Var.est error with gamma manyglm

[HaydenSchilling]

Hi,

I am trying to run a manyglm with a gamma distribution but am getting a strange error which I don't understand.

My error can be replicated with (shortened version):

mydata <- read.csv("https://raw.githubusercontent.com/HaydenSchilling/MGLMs-Otoliths/master/Data/Otolith_data_mmol_mol_Ca.csv")
E_data <- mydata[,c(2:8,10:14)]
elements <- mvabund(E_data) # select only element data
gamma_glm <- manyglm(elements ~ 1, family="gamma")

The error says:

Error in if (any(z$var.est == 0)) { : 
  missing value where TRUE/FALSE needed

It is possible to run a univariate glm with gamma distribution on all of variables I'm trying to use in the manyglm

glm1 <- glm(elements[,1] ~ 1, family =Gamma(link="log"))

Any ideas would be very much appreciated,
Thanks!

[harshadkarandikar]

I'm facing the same issue. Could you figure out a workaround? Thanks!

[HaydenSchilling]

I ended up using the manyany function and passing the gamma distribution as a special form of the Tweedie distribution.
A Tweedie distribution with variance power 2 which is equivalent to a gamma distribution.

An example of my null model was then:

fitN <- manyany("glm", elements, data = mydata, elements ~ 1, family = tweedie(var.power = 2, link.power = 0), var.power = 2)

[harshadkarandikar]

Thanks a ton! I tried that (and the Tweedie distribution might actually be more appropriate for my data), but I am now struggling to get the multivariate and univariate results through the anova function, which insists on a second manyany object for comparison (unlike for manyglm, where the anova call outputs all that). How did you get that?

[HaydenSchilling]

you can use the anova.manyany function to test between two models, one of which is nested in the other. The p-values given will then be for the factor which varies between the models. My example was as follows to test the term pop.

# Gamma Function
fit3 <- manyany("glm", elements, data = mydata, elements ~ pop, 
                family = tweedie(var.power = 2, link.power = 0), var.power = 2)
# # Null model for Gamma
fitN <- manyany("glm", elements, data = mydata, elements ~ 1, 
                 family = tweedie(var.power = 2, link.power = 0), var.power = 2)

anova_results <- anova(fitN, fit3, p.uni = "unadjusted", nBoot = 999)

Note it can be particularly slow.

[NJOram]

Hello! I am having a similar issue now, and wonder if there has been any update since 2021.
My multivariate dataset is example_data1, and I am trying to explain these with three factors in example_data2
example_data1.csv
example_data2.csv

This is my code:
remotes::install_github("aliceyiwang/mvabund")
library(tidyverse)
library(mvabund)

########## Example

load matrix data and explanatory variables

data1<-read.csv("example_data1.csv")%>%
column_to_rownames(var = "obs")%>%
as.matrix()
data2<-read.csv("example_data2.csv")

manyglm

set.seed(210)
mod1 <- manyglm(data1 ~ factor1 x factor2 x factor3,
family=Gamma(link = "log"),
data=data2
)

with family = Gamma, I get the following error:

Error in if (any(z$var.est == 0)) { :
missing value where TRUE/FALSE needed

---

Any idea why this is happening or how to fix it? all univariate GLMs run fine.

Many thanks!