pcorr 0.6.0

Pairwise Pearson correlations with p-values and multiple-comparison correction

pcorr

Compute all pairwise Pearson/Spearman correlations between numeric columns in a pandas.DataFrame — like pandas.DataFrame.corr(), but with p-values for each pair and multiple-comparison correction in one call.

Installation

pip install pcorr
pip install "pcorr[full]"

Usage

import pandas as pd
from pcorr import corr_pairwise, corr_table, show_table

df = pd.read_csv("data.csv")

# Long (tidy) format: one row per pair
table = corr_pairwise(df, method="fdr_bh", alpha=0.05)
print(table)
#   var1 var2    n      r  p_corrected  p_value  significant
# 0    x    y  200  0.967       0.0000  0.0000         True
# 1    x    z  200 -0.894       0.0000  0.0000         True
# ...

# Two square tables (r and p) rendered as lower-triangle output
tables = corr_table(df, method="bonferroni")
tables["r"]  # coefficients
tables["p"]  # p-values (raw when method="none", otherwise corrected)

# Convenience display (renders nicely in Jupyter, prints in console)
show_table(df, method="bonferroni")

API

• corr_pairwise(df, ...) — tidy table: one row per column pair.
• corr_table(df, ...) — two square tables (coefficients and p-values) in a lower-triangle style.
• show_table(df, ...) — displays/prints corr_table(...).
• corr_matrices(df, ...) — alias for corr_table(...) (compatibility).

Output format

corr_pairwise

Always returns:

• var1, var2 — column names
• n — number of valid observations in the pair (after pairwise NaN deletion)
• r — correlation coefficient
• significant — p < alpha for the p-value used for significance

P-value columns depend on method:

• method="none" returns p_value (raw p-value)
• any correction (e.g. bonferroni, fdr_bh, holm) returns p_corrected and p_value (raw)

Raw p-values example:

from pcorr import corr_pairwise
out = corr_pairwise(df, method="none")

corr_table / corr_matrices

Returns a dict with two DataFrames:

• tables["r"] — coefficients (lower triangle filled), diagonal "1", upper triangle blank
• tables["p"] — p-values in the same layout, diagonal "—"
  • raw p-values when method="none"
  • corrected p-values for any correction method

Parameters

parameter	meaning
columns	which columns to use (default: all numeric)
corr	"pearson" or "spearman"
method	bonferroni, fdr_bh, holm, sidak, none, ...
alpha	threshold for significant
min_n	minimum valid observations per pair (pairwise NaN deletion)
round_to	rounding for r/p columns (None disables rounding)

Notes

• Pairwise NaN deletion: each pair has its own n.
• Constant columns and pairs with n < min_n are handled without crashing.
• corr_pairwise is sorted by the p-value used for ranking (corrected when applicable).

P-value correction methods

• method="none" and method="bonferroni" work without statsmodels.
• Other methods use statsmodels.stats.multitest.multipletests and require statsmodels (install via the full extra).