Cook’s Distance

Cook’s Distance is a measure of an observation or instances’ influence on a linear regression. Instances with a large influence may be outliers and datasets that have a large number of highly influential points might not be good predictors to fit linear models. The CooksDistance visualizer shows a stem plot of all instances by index and their associated distance score, along with a heuristic threshold to quickly show what percent of the dataset may be impacting OLS regression models.

from yellowbrick.regressor import CooksDistance
from yellowbrick.datasets import load_concrete

# Load the regression dataset
X, y = load_concrete()

# Instantiate and fit the visualizer
visualizer = CooksDistance(), y)

(Source code, png, pdf)

Cook's distance using concrete dataset

API Reference

Visualize the influence and leverage of individual instances on a regression model.

class yellowbrick.regressor.influence.CooksDistance(ax=None, draw_threshold=True, linefmt='C0-', markerfmt=', ', **kwargs)[source]

Bases: yellowbrick.base.Visualizer

Cook’s Distance is a measure of how influential an instance is to the computation of a regression, e.g. if the instance is removed would the estimated coeficients of the underlying model be substantially changed? Because of this, Cook’s Distance is generally used to detect outliers in standard, OLS regression. In fact, a general rule of thumb is that D(i) > 4/n is a good threshold for determining highly influential points as outliers and this visualizer can report the percentage of data that is above that threshold.

This implementation of Cook’s Distance assumes Ordinary Least Squares regression, and therefore embeds a sklearn.linear_model.LinearRegression under the hood. Distance is computed via the non-whitened leverage of the projection matrix, computed inside of fit(). The results of this visualizer are therefore similar to, but not as advanced, as a similar computation using statsmodels. Computing the influence for other regression models requires leave one out validation and can be expensive to compute.

See also

For a longer discussion on detecting outliers in regression and computing leverage and influence, see linear regression in python, outliers/leverage detect by Huiming Song.

ax : matplotlib Axes, default: None

The axes to plot the figure on. If None is passed in the current axes will be used (or generated if required).

draw_threshold : bool, default: True

Draw a horizontal line at D(i) == 4/n to easily identify the most influential points on the final regression. This will also draw a legend that specifies the percentage of data points that are above the threshold.

linefmt : str, default: ‘C0-‘

A string defining the properties of the vertical lines of the stem plot, usually this will be a color or a color and a line style. The default is simply a solid line with the first color of the color cycle.

markerfmt : str, default: ‘,’

A string defining the properties of the markers at the stem plot heads. The default is “pixel”, e.g. basically no marker head at the top of the stem plot.

kwargs : dict

Keyword arguments that are passed to the base class and may influence the final visualization (e.g. size or title parameters).


Cook’s Distance is very similar to DFFITS, another diagnostic that is meant to show how influential a point is in a statistical regression. Although the computed values of Cook’s and DFFITS are different, they are conceptually identical and there even exists a closed-form formula to convert one value to another. Because of this, we have chosen to implement Cook’s distance rather than or in addition to DFFITS.

distance_ : array, 1D

The Cook’s distance value for each instance specified in X, e.g. an 1D array with shape (X.shape[0],).

p_values_ : array, 1D

The p values associated with the F-test of Cook’s distance distribution. A 1D array whose shape matches distance_.

influence_threshold_ : float

A rule of thumb influence threshold to determine outliers in the regression model, defined as It=4/n.

outlier_percentage_ : float

The percentage of instances whose Cook’s distance is greater than the influnce threshold, the percentage is 0.0 <= p <= 100.0.


Draws a stem plot where each stem is the Cook’s Distance of the instance at the index specified by the x axis. Optionaly draws a threshold line.


Prepares the visualization for presentation and reporting.

fit(self, X, y)[source]

Computes the leverage of X and uses the residuals of a sklearn.linear_model.LinearRegression to compute the Cook’s Distance of each observation in X, their p-values and the number of outliers defined by the number of observations supplied.

X : array-like, 2D

The exogenous design matrix, e.g. training data.

y : array-like, 1D

The endogenous response variable, e.g. target data.

self : CooksDistance

Fit returns the visualizer instance.