOut<- simulateResiduals(fittedModel = M1, n = 250,refit = FALSE)
plot(simulationOut) # bad fit
testZeroInflation(simulationOut) # overdispersion by zeros
testOverdispersion(simulationOut) # overdispersion by not zeros is good

M2=pglmm(cbind(Neoplasia, knownDeaths-Neoplasia) ~ log(BodyMass)+log(lifeexp)+ (1|OLRE), data = data, family = " zeroinflated.binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE,bayes=TRUE)
`
dataIZ.txt

[daijiang]

It seems that @rdinnager is working on this. @Flaiba if you don't need random terms, I think the phylolm package should work.

[rdinnager]

Hi @Flaiba ,

Thanks for the detailed bug report. I believe the problem here is that your formula is not getting translated into an INLA formula properly, and it's because currently doing an operation inside cbind in a formula is not supported. I will work on fixing the bug (and also providing a way to pass Ntrials argument through to INLA directly). As a workaround you should be able to create a new column in your data.frame for the failed trials and then use that in the cbind, and it should work. So something like this:

data$fails <- data$knownDeaths - data$Neoplasia
M2 <- pglmm(cbind(Neoplasia, fails) ~ log(BodyMass) + log(lifeexp) + (1|OLRE), data = data, family = "zeroinflated.binomial", cov_ranef = list(sp = phy), add.obs.re = FALSE, bayes = TRUE)

Let me know if the above does not work.

[Flaiba]

Hi daijiang and rdinnager
Thanks for the reply.
As for the model with random effects, I may be wrong, but phylolm is not just for logistic regression and Poisson distribution. In other words, this package only can be used for the binomial distribution with a dichotomic response (success/fail), not for proportions like as (cbind(success, fails). Am I correct?

On the other hand, I think the problem is as you said rdinnager. Unfortunately, I used your code, but I still have the same problem.
When moving to INLA, there is a problem with the code concerning how to codify the variable response. Again appear the Error in the formula:

"Error in formula.character(object, env = baseenv()) :
invalid formula "Neoplasia": not a call "

The response variable is not detected. :(

Thank you both very much. Packages like this are essential to making better models and the possibility of asking them for help is an unbeatable chance to continue learning. Thanks.

[rdinnager]

Hi @Flaiba , Thanks for giving that a try. After the error occurs, could you run traceback() and paste the results here? It would also be helpful if you could make a minimal reproducible example. I can't run the code you pasted here because I don't have the tree file you used ("consensus_phylogeny.tre").

[Flaiba]

Hi! first of all, I apologize for not attaching the phylogenetic tree.
I copy the link to the phylogenetic tree.
https://github.com/OrsolyaVincze/VinczeEtal2021Nature/blob/d2f8e2dff8e64393ce3512f8b8490507fa9ae803/consensus_phylogeny.tre
These are the syntaxes I tried with the subsequent error message:

M2=pglmm(cbind(Neoplasia, knownDeaths-Neoplasia) ~ log(BodyMass)+log(lifeexp)+ (1|OLRE), data = data, family = "zeroinflated.binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE,bayes=TRUE,verbose=TRUE)
Error in formula.character(object, env = baseenv()) :
invalid formula "Neoplasia": not a call

data$fails <- data$knownDeaths - data$Neoplasia
M3 <- pglmm(cbind(Neoplasia, fails) ~ log(BodyMass) + log(lifeexp) + (1|OLRE), data = data, family = "zeroinflated.binomial", cov_ranef = list(sp = phy), add.obs.re = FALSE, bayes = TRUE)

Error in formula.character(object, env = baseenv()) :
invalid formula "Neoplasia": not a call

and Finally, I tried used the ratio of Neoplasia/KnownDeadths (CMR), but the problem is with the INLA

Error in inla.inlaprogram.has.crashed() :
The inla-program exited with an error. Unless you interupted it yourself, please rerun with verbose=TRUE and check the output carefully.
If this does not help, please contact the developers at help@r-inla.org.
Error in inla.core.safe(formula = formula, family = family, contrasts = contrasts, :
*** Fail to get good enough initial values. Maybe it is due to something else.

Thanks again, I hope that with the information from the phylogenetic tree you can replicate this problem, and finally get a solution.

[rdinnager]

Hi @Flaiba. I've now made a patch which I think should fix your issue. To try it just install the development version of phyr from github using remotes::install_github("daijiang/phyr"). Please let me know if you if this resolves your problem.

[Flaiba]

I am very grateful for your help. Both are very helpful. I can run this model perfectly.

M1=pglmm(cbind(Neoplasia, Nodeadbycancer) ~ LitterSize +log(BodyMass)+log(lifeexp)+(1|OLRE), data = datapaper, family = "zeroinflated.binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE,bayes = TRUE)

However, Can I use this space to ask two things related to this model? Firstly, I saw in this blog/help (https://daijiang.github.io/phyr/articles/phyr_example_empirical.html) for the package that the models can be checked using the Dharma package. In the case of this model, perhaps after this actualization, I can not check the model by Dharma.

simulationOut<- simulateResiduals(fittedModel = M1, plot = FALSE)
plot(simulationOut)

Maybe the problem is why the distribution of the model is zero-inflated? In this case, I probed with a binomial distribution "common" and finally, I saw this problem:

M2=pglmm(cbind(Neoplasia, Nodeadbycancer) ~ LitterSize +log(BodyMass)+log(lifeexp)+(1|OLRE), data = datapaper, family = "binomial",cov_ranef = list(sp = phy),add.obs.re = FALSE,bayes = TRUE)
summary(M17)
simulationOut<- simulateResiduals(fittedModel = M2, plot = FALSE)
plot(simulationOut)

Is it possible using Dharma to check these models?

Additionally, and I promise this is the last asked.
In my analysis, I have a variable that has 3 levels. Since this variable is significant, I need to do contrasts, for example, Tukey. So, what package can I use to do the contrasts that support objects of phyr package? Because the emmeans does not support this package

Thanks, thanks a lot for your time

[rdinnager]

Hmm.. I'm not surprised the zero inflated model didn't work with DHARMa, but the binomial one definitely should work. I'll look into that and get back to you. I will see if it is possible to get the zero inflated one working too. As for your final question, there are several possibilities. phyr doesn't support emmeans currently. For the Bayesian model, what I would do is extract the posterior distribution of the relevant coefficients to reconstruct the categorical means and plot them. It should be obvious which categories differ. If you have an a-priori hypothesis for which levels should be compared, there are two options:

1. Reparameterise your model to make the comparison (that is, change how dummy variables are generated to make the right contrast(s)). This isn't always possible.
2. Calculate the required contrasts from a sample from the posterior distribution, and then summarise across many posterior samples to get a distribution of the contrast values.

We intent to add more user friendly options for doing this sort of thing in the future, but at the moment it might require manually extracting useful quantities from the model object. Supporting emmeans would probably be a good way forward for supporting these use cases.

How did you determine that the categorical variable was 'significant'? Were you using the maximum likelihood or Bayesian methods for this?

[rdinnager]

In the mean-time, you may want to look into the emmeans::qdrg() function (https://rdrr.io/cran/emmeans/man/qdrg.html), and see if you can manually extract the useful quantities from the communityPGLMM object.