Nested design

[aston9090]

Hello, thanks for developing the package! Unfortunately I am new to microbiome analysis and I encountered some errors using ANCOMBC2.

I have data from 12 species (each with 10 samples) that are divided into 3 groups, with 4 species in each group. I would like to test if there are differentially abundant microbiota in each group, accounting for the within-group variations (caused by animal species) at the same time. I tried the following code:

`
Sample <- paste0("Sample", 1:120)
Group <- rep(c("A", "B", "C"), each = 40)
Subgroup <- rep(rep(c("A1", "A2","A3","A4", "B1", "B2", "B3","B4", "C1", "C2","C3","C4"), each = 10))

combine the columns into a data frame

meta <- data.frame(Sample, Group, Subgroup)
rownames(meta)<-meta$Sample

Count generation

number of ASVs (na)

na <- 100
x <- rnbinom(120*na, .1, mu = 100)
x <- matrix(x, nrow = 120, ncol = na)
colnames(x) <- paste0("ASV", 1:na)
rownames(x) <- paste0("Sample", 1:120)

Final phyloseq experiment

exp <- phyloseq(
otu_table(x, taxa_are_rows = FALSE),
sample_data(meta)
)

#ANCOMBC2
trial <- ancombc2(data = exp, fix_formula = "Subgroup + Group", tax_level="Species")
`

And got an error:
Error: Estimation failed for the following covariates: GroupB, GroupC Consider removing these covariates

Any help and advice would be much appreciated. Thank you!

[FrederickHuangLin]

Hi @aston9090,

In your case, it seems that the variables Group and Subgroup are highly correlated, leading to multicollinearity issues in the linear regression model. To confirm this, you can perform the following steps:

df <- data.frame(y = log(x[, 1] + 1), meta)
fit <- lm(y ~ Subgroup + Group, data = df)
summary(fit)

If there is a high correlation between Group and Subgroup, you will notice that the coefficients for GroupB and GroupC are likely to be listed as NAs in the model summary.

It is generally recommended to avoid including highly correlated independent variables in regression models to prevent multicollinearity. Consider choosing one of the variables or exploring alternative modeling approaches if both variables are important for your analysis.

Best regards,
Huang

[sihanbu]

Hi Huang, are there recommended ways to include these correlated variables into ancombc2?