Phyloseq CCA: 'x' and 'w' must have the same length

[pauGuas]

Hi, I am trying to run a CCA on the numerical environmental factors that influence microbiome composition (phyloseq object). The environmental factors are summarized in "var". When I run the code below, I get "Error in weighted.mean.default(newX[, i], ...) : 'x' and 'w' must have the same length". I don't know what 'w' is in this. Does anyone know what to do with this? Thank you!

ps <- phyloseq_object
prop= transform_sample_counts(ps, function(OTU) OTU/sum(OTU))
prop_df= as.data.frame(otu_table(prop))
prop_dft= as.data.frame(t(prop_df))

prop_dft <- ps_prop_dft[rowSums(prop_dft[, -1])>0, ]

var <- c("Chloride", "Inorganic.N", "Kjeldahl.N",  "Dissolved.O", "Temp",  "TSS", "Velocity.of.Discharge", "pH")

ccamodel= cca(prop_dft ~ var)

[MarkReader]

The "w" field will be, I guess, the Population Weights. One value (weight) for each n (in sample), hence w has to be same length as your obs. They will be different if they are population weights, such as (often) much higher weights for known demographics that are undersampled (higher n/r rates, etc). If you just want the arithmetic (sample, not population) mean, just set them all to 1.000. :)

[gavinsimpson]

@MarkReader CCA involves abundance-weighting the data so the error could just as easily be being raised from functions internal to cca.

@pauGuas Solve the obvious problem first; that's not how you fit a CCA model. The RHS of the formula should be a standard R formula: ~ x1 + x2 + x3 etc, not a character vector.

If you want to proceed using the defined var, use:

ord <- cca(prop_dft ~ ., data = ENV[, var])

replacing ENV with your data frame containing the environmental variables that you want to use in the model.

[MarkReader]

Very likely so. (But, out of curiosity @gavinsimpson what are the "w"? Will be really grateful if you can DM me, if you like.).

[jarioksa]

Yes, the error is triggered in an internal function: user should not and cannot supply weights to cca. If you are interested in details, look at the function vegan:::ordConstrain which has the following code block that generated the error:

    if (!is.null(RW)) {
        envcentre <- apply(X, 2, weighted.mean, w = RW)
        X <- scale(X, center = envcentre, scale = FALSE)
        X <- sweep(X, 1, sqrt(RW), "*")
    }

The weights RW are the row weights of the dependent data used as internal weights in Correspondence Analysis (the method, irrespective of its software implementation). These weights are calculated in vegan:::initCA function.

However, I cannot understand how you reach this error. Like @gavinsimpson wrote, the model definition is wrong. Moreover, we do check the dimensions before going to vegan:::ordConstrained, and the commands in the original message should give this error message:

Error in ordConstrained(X, Y, Z, method = "cca") : 
  dependent data and constraints must have the same number of rows

(and I got this when I tried to mimick the original commands – I could not reproduce them as there was nothing reproducible.) The command should only work if length(var) == NROW(prop_dft), and then you would get a useless result identical to unconstrained CA (as the number of levels of var factor equals the number of observations).

The way to proceed is to follow Gav's instruction, and if it still fails, provide a reproducible example (where "reproducible" means something we can run and get the same error).