Compare a numeric variable between groups at each level of a column

Runs the group test of Analyze_GroupDifference() within each level of one more column, a visit say, in one call, and returns one row per level: the test's p-value, that p-value adjusted across the levels, the counts, and for two groups the difference in means with its interval.

Usage

Analyze_GroupDifferenceBy(
  dfData,
  strValueCol,
  strGroupCol,
  strByCol,
  strMethod = "t",
  chrGroups = NULL,
  chrBy = NULL,
  strPAdjust = "none",
  nConfLevel = 0.95,
  nMinGroup = nMinGroupDefault
)

Arguments

dfData: data.frame One row per participant and level.
strValueCol: character Name of the numeric column to compare.
strGroupCol: character Name of the column holding each participant's group.
strByCol: character Name of the column holding the level of each row, such as the visit. The test is run once within each level.
strMethod: character The test: "t", "wilcoxon", "anova" or "kruskal". Default: "t".
chrGroups: character The groups to compare at every level, in order; the difference in means is the first minus the second. Rows in any other group are dropped and counted. Default: NULL, every group present among the levels answered, in sorted order.
chrBy: character The levels to answer, in order. Rows at any other level are dropped and counted. A level named here that the data do not hold is a row with nobody in any group. Default: NULL, every level present, in sorted order.
strPAdjust: character The adjustment across the levels, one of stats::p.adjust.methods. Default: "none".
nConfLevel: numeric Confidence level of the intervals. Default: 0.95.
nMinGroup: numeric The smallest group the test is computed for. A level where any group has fewer participants with a value has status "too_small" in its row and no numbers. Default: nMinGroupDefault, which is 5. See StatisticsResult.

Value

The fixed result described in StatisticsResult. Here p_value is NA and estimates and statistic have no rows; method is the first computed level's; counts is a named list of level to participants used; dropped counts the rows with no level or a level not asked for, and then, within the levels answered, the rows the test left out; and rows has one row per level, with the columns:

by, the level.
group_1, group_2 and so on, one per group, the groups compared, the same on every row; then n_1, n_2 and so on, the participants used in each at that level. With two groups these are the columns a pairwise comparison of Analyze_GroupDifference() has.
counts (the groups together) and dropped (the rows of that level the test left out).
estimate, lower, upper and level: with two groups, the difference in means, group_1 minus group_2, and its interval, from t.test() (Welch) whatever the test; NA with more than two.
method and statistic, as R reports them for that level.
p_unadjusted, p_value (adjusted across the levels), adjustment and adjusted_over.
status, reason and warning for that level.

The result's own status is "ok" when any level is, "too_small" when every level is too small, and "error" when no level could be computed for any other reason. Each row gives its own reason.

Details

The data are long: one row per participant and level, as a results table holds one biomarker across its visits. Each row of the answer is Analyze_GroupDifference()'s own answer on the rows of that level, with bPairwise off and the same groups named, so its unadjusted p-value is the one printed when the level is looked at alone with those groups.

The groups are the same at every level: the ones named in chrGroups, or every group present anywhere among the levels answered. A group with nobody at a level is too small there; the level is not quietly compared without it. So with three or more groups, a level where one of them has nobody has no p-value here, while Analyze_GroupDifference() on that level's rows alone, with chrGroups left out, compares the groups that are there.

The adjustment is stats::p.adjust() across the levels that have a p-value. A level that could not be computed, because a group is below the minimum size or its values do not vary, has its own status and reason, has no p-value, and is left out of the adjustment; notes says how many levels the adjustment covered, and each adjusted row carries that number as adjusted_over. The default is no adjustment: p_value is then each level's own p-value. Holm guards against any false lead among the levels; Benjamini-Hochberg controls the share of false leads among the levels picked out.

The rows come back in the order of chrBy. R does not know the order of visits, so name them in chrBy to have them in theirs.

Examples

# IL-6 by arm at every visit, with the p-values adjusted across the visits
dfIL6 <- merge(
  Synthetic_Results[Synthetic_Results$TEST == "IL-6", ],
  Synthetic_Participants[c("USUBJID", "ARM")]
)

lResult <- Analyze_GroupDifferenceBy(
  dfIL6, "STRESN", "ARM", "VISIT",
  chrGroups = c("Treatment", "Placebo"),
  chrBy = c("Baseline", "Week 2", "Week 4", "Week 8", "Week 12"),
  strPAdjust = "holm"
)
lResult$rows[c("by", "n_1", "n_2", "estimate", "p_unadjusted", "p_value", "adjustment")]
#>         by n_1 n_2  estimate p_unadjusted      p_value adjustment
#> 1 Baseline 100 100 -0.252230 2.214152e-01 2.214152e-01       holm
#> 2   Week 2  93  92 -1.526236 4.840136e-10 9.680271e-10       holm
#> 3   Week 4  91  95 -1.544451 2.215792e-10 6.647376e-10       holm
#> 4   Week 8  95  93 -1.577415 5.542038e-12 2.216815e-11       holm
#> 5  Week 12  92  92 -1.596728 3.459919e-12 1.729960e-11       holm

# The same p-value as the single function on one visit's rows
Analyze_GroupDifference(
  dfIL6[dfIL6$VISIT == "Week 4", ], "STRESN", "ARM",
  chrGroups = c("Treatment", "Placebo")
)$p_value
#> [1] 2.215792e-10