Does the order of variables matter in ANCOM-BC

Emily_Yu · April 7, 2024, 5:07pm

Hi, when I am using "--p-formula" in ANCOM-BC, I was wondering if the order of the variables matter? For example, if my variable of interest is sex and I want to add body-site and treatment as covariates, is there any difference between:

--p-formula "Sex+Body-site+Treatment"
--p-formula "Sex+Treatment+Body-site"
--p-formula "Treatment+Body-site+Sex"
--p-formula "Body-site+Treatment+Sex"

colinbrislawn · April 11, 2024, 10:01pm

It's such a good question!

This page discibes the formula inputs as 'fixed-effects' which implies independant, but that's ANCOM-BC2
ANCOM-BC2 Tutorial

The page for ANCOM-BC does not mention fixed effects at all... ANCOM-BC Tutorial

github.com

FrederickHuangLin/ANCOMBC/blob/bugfix/R/ancombc.R#L291


      
          vcov_hat = para$vcov_hat
          var_hat = para$var_hat
          
          # 4. Estimation of the sample-specific bias
          if (verbose) {
              message("Estimating sample-specific biases ...")
          }
          fun_list = list(.bias_em)
          bias = foreach(i = seq_len(ncol(beta)), .combine = rbind) %dorng% {
              output = fun_list[[1]](beta = beta[, i],
                                     var_hat = var_hat[, i],
                                     tol = tol,
                                     max_iter = max_iter)
          }
          bias = data.frame(bias, row.names = covariates, check.names = FALSE)
          delta_em = bias$delta_em
          delta_wls = bias$delta_wls
          var_delta = bias$var_delta
          
          # 5. Obtain coefficients, standard errors, and sampling fractions
          beta_hat = beta

I think the function stats::model.matrix() ignores order. Can't find stuff on that either!

colinbrislawn · April 12, 2024, 5:22am

I tried running this with data from PD-mice

In this example, donor and genotype are fully blocked and perfectly balanced:

$ cut -f 4,6 metadata.tsv | sort | uniq -c
      1 categorical     categorical
      1 genotype        donor
     12 susceptible     hc_1
     12 susceptible     pd_1
     12 wild type       hc_1
     12 wild type       pd_1

qiime composition ancombc \
  --i-table table_2k_abund.qza \
  --m-metadata-file metadata.tsv \
  --p-formula 'donor + genotype' \
  --o-differentials ancombc_donor_first.qza

qiime composition ancombc \
  --i-table table_2k_abund.qza \
  --m-metadata-file metadata.tsv \
  --p-formula 'genotype + donor' \
  --o-differentials ancombc_genotype_first.qza

# Then qiime composition da-barplot to make these:

ancombc_genotype_first.qzv (222.0 KB)
ancombc_donor_first.qzv (222.0 KB)

On first inspection, these look the same...

Remember that

this study is fully blocked (no confounding factors) and
all cohorts are balanced (n = 12 for subgroups)
which is cleaner than most real studies!

What I don't have is a citation that says formula order doesn't matter.

Remember that different packages run different tests, so formula order can absolutely matter!

system · May 13, 2024, 11:22am

This topic was automatically closed 31 days after the last reply. New replies are no longer allowed.