Subject: Question about uniqueness_gtest / gtest.standard with multi-group designs (ftmsRanalysis 1.1.0)

[ycsong-pnnl]

I have recently used ftmsRanalysis (v1.1.0) for G-tests of uniqueness on FT-ICR MS data,
and I would like to understand two behaviours in gtest.standard(). I would appreciate your
view on whether they're intended, and the reasoning if so.

Workflow

  peakObj <- as.peakData(edata, fdata, emeta, ..., data_scale = "pres")
  peakObj <- group_designation(peakObj, main_effects = c("Layer", "Respiration"))
      # -> 4 groups: BTM_High, BTM_Low, TOP_High, TOP_Low
  comp <- divideByGroupComparisons(peakObj,
                                   comparisons = list(c("BTM_High", "BTM_Low")))[[1]]
  res  <- summarizeGroupComparisons(comp,
            summary_functions = "uniqueness_gtest",
            summary_function_params = list(uniqueness_gtest = list(
              pres_fn = "nsamps", pres_thresh = 2, pvalue_thresh = 0.05)))

What I observe

After divideByGroupComparisons() extracts a two-group comparison, Group is still a
factor with all four levels. In gtest.standard():

grps = as.factor(group)
num.grps = length(levels(grps)) # 4
...
df = num.grps - 1 # 3
major.group[j] = ifelse(obs.miss[2, 1] > obs.miss[2, 2],
grp.names[1], grp.names[2])

1. Degrees of freedom: a two-group comparison is tested with df = 3. If only the
   two levels being compared are kept, df = 1. The G statistic is the same either
   way; only the p-value changes.

2. Major group: major.group comes from factor levels 1 and 2 (here BTM_High and
   BTM_Low) whatever pair is compared. For a pair that doesn't include those levels
   (e.g. TOP_High vs TOP_Low), the result names a group outside the comparison, and
   uniqueness_gtest records it as NA. In our data, TOP_High vs TOP_Low gives 0
   compounds unique to either group, while 2,350 compounds have p <= 0.05.

Minimal example

One compound present in all 11 samples of group A and absent from all 10 of group B:

  library(ftmsRanalysis)
  x   <- matrix(c(rep(1, 11), rep(NA, 10)), nrow = 1)
  lv  <- c("BTM_High", "BTM_Low", "TOP_High", "TOP_Low")
  g_a <- factor(c(rep("BTM_High", 11), rep("BTM_Low", 10)), levels = lv)
  g_b <- factor(c(rep("TOP_High", 11), rep("TOP_Low", 10)), levels = lv)

  ftmsRanalysis:::gtest.standard(x, g_a)
  #   pvals = 2.17e-06, major.group = "BTM_High"
  ftmsRanalysis:::gtest.standard(x, droplevels(g_a))
  #   pvals = 7.0e-08,  major.group = "BTM_High"
  ftmsRanalysis:::gtest.standard(x, g_b)
  #   pvals = 2.17e-06, major.group = "BTM_Low"

Questions

1. Is the df meant to reflect the number of groups in the full design rather than
   in the comparison? If so, could you explain the reasoning?
2. Is major.group meant to come from factor levels 1 and 2 regardless of which pair
   is compared?
3. If the two-group behaviour is intended, is applying droplevels() to Group before
   the test the recommended approach? And has this changed in any release after
   1.1.0?

Thanks very much for your time, and for the package.

Best regards,
Young