Calculate pairwise correlations

Calculate pairwise correlations, like Stata's pwcorr, sig obs. Each pair of variables uses every row where both are observed (pairwise deletion), and the table reports the correlation, the number of observations and the p-value.

Usage

pwcorr(
  .data,
  ...,
  .by = NULL,
  .method = c("pearson", "spearman"),
  .wide = FALSE,
  .wt = NULL,
  .wt_type = c("frequency", "analytic")
)

Arguments

.data

A data frame, tibble, or data.table.

...

<tidy-select> or <data-masking> Columns to select or compute with select_or_transmute(). Existing columns, selection helpers, and named expressions can be mixed. Results must be numeric or logical vector columns. Matrix columns are unsupported. Omit to use every numeric or logical vector column; other columns are then skipped. Selecting another type is an error.

.by

<tidy-select> Optional grouping variables. Grouped data use their groups unless .by is supplied. Select existing columns only; create new grouping columns with dplyr::mutate() first.

.method

"pearson" (default) or "spearman", like Stata's pwcorr and spearman.

.wide

If TRUE, return the correlation matrix as Stata prints it: one row per variable and the lower triangle filled, without n or p.

.wt

Optional weights: a numeric column of .data (or a vector with one value per row), like Stata's pwcorr [fweight/aweight]. Pearson correlations only. Rows with a missing weight are excluded with a message.

.wt_type

"frequency" (default) or "analytic". The correlation is the same under both; with frequency weights, n and the p-value use the summed weight, and with analytic weights the number of rows.

Value

An ungrouped tibble with:

• By default, one row per distinct variable pair and group, with any grouping keys, var1, var2, r (correlation), n (observations or summed frequency weight), and p (two-sided p-value).

• With .wide = TRUE, one row per variable and group, with grouping keys, variable, and one correlation column per variable. The diagonal is 1 and the upper triangle is NA; n and p are omitted.

Undefined correlations and p-values are NA (see Details).

Details

Computed columns are evaluated on ungrouped working data before correlations are calculated within groups. Create columns with grouped dplyr::mutate() first when a calculation needs to run separately within each group.

The p-value is the two-sided test of zero correlation that Stata reports, based on t = r * sqrt((n - 2) / (1 - r^2)) with n - 2 degrees of freedom. For Pearson correlations it matches cor.test(); for Spearman correlations it is the same large-sample approximation that Stata uses. Correlations with a constant variable, and p-values with fewer than three observations, are NA. Labelled numeric columns are correlated on their codes, excluding tagged and SPSS user-missing values. Use corr() for listwise Pearson correlations, like Stata's corr (correlate).

See also

Other tabulation functions: corr(), tab()

Examples

pwcorr(lifeexp, lexp, gnppc, safewater)
#> # A tibble: 3 x 5
#>   var1  var2          r     n        p
#>   <chr> <chr>     <dbl> <int>    <dbl>
#> 1 lexp  gnppc     0.733    67 1.79e-12
#> 2 lexp  safewater 0.760    54 2.74e-11
#> 3 gnppc safewater 0.657    53 8.95e- 8

pwcorr(lifeexp, lexp, gnppc, safewater, .wide = TRUE)
#> # A tibble: 3 x 4
#>   variable   lexp  gnppc safewater
#>   <chr>     <dbl>  <dbl>     <dbl>
#> 1 lexp      1     NA            NA
#> 2 gnppc     0.733  1            NA
#> 3 safewater 0.760  0.657         1

pwcorr(lifeexp, lexp, gnppc, .by = region)
#> # A tibble: 3 x 6
#>   region               var1  var2      r     n        p
#>   <dbl+lbl>            <chr> <chr> <dbl> <int>    <dbl>
#> 1 1 [Europe & C. Asia] lexp  gnppc 0.792    43 2.57e-10
#> 2 2 [North America]    lexp  gnppc 0.658    14 1.05e- 2
#> 3 3 [South America]    lexp  gnppc 0.640    10 4.60e- 2

pwcorr(lifeexp, lexp, gnppc, .method = "spearman")
#> # A tibble: 1 x 5
#>   var1  var2      r     n        p
#>   <chr> <chr> <dbl> <int>    <dbl>
#> 1 lexp  gnppc 0.861    67 9.53e-21

# Correlate an existing column with a computed column
pwcorr(lifeexp, lexp, gnppc_thousands = gnppc / 1000)
#> # A tibble: 1 x 5
#>   var1  var2                r     n        p
#>   <chr> <chr>           <dbl> <int>    <dbl>
#> 1 lexp  gnppc_thousands 0.733    67 1.79e-12