Quick Answer: Cook’s Distance Result

Cook’s Distance, often written as Cook’s D, measures how much a single observation influences a regression model. It combines information from residual size and leverage. A point with a large residual may be unusual in the outcome direction. A point with high leverage may be unusual in the predictor space. A point with both high residual influence and high leverage can become highly influential.

In practical regression diagnostics, Cook’s Distance is not interpreted alone. It should be read together with residual plots, leverage diagnostics, DFBETAs, predicted values, model fit, subject-matter meaning, and data-entry checks. If a case has a high Cook’s Distance, the correct response is not automatic deletion. The correct response is investigation: check whether the case is a data error, a valid but unusual observation, a subgroup pattern, or evidence that the model form is incomplete.

Hypothesis-style interpretation: Cook’s Distance is not a traditional null-hypothesis test like a one-sample z test or one-tailed t test. Instead, it is a diagnostic rule. The working null assumption is that no single case has excessive influence on the regression result. The diagnostic alternative is that one or more cases exert enough influence to deserve investigation, sensitivity analysis, or separate reporting.

What Is Cook’s Distance?

Cook’s Distance is a regression influence statistic that estimates how much the fitted regression model would change if a particular case were removed. It is especially useful in linear regression and multiple regression because a single unusual observation can change the slope, intercept, fitted values, p-values, confidence intervals, and practical conclusion.

Cook’s Distance is different from a simple outlier check. A case may have a large residual but low leverage, meaning it is unusual in the outcome value but not necessarily powerful enough to change the regression line. Another case may have high leverage but a small residual, meaning it sits far away in predictor space but still follows the model pattern. The most concerning cases often combine high leverage with a large residual. Cook’s Distance summarizes that joint influence.

Cook’s Distance belongs to the broader family of regression diagnostics. It should be interpreted alongside residuals versus fitted plots, leverage plots, normality diagnostics, variance checks, model specification checks, and assumption tests. For related checks, compare this guide with Ramsey RESET Test, Goldfeld-Quandt Test, Brown-Forsythe Test, Levene Test, Q-Q Plot Normality Check, and Histogram Interpretation.

Cook’s Distance Formula

The most common Cook’s Distance formula for case i can be written as:

Here, ŷ_j is the fitted value from the model using all observations, ŷ_j(i) is the fitted value after removing case i, p is the number of predictors, and MSE is the mean squared error. This version emphasizes the idea that Cook’s Distance measures how much the fitted values change when a case is removed.

A commonly used leverage-residual form is:

Here, r_i is a standardized or studentized residual and h_ii is leverage for case i. This formula shows why Cook’s Distance increases when a case has both a large residual and high leverage.

Null and Alternative Hypothesis for Cook’s Distance

Cook’s Distance is a diagnostic statistic rather than a formal hypothesis test. Still, it can be explained in a hypothesis-style decision framework for reporting and teaching.

Dataset and Variables Used

This Cook’s Distance guide is written for a regression dataset where one continuous dependent variable is predicted from one or more independent variables. The exact model may be simple linear regression or multiple linear regression. Cook’s Distance becomes more important as models become more sensitive to unusual combinations of predictor values and outcome values.

Before interpreting Cook’s Distance, review the dataset with descriptive statistics, five-number summary, frequency distribution, histogram interpretation, and box plot interpretation. These summaries help identify whether flagged cases are extreme, plausible, miscoded, or part of a meaningful subgroup.

Verified SPSS Output Interpretation

The SPSS output for this Cook’s Distance guide is available here: Cook’s Distance SPSS Output PDF.

In SPSS, Cook’s Distance is usually generated from the Linear Regression procedure by saving influence statistics. After running the regression model, SPSS can save Cook’s Distance, leverage values, predicted values, residuals, standardized residuals, studentized residuals, and other diagnostic variables into the dataset. These saved variables can then be inspected in tables, sorted by size, and plotted against fitted values or leverage.

SPSS Interpretation Table

SPSS Decision Summary

Python Chart-by-Chart Interpretation

The Python charts show the Cook’s Distance workflow visually. They explain influence by case, leverage-residual influence, residuals versus fitted values, top cases, distribution of Cook’s Distance, leverage versus Cook’s Distance, and DFBETAs compared with Cook’s Distance. Each chart below includes a figure block, descriptive alt text, title, lazy loading, caption, detailed interpretation, and decision/reporting conclusion.

Python Chart 1: Cook’s Distance by Case

Detailed interpretation: This chart plots Cook’s Distance for every case in the regression dataset. Most cases usually have small influence values near zero. Cases that rise above the rest deserve attention because they have more potential to change the regression result. If a reference line such as 4/n or 1 is included, cases above the line should be checked more carefully. The chart is useful because it shows whether influence is spread across many cases or concentrated in a few observations.

Decision/reporting conclusion: Use this chart to identify candidate influential cases. Cases with unusually high Cook’s Distance should be investigated through raw data checking, residual inspection, leverage inspection, and sensitivity analysis.

Python Chart 2: Leverage, Residual and Cook Bubble Plot

Detailed interpretation: This bubble plot is one of the most informative Cook’s Distance visuals. The horizontal axis usually represents leverage, while the vertical axis represents residual size. Bubble size represents Cook’s Distance. A case with high leverage but a small residual may not be highly influential. A case with a large residual but low leverage may be an outcome outlier but may not strongly affect the regression line. The most important points are often large bubbles located in high-leverage or high-residual regions.

Decision/reporting conclusion: Use this chart to decide whether flagged cases are influential because of leverage, residual size, or both. Large bubbles should be examined first in the sensitivity analysis.

Python Chart 3: Residuals vs Fitted Values with Flagged Cases

Detailed interpretation: The residuals versus fitted chart checks model fit and influential case behavior. A good linear regression pattern usually has residuals scattered around zero without a strong curve or funnel shape. Flagged influential cases in this plot show where the model fits poorly or where unusual observations may affect the fitted regression line. If flagged cases are also part of a curved residual pattern, the issue may be model misspecification rather than only isolated influence.

Decision/reporting conclusion: Use this chart to decide whether influential cases are isolated observations or symptoms of a broader model problem. If residual patterns show curvature, consider checking model specification with a guide such as the Ramsey RESET Test.

Python Chart 4: Top Cases by Cook’s Distance

Detailed interpretation: This chart ranks the most influential cases from highest to lower Cook’s Distance. It is useful because it turns a full diagnostic list into a focused investigation list. Instead of reviewing every observation, the researcher can begin with the most influential cases. The ranking also helps determine whether one case dominates the influence pattern or whether several cases have similar influence.

Decision/reporting conclusion: Use this chart to document the top influential cases. Investigate the top-ranked cases for data errors, unusual predictor combinations, extreme residuals, and impact on regression coefficients.

Python Chart 5: Cook’s Distance Distribution

Detailed interpretation: The Cook’s Distance distribution chart shows whether influence values are generally small or whether the distribution has a long right tail. In many regression datasets, most Cook’s Distance values cluster near zero, with a small number of cases extending to the right. A long right tail indicates that a few cases are much more influential than the majority. This chart complements the case-by-case plot by showing the overall influence distribution.

Decision/reporting conclusion: Use this chart to summarize the influence pattern. If the distribution has a clear extreme tail, report that selected cases were examined further through sensitivity checks.

Python Chart 6: Leverage vs Cook’s Distance

Detailed interpretation: This chart directly compares leverage with Cook’s Distance. High leverage means that a case is unusual in predictor space, but high leverage alone does not always mean high influence. If points with high leverage also have high Cook’s Distance, they are more important for diagnosis. If leverage is high but Cook’s Distance remains low, the case may be unusual but still consistent with the fitted regression pattern.

Decision/reporting conclusion: Use this chart to separate high-leverage cases from truly influential cases. Cases in the upper-right region deserve the strongest diagnostic attention.

Python Chart 7: DFBETAs vs Cook’s Distance

Detailed interpretation: DFBETAs show how much a regression coefficient changes when a case is removed. Cook’s Distance provides an overall influence measure, while DFBETAs help identify which predictor coefficient is affected. A case with high Cook’s Distance and large DFBETAs may be changing the slope of one or more predictors. This chart is especially useful in multiple regression, where one influential case may affect some coefficients more than others.

Decision/reporting conclusion: Use this chart to move beyond “which case is influential” toward “which coefficient is affected.” Report DFBETAs when influential cases change the interpretation of specific predictors.

R Chart-by-Chart Validation

The R charts validate the Cook’s Distance analysis using a separate statistical environment. R regression diagnostics are widely used because base R and related packages provide influence measures such as Cook’s Distance, leverage, residuals and DFBETAs. The R charts below mirror the Python workflow and confirm the influence-diagnostic story.

R Chart 1: Cook’s Distance by Case

Detailed interpretation: The R case-level Cook’s Distance chart confirms which observations stand out in the regression model. Consistency between Python and R increases confidence that the flagged cases are not software artifacts. Most cases should have low Cook’s Distance, while influential observations rise above the general pattern.

Decision/reporting conclusion: Use this R chart as validation for the influential case list. Cases that stand out in both Python and R should receive priority in sensitivity analysis.

R Chart 2: Leverage, Residual and Cook Bubble Plot

Detailed interpretation: The R bubble plot validates whether influential cases are driven by leverage, residual size, or both. Large bubbles in extreme regions are the most important for diagnostics. If the R chart matches the Python chart, the influence pattern is stable across software and should be taken seriously.

Decision/reporting conclusion: Use this chart to confirm the mechanism of influence. Report whether flagged cases were high leverage, large residual, or both.

R Chart 3: Residuals vs Fitted Values with Flagged Cases

Detailed interpretation: The R residuals versus fitted chart confirms whether flagged influential cases occur in areas of poor model fit. If residuals are randomly scattered around zero, the model form is more defensible. If residuals show curvature or changing spread, influence may be connected to nonlinearity or heteroscedasticity rather than isolated cases only.

Decision/reporting conclusion: Use this chart to decide whether the issue is case influence, model form, unequal variance, or a combination. For variance-related checks, compare with Goldfeld-Quandt Test, Levene Test, and Brown-Forsythe Test.

R Chart 4: Top Cases by Cook’s Distance

Detailed interpretation: The R ranking chart confirms which cases have the largest Cook’s Distance values. This ranking is helpful for direct case review, especially when regression output contains many rows. The most influential cases should be checked for data-entry errors, unusual predictor combinations, and coefficient impact.

Decision/reporting conclusion: Use the R top-cases chart to validate the final influential case list. If the same top cases appear across platforms, report them as the primary influence candidates.

R Chart 5: Cook’s Distance Distribution

Detailed interpretation: The R distribution chart validates whether influence is concentrated in a small number of cases. A distribution clustered near zero with a right tail suggests that most cases have little influence while a few cases deserve closer inspection. This chart is helpful for summarizing the diagnostic pattern in a simple visual.

Decision/reporting conclusion: Use this chart to report whether the regression model is influenced by a small number of cases or whether influence is generally low across the dataset.

R Chart 6: Leverage vs Cook’s Distance

Detailed interpretation: This R chart confirms whether high leverage is associated with high Cook’s Distance. Some high-leverage points may not be influential if they fit the model well. Other points may become influential when leverage is paired with large residuals. This distinction is essential because leverage alone is not the same as influence.

Decision/reporting conclusion: Use this chart to avoid overreacting to leverage alone. Focus on cases that combine high leverage with high Cook’s Distance or coefficient-level impact.

R Chart 7: DFBETAs vs Cook’s Distance

Detailed interpretation: The R DFBETAs versus Cook’s Distance chart validates whether influential cases affect specific regression coefficients. A high Cook’s Distance value tells us that the overall fitted model may change, while DFBETAs show which slope or intercept is affected. In multiple regression, this distinction is important because one case may strongly affect one predictor but not another.

Decision/reporting conclusion: Use this chart when the final report needs to explain which regression coefficient is sensitive to influential cases. If DFBETAs are large, report coefficient sensitivity rather than only a general Cook’s Distance warning.

SPSS, R, Python and Excel Workflows for Cook’s Distance

The Cook’s Distance workflow is similar across software. First, fit the regression model. Second, save or calculate influence diagnostics. Third, identify cases with high Cook’s Distance. Fourth, inspect leverage, residuals and DFBETAs. Finally, run sensitivity analysis and decide whether the original conclusion is stable.

SPSS Workflow

R Workflow

Python Workflow

Excel Workflow

Code Blocks for Cook’s Distance

SPSS Syntax for Cook’s Distance

* Cook's Distance in SPSS.
* Replace outcome x1 x2 x3 with your actual variables.

TITLE "Cook's Distance Regression Diagnostics".

REGRESSION
  /DEPENDENT outcome
  /METHOD=ENTER x1 x2 x3
  /STATISTICS COEFF OUTS R ANOVA COLLIN CI(95)
  /SAVE PRED(Predicted_Value)
        RESID(Residual_Value)
        ZRESID(Standardized_Residual)
        SRESID(Studentized_Residual)
        LEVER(Leverage_Value)
        COOK(Cooks_Distance)
  /SCATTERPLOT=(*ZPRED,*ZRESID)
  /RESIDUALS HISTOGRAM(ZRESID) NORMPROB(ZRESID).

SORT CASES BY Cooks_Distance(D).

LIST VARIABLES=Cooks_Distance Leverage_Value Standardized_Residual Studentized_Residual Predicted_Value Residual_Value
  /CASES=FROM 1 TO 20.

OUTPUT SAVE OUTFILE="Cooks-Distance-SPSS-Output.spv".

OUTPUT EXPORT
  /CONTENTS EXPORT=VISIBLE
  /PDF DOCUMENTFILE="Cooks-Distance-SPSS-Output.pdf".

Python Code for Cook’s Distance

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import statsmodels.api as sm

# Load data
df = pd.read_csv("dataset.csv")

# Replace these with your actual variables
y = df["outcome"]
X = df[["x1", "x2", "x3"]]
X = sm.add_constant(X)

# Fit regression model
model = sm.OLS(y, X, missing="drop").fit()
print(model.summary())

# Influence diagnostics
influence = model.get_influence()
cooks_d, p_values = influence.cooks_distance
leverage = influence.hat_matrix_diag
standard_resid = influence.resid_studentized_internal
dfbetas = influence.dfbetas

n = len(cooks_d)
threshold = 4 / n

diagnostics = pd.DataFrame({
    "case": np.arange(1, n + 1),
    "cooks_distance": cooks_d,
    "leverage": leverage,
    "standardized_residual": standard_resid
})

diagnostics["flagged"] = diagnostics["cooks_distance"] > threshold
print(diagnostics.sort_values("cooks_distance", ascending=False).head(10))

# Cook's Distance by case
plt.figure(figsize=(10, 5))
plt.stem(diagnostics["case"], diagnostics["cooks_distance"], basefmt=" ")
plt.axhline(threshold, linestyle="--", label="4/n threshold")
plt.title("Cook's Distance by Case")
plt.xlabel("Case Number")
plt.ylabel("Cook's Distance")
plt.legend()
plt.tight_layout()
plt.show()

# Leverage-residual bubble plot
plt.figure(figsize=(8, 6))
bubble_size = 1000 * diagnostics["cooks_distance"] / diagnostics["cooks_distance"].max()
plt.scatter(diagnostics["leverage"], diagnostics["standardized_residual"], s=bubble_size, alpha=0.5)
plt.axhline(0, linewidth=1)
plt.title("Leverage, Residual and Cook's Distance Bubble Plot")
plt.xlabel("Leverage")
plt.ylabel("Standardized Residual")
plt.tight_layout()
plt.show()

R Code for Cook’s Distance

# Cook's Distance in R

library(tidyverse)

# Load data
df <- read.csv("dataset.csv")

# Replace outcome, x1, x2, x3 with your actual variables
model <- lm(outcome ~ x1 + x2 + x3, data = df)

summary(model)

# Influence diagnostics
cooks_d <- cooks.distance(model)
leverage <- hatvalues(model)
std_resid <- rstandard(model)
stud_resid <- rstudent(model)
dfbetas_values <- dfbetas(model)

n <- length(cooks_d)
threshold <- 4 / n

diagnostics <- data.frame(
  case = seq_along(cooks_d),
  cooks_distance = cooks_d,
  leverage = leverage,
  standardized_residual = std_resid,
  studentized_residual = stud_resid,
  flagged = cooks_d > threshold
)

diagnostics %>%
  arrange(desc(cooks_distance)) %>%
  head(10) %>%
  print()

# Cook's Distance by case
ggplot(diagnostics, aes(x = case, y = cooks_distance)) +
  geom_col() +
  geom_hline(yintercept = threshold, linetype = "dashed") +
  labs(
    title = "Cook's Distance by Case",
    x = "Case Number",
    y = "Cook's Distance"
  )

# Leverage vs Cook's Distance
ggplot(diagnostics, aes(x = leverage, y = cooks_distance)) +
  geom_point() +
  geom_hline(yintercept = threshold, linetype = "dashed") +
  labs(
    title = "Leverage vs Cook's Distance",
    x = "Leverage",
    y = "Cook's Distance"
  )

Excel Formulas and Steps for Regression Influence Support

Excel is not the best tool for formal Cook's Distance, but it can support regression diagnostics.

Assume:
Y is in column B
Predicted Y is in column C
Residual is in column D

Residual:
=B2-C2

Squared residual:
=D2^2

Mean squared error:
=SUM(E2:E101)/(COUNT(E2:E101)-number_of_parameters)

Standardized residual approximation:
=D2/STDEV.S($D$2:$D$101)

Reference rule for Cook's Distance when calculated elsewhere:
=4/COUNT(B2:B101)

Sensitivity analysis idea:
1. Fit the model with all cases.
2. Identify flagged cases using SPSS, R or Python.
3. Refit the model without flagged cases only if justified.
4. Compare coefficients, p-values and R Square.
5. Report whether the conclusion changed.

APA Reporting Wording for Cook’s Distance

When reporting Cook’s Distance, the goal is to show that the regression was checked for influential observations. Report the diagnostic rule used, whether any cases were flagged, how flagged cases were investigated, and whether the substantive regression conclusion changed.

APA-Style Reporting Template

Short Report Sentence

Student-Friendly Report Example

For stronger reporting, combine Cook’s Distance with confidence interval, effect size, coefficient of variation, residual plots, and model assumption diagnostics.

Common Mistakes in Cook’s Distance Interpretation

When to Use Cook’s Distance

Use Cook’s Distance whenever a regression model may be affected by unusual observations. It is especially important in small or moderate samples, models with several predictors, datasets with extreme predictor values, and applied research where a few cases may strongly affect conclusions.

For related diagnostic topics, compare this guide with Ramsey RESET Test, Goldfeld-Quandt Test, Kolmogorov-Smirnov Test, D’Agostino-Pearson Test, Cramer-von Mises Test, and Ryan-Joiner Test.

Downloads and Resources for Cook’s Distance

The resources below include the SPSS output PDF, Python charts, and R validation charts used in this Cook’s Distance guide.

Related Statistical Guides

Use these related guides to connect Cook’s Distance with model diagnostics, normality checks, variance tests, effect size, confidence intervals, hypothesis testing and broader regression reporting.

Final Conclusion

Cook’s Distance is one of the most practical regression diagnostics for detecting influential observations. It helps answer whether a single case is powerful enough to affect the fitted regression model. The most useful interpretation combines Cook’s Distance with leverage, residuals, DFBETAs, residual plots, top-case rankings and sensitivity analysis.

The SPSS output, Python charts and R validation charts in this guide show a complete influence-diagnostic workflow. The correct conclusion is not simply “remove influential cases.” The correct conclusion is to investigate flagged cases, check whether they are valid, test whether the regression result changes, and report the decision transparently. This approach makes Cook’s Distance a professional diagnostic tool rather than a mechanical outlier-removal rule.

Prepared for Salar Cafe — onlineinternetcafe.com