Computes the correlation of every pair of numeric columns and returns it
in long form – one row per pair, which is easier to sort, filter and join
than a matrix – with the number of complete pairs each figure rests on.
The counterpart of corrr::correlate().
Usage
cortbl(data, meth = c("spearman", "pearson"), minprs = 3L)Arguments
- data
A data frame; its numeric columns are correlated in pairs.
- meth
"spearman"(default) or"pearson". Seetopcors()for why the rank correlation is the default.- minprs
Minimum complete pairs required before a correlation is computed, a whole number of at least 2 (default 3). Below it the pair is
NArather than a number computed from almost nothing.
Value
A data frame of class c("bricklayer_cortable", "data.frame")
with one row per pair of numeric columns and the columns x and
y (character, the column names), correlation (numeric, in
\([-1, 1]\) or NA) and nprs (integer); the method is
in attr(, "method").
Details
Non-numeric columns are ignored. For each pair, rows where either value is
missing are dropped (pairwise deletion), so different pairs can rest on
different rows; nprs says how many. Spearman's rank correlation
(the default) is the Pearson correlation of the ranks, robust to outliers
and to monotone transformations; see
topcors() for why it is
the default. A pair with fewer than minprs complete rows, or
with a constant column (no variance), gets NA. The pairs are listed
in column order (utils::combn()). A progress bar is drawn on stderr
for many pairs unless options(morie.quiet = TRUE).
Errors: data not a data frame; fewer than two numeric columns;
minprs not a whole number of at least 2; an invalid
meth.
Examples
set.seed(1)
df <- data.frame(a = stats::rnorm(100), b = stats::rnorm(100))
df$c <- df$a + stats::rnorm(100, sd = 0.2)
cortbl(df)
#> ╭── Correlations (spearman) ──────────────────────────────────────────────╮
#> │ Method Spearman (rank) │
#> │ Pairs 3 │
#> │ Strongest |r| 0.9665 │
#> │ │
#> │ x y correlation nprs │
#> │ - - ----------- ---- │
#> │ a b 0.0387 100 │
#> │ a c 0.9665 100 │
#> │ b c 0.0151 100 │
#> │ │
#> │ Each row is one pair of numeric columns, over the rows where both are │
#> │ present (nprs); a correlation is blank when too few pairs remain. │
#> ╰─────────────────────────────────────────────────────────────────────────╯
# Pearson instead of Spearman
cortbl(df, meth = "pearson")
#> ╭── Correlations (pearson) ───────────────────────────────────────────────╮
#> │ Method pearson │
#> │ Pairs 3 │
#> │ Strongest |r| 0.9747 │
#> │ │
#> │ x y correlation nprs │
#> │ - - ----------- ---- │
#> │ a b -0.0010 100 │
#> │ a c 0.9747 100 │
#> │ b c -0.0120 100 │
#> │ │
#> │ Each row is one pair of numeric columns, over the rows where both are │
#> │ present (nprs); a correlation is blank when too few pairs remain. │
#> ╰─────────────────────────────────────────────────────────────────────────╯
# nprs shows how much data each figure rests on.
gappy <- df
gappy$a[1:80] <- NA
cortbl(gappy)
#> ╭── Correlations (spearman) ──────────────────────────────────────────────╮
#> │ Method Spearman (rank) │
#> │ Pairs 3 │
#> │ Strongest |r| 0.9594 │
#> │ │
#> │ x y correlation nprs │
#> │ - - ----------- ---- │
#> │ a b 0.165 20 │
#> │ a c 0.959 20 │
#> │ b c 0.015 100 │
#> │ │
#> │ Each row is one pair of numeric columns, over the rows where both are │
#> │ present (nprs); a correlation is blank when too few pairs remain. │
#> ╰─────────────────────────────────────────────────────────────────────────╯
# A pair with too little overlap is NA, not a number from nothing.
thin <- df
thin$a[1:99] <- NA
cortbl(thin)$correlation
#> [1] NA NA 0.01508551
# one numeric column cannot be correlated
try(cortbl(data.frame(a = 1:5, b = letters[1:5])))
#> Error : cortbl: need at least two numeric columns to correlate