Ranks the numeric column pairs by the strength of their association and returns the strongest, so a wide table's structure can be read without squinting at a correlation matrix. Use it on a new data set to find near-duplicate columns, derived columns and the relationships worth a closer look.
Usage
topcors(data, n = 10L, meth = c("spearman", "pearson"), minabs = 0)Value
A data frame of class c("bricklayer_correlations",
"data.frame") with x and y (character, column names),
correlation (numeric) and abscor (numeric),
strongest first, at most n rows; the method is in
attr(, "method"). The print method draws a signed bar per pair.
Details
Spearman is the default deliberately. Pearson measures LINEAR association only, so it understates a relationship that is perfectly monotone but curved, and a single outlier can manufacture or destroy it. On data you have not yet inspected – which is the situation this function is for – the rank correlation is the safer question to ask.
Non-numeric columns are ignored. Each pair uses the rows where both
values are present; a pair with fewer than three such rows, or with a
constant column, has no correlation and is left out. Pairs are ranked by
absolute correlation, those below minabs dropped, and the first
n returned. cortbl() gives every
pair, with the
number of rows behind each. 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;
n not one positive whole number; minabs not one number in
\([0, 1]\); an invalid meth.
Examples
set.seed(1)
n <- 200
df <- data.frame(
a = stats::rnorm(n),
b = stats::rnorm(n),
grade = sample(letters[1:3], n, TRUE)
)
df$c <- df$a * 2 + stats::rnorm(n, sd = 0.1) # strongly related to a
df$d <- exp(df$a) # monotone but curved
topcors(df)
#> ╭── Top correlations (spearman) ──────────────────────────────────────────────╮
#> │ Method spearman │
#> │ Pairs 6 │
#> │ │
#> │ x y correlation plot │
#> │ - - ----------- --------------------- │
#> │ a d 1.000 |########## │
#> │ a c 0.998 |########## │
#> │ c d 0.998 |########## │
#> │ a b -0.062 #| │
#> │ b d -0.062 #| │
#> │ b c -0.060 #| │
#> │ │
#> │ The pairs most strongly associated, strongest first; a correlation near │
#> │ 1 or -1 often means one column is derived from the other. │
#> ╰─────────────────────────────────────────────────────────────────────────────╯
# The curved pair is ranked correctly by Spearman and understated by
# Pearson, which is why Spearman is the default.
tc <- topcors(df, meth = "spearman")
tc[tc$x == "a" & tc$y == "d", "correlation"]
#> [1] 1
tp <- topcors(df, meth = "pearson")
tp[tp$x == "a" & tp$y == "d", "correlation"]
#> [1] 0.8496382
# Filter to the pairs worth looking at, and keep the top two.
topcors(df, n = 2, minabs = 0.5)
#> ╭── Top correlations (spearman) ──────────────────────────────────────────────╮
#> │ Method spearman │
#> │ Pairs 2 │
#> │ │
#> │ x y correlation plot │
#> │ - - ----------- --------------------- │
#> │ a d 1.000 |########## │
#> │ a c 0.998 |########## │
#> │ │
#> │ The pairs most strongly associated, strongest first; a correlation near │
#> │ 1 or -1 often means one column is derived from the other. │
#> ╰─────────────────────────────────────────────────────────────────────────────╯
# a threshold above 1 is an error
try(topcors(df, minabs = 2))
#> Error : `minabs` must be a single value in [0, 1]