Quick Answer: Studentized Residual Result

The worked regression model predicted G3 final grade from six predictors: G1, G2, studytime, failures, absences and age. The model was strong overall, with R = .922, R² = .851, adjusted R² = .849, standard error of estimate = 1.254, and F(6, 642) = 609.353, p < .001. However, the purpose of Studentized Residual analysis is not simply to confirm model fit. The main purpose is to check whether some individual cases were predicted unusually poorly.

The SPSS residual diagnostics showed that the externally studentized deleted residual values ranged from approximately -7.59071 to 4.72741. Using the common diagnostic rule |external Studentized Residual| > 3, 10 cases were flagged as strong residual outlier candidates, representing about 1.5% of the dataset. Standardized residuals above |2| flagged 18 cases, and Cook’s distance above 4/n flagged 26 cases.

What Is a Studentized Residual?

A Studentized Residual is a residual divided by an estimate of its standard error. A raw residual tells how far an observed value is from its predicted value. A standardized residual divides that prediction error by the overall residual standard deviation. A Studentized Residual improves the diagnostic by also considering the fact that different cases can have different leverage and therefore different residual variance.

In regression, some observations are easier for the model to predict than others. A case with an unusual predictor pattern can have high leverage. A case can also have a very unusual outcome compared with the fitted model. Studentized Residual analysis helps identify the second issue: cases where the observed outcome is far from the model prediction after proper scaling.

The most useful version for outlier screening is often the externally Studentized Residual, also called the studentized deleted residual. It is calculated by excluding the case from the residual variance estimate. This makes it more sensitive for identifying observations that may be unusually inconsistent with the rest of the regression model.

Studentized Residual analysis belongs inside a broader regression diagnostic workflow. It should be interpreted together with residual plots, leverage, Cook’s distance, DFFITS, normality checks, homoscedasticity checks and model specification checks. Useful related guides include the Ramsey RESET test, Goldfeld-Quandt test, Q-Q plot normality check, and P-P plot normality check.

Studentized Residual Formula

The raw residual for case i is the difference between the observed outcome and the predicted outcome:

The internally Studentized Residual divides the raw residual by the residual standard error adjusted for leverage:

The externally Studentized Residual, also called the studentized deleted residual, uses the residual standard error calculated without case i:

Diagnostic Null and Alternative Hypothesis for Studentized Residual

A Studentized Residual is not usually reported as a single traditional p-value test. It is interpreted through diagnostic thresholds. However, a hypothesis-style statement helps readers understand the decision rule.

Dataset and Regression Variables Used

The worked example uses a student performance regression model. The dependent variable is G3 final grade. The predictors are G1, G2, studytime, failures, absences and age. The sample contains 649 valid cases. The model is useful for a Studentized Residual lesson because it predicts G3 strongly while still containing a small set of cases with unusual prediction errors.

Before interpreting Studentized Residuals, it is useful to understand the distribution of the variables using descriptive statistics, frequency distributions, histograms, box plots, and the five-number summary.

SPSS Output Interpretation for Studentized Residual

The SPSS output provides the main verification for the Studentized Residual example. SPSS fitted the regression model, saved residual diagnostics, produced residual descriptive statistics, tested residual normality, counted diagnostic flags, listed the largest absolute studentized deleted residuals and summarized the case-level review problem.

SPSS Regression Model Summary

SPSS Coefficients and Predictor Context

SPSS Residual Statistics

SPSS Diagnostic Flag Counts

Largest Absolute Studentized Deleted Residuals

SPSS Normality and Residual Shape

SPSS residual normality checks showed that the studentized deleted residual distribution was not perfectly normal. This is expected when the distribution contains a small number of extreme residuals. The issue does not automatically invalidate the regression model, especially with a large sample, but it does support visual inspection through histograms, Q-Q plots, P-P plots and case-level residual charts.

Python Chart-by-Chart Interpretation

The Python charts below show the complete Studentized Residual diagnostic workflow. They include residuals versus fitted values, case-index residuals, residual distribution, leverage comparison, top residual cases, diagnostic flag counts and Cook’s distance support.

Python Chart 1: Studentized Residuals vs Fitted Values

This chart compares Studentized Residuals with fitted G3 values. A good residual plot should mostly show points scattered around zero without strong curves, funnels or systematic clusters. In this example, most cases sit near the center, but several points fall far below the usual ±2 and ±3 review zones. These negative residuals show cases where the model predicted a higher G3 score than the student actually received.

The chart is important because a single residual table does not show where unusual cases occur across the fitted range. If extreme residuals appear only in one fitted-value region, it may suggest model misspecification, nonlinear relationships, subgroup problems or missing predictors. The Python plot supports the SPSS conclusion that the model is generally strong but has a small set of unusual prediction errors.

Python Chart 2: External Studentized Residual Index

The index chart makes it easier to locate case-level spikes. The strongest negative residuals appear as sharp downward points, while one strong positive residual appears above the upper threshold. This matches the SPSS list of largest absolute studentized deleted residuals, where case 164, case 173, case 587, case 640, case 520, case 638, case 641, case 584, case 62 and case 627 were among the most important review cases.

This chart is especially useful for audit work. A researcher can connect the case index to the original record, check whether the values are valid, and decide whether the residual is caused by data entry, real unusual behavior, a subgroup effect or a missing explanatory variable.

Python Chart 3: Studentized Residual Distribution

The distribution chart shows that most Studentized Residuals are near zero, which is expected in a fitted regression model. However, the distribution has clear tail behavior caused by extreme residual cases. The negative tail is especially important because several students had much lower G3 scores than the model predicted from G1, G2 and other predictors.

This chart supports the SPSS residual statistics. The residual distribution is centered near zero, but it is not perfectly normal because of extreme cases. In a large sample, the model may still be useful, but the diagnostic report should mention the unusual residual cases.

Python Chart 4: Leverage vs Absolute Studentized Residual

This chart combines two diagnostic ideas. Studentized Residuals show outcome unusualness, while leverage shows predictor-profile unusualness. A case with a high absolute Studentized Residual but low leverage is mainly an outcome outlier. A case with high leverage but a small residual may affect the fitted model shape but not appear as a large prediction error. A case with both high leverage and high absolute residual is a stronger diagnostic concern.

The plot helps distinguish “unusual outcome” from “influential case.” This matters because not every residual outlier changes the regression model strongly. Cook’s distance is then used as support to decide whether the unusual residual also has influence.

Python Chart 5: Top External Studentized Residuals

This chart ranks the strongest residual outlier candidates. The largest negative cases are students whose final G3 grade was much lower than the model predicted. The strongest positive case is a student whose G3 grade was much higher than predicted. The most extreme case in the SPSS list is case 164, with an external Studentized Residual near -7.59.

The chart is useful for reporting because it gives a readable case-level summary. Instead of saying only “10 cases exceeded |3|,” the chart shows which cases were most important and in which direction the model failed to predict accurately.

Python Chart 6: Diagnostic Flag Counts

The diagnostic flag count chart summarizes the main decision rules. The SPSS output showed 18 cases above |standardized residual| > 2, 10 cases above |external Studentized Residual| > 3, and 26 cases above Cook’s distance > 4/n. These counts show that the dataset contains a small but important group of cases requiring review.

The chart also shows why multiple diagnostics are needed. Cook’s distance flagged more cases than external Studentized Residuals. This means some cases may be influential because of leverage and residual combination even if they are not among the most extreme residual outliers.

Python Chart 7: Cook’s Distance Support

This chart links Studentized Residual size with influence. A large residual alone means the model predicted poorly for a case. A large Cook’s distance means the case may also affect the fitted regression coefficients or predicted values. The most important review cases are those that combine large absolute Studentized Residuals with nontrivial Cook’s distance.

In the SPSS output, case 164 had both the most extreme external Studentized Residual and the highest Cook’s distance. This makes it a priority case for review. However, even this does not mean automatic deletion. The correct approach is to inspect the original record, understand why it is unusual, run sensitivity analysis if needed, and report the decision transparently.

R Chart-by-Chart Validation

The R charts validate the Python and SPSS conclusions using a separate workflow. The R visual pattern is the same: the model fits G3 strongly, most residuals are close to zero, and a small group of cases has large external Studentized Residuals. This software-to-software agreement strengthens confidence in the diagnostic interpretation.

R Chart 1: Studentized Residuals vs Fitted Values

The R residuals-versus-fitted chart confirms the Python pattern. Most cases cluster around the zero residual line, but a small number of cases sit far from the expected range. The extreme negative cases show that the model predicted higher G3 values than the observed values for those students.

The R chart validates that the residual pattern is not a software artifact. It is a real diagnostic feature of the regression model and should be discussed in the residual diagnostics section of the report.

R Chart 2: External Studentized Residual Index

The R index chart confirms the case-level residual spikes. The same strong negative residual group appears, and the same diagnostic logic applies: cases beyond the ±3 threshold should be reviewed as strong residual outlier candidates.

Case-index charts are useful because they allow the analyst to trace back from the plotted diagnostic point to the original row in the dataset. That is essential for transparent outlier review.

R Chart 3: Studentized Residual Distribution

The R distribution chart confirms that the Studentized Residuals are centered near zero but contain long-tail cases. This supports the SPSS descriptive statistics, where the studentized deleted residual mean was close to zero and the standard deviation was close to one, but the minimum and maximum values were extreme.

The chart supports a balanced conclusion: the model is useful overall, but a few cases create tail behavior that should be reviewed and disclosed.

R Chart 4: Leverage vs Absolute Studentized Residual

The R leverage chart confirms that residual size and leverage are different diagnostic dimensions. A case can have a large residual without high leverage, or high leverage without a large residual. The strongest follow-up cases are those that are unusual in both outcome and predictor pattern.

This chart should be interpreted with Cook’s distance because influence is a combination of residual size and leverage. A large residual is important, but influence depends on whether the case meaningfully changes the fitted model.

R Chart 5: Top External Studentized Residuals

The R top-case chart validates the same review list. It confirms that the outlier concern is not caused by one software package. The largest negative residuals correspond to students whose final grades were much lower than the model predicted.

The visual ranking is helpful for final reporting because it turns a long residual output table into a clear diagnostic summary. The analyst can identify the most important cases and discuss the review process.

R Chart 6: Diagnostic Flag Counts

The R flag-count chart validates the same diagnostic summary as SPSS and Python. A small percentage of cases were flagged by residual and influence rules. This is common in many real datasets, but it must be disclosed when reporting regression diagnostics.

Flag counts are also useful because they prevent overreacting to a single case. The report can say how many cases exceeded each rule, then discuss whether those cases materially affected the analysis.

R Chart 7: Cook’s Distance Support

The R Cook’s distance chart confirms that not all residual outliers are equally influential. Cases with high residuals and meaningful Cook’s distance are the most important for sensitivity review. This chart supports the conclusion that the residual diagnostics require case review, not automatic model rejection.

The final R validation message matches Python and SPSS: the model performs strongly overall, but the Studentized Residual diagnostics identify a small set of cases that should be checked carefully.

SPSS, R, Python and Excel Workflows for Studentized Residual

The same Studentized Residual workflow can be reproduced in SPSS, R, Python and Excel. SPSS can save studentized deleted residuals directly. R and Python can calculate externally Studentized Residuals through model influence functions. Excel can calculate a simplified residual diagnostic workflow by computing residuals, leverage, standard errors and internally Studentized Residuals manually.

SPSS Workflow

R Workflow

Python Workflow

Excel Workflow

Code Blocks for Studentized Residual

SPSS Syntax for Studentized Residual

* Studentized Residual and regression diagnostics in SPSS.
* Dependent variable: G3.
* Predictors: G1 G2 studytime failures absences age.

TITLE "Studentized Residual Regression Diagnostics".

REGRESSION
  /DEPENDENT G3
  /METHOD=ENTER G1 G2 studytime failures absences age
  /STATISTICS COEFF OUTS R ANOVA COLLIN TOL
  /SAVE PRED RESID ZRESID SRESID DRESID SDRESID MAHAL COOK LEVER
  /CRITERIA=PIN(.05) POUT(.10)
  /NOORIGIN.

* Flag diagnostic cases.
COMPUTE flag_abs_standardized_gt_2 = (ABS(ZRE_1) > 2).
COMPUTE flag_abs_studentized_deleted_gt_3 = (ABS(SDR_1) > 3).
COMPUTE flag_cooks_distance_gt_4_over_n = (COO_1 > (4 / 649)).
EXECUTE.

FREQUENCIES VARIABLES=
  flag_abs_standardized_gt_2
  flag_abs_studentized_deleted_gt_3
  flag_cooks_distance_gt_4_over_n.

EXAMINE VARIABLES=SDR_1 ZRE_1 RES_1
  /PLOT BOXPLOT HISTOGRAM NPPLOT
  /COMPARE GROUPS
  /STATISTICS DESCRIPTIVES
  /CINTERVAL 95
  /MISSING LISTWISE
  /NOTOTAL.

SORT CASES BY SDR_1 (A).

OUTPUT EXPORT
  /CONTENTS EXPORT=VISIBLE
  /PDF DOCUMENTFILE="Studentized-Residuals-SPSS-Output.pdf".

Python Code for Studentized Residual

import pandas as pd
import numpy as np
import statsmodels.api as sm

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

dependent = "G3"
predictors = ["G1", "G2", "studytime", "failures", "absences", "age"]

model_data = df[[dependent] + predictors].apply(pd.to_numeric, errors="coerce").dropna()

X = sm.add_constant(model_data[predictors])
y = model_data[dependent]

model = sm.OLS(y, X).fit()
influence = model.get_influence()

diagnostics = model_data.copy()
diagnostics["predicted_G3"] = model.fittedvalues
diagnostics["raw_residual"] = model.resid
diagnostics["standardized_residual"] = influence.resid_studentized_internal
diagnostics["external_studentized_residual"] = influence.resid_studentized_external
diagnostics["leverage"] = influence.hat_matrix_diag
diagnostics["cooks_distance"] = influence.cooks_distance[0]
diagnostics["abs_external_studentized"] = diagnostics["external_studentized_residual"].abs()

n = len(diagnostics)

diagnostics["flag_abs_standardized_gt_2"] = diagnostics["standardized_residual"].abs() > 2
diagnostics["flag_abs_external_studentized_gt_3"] = diagnostics["external_studentized_residual"].abs() > 3
diagnostics["flag_cooks_gt_4_over_n"] = diagnostics["cooks_distance"] > (4 / n)

print(model.summary())

summary_table = diagnostics[
    [
        "raw_residual",
        "standardized_residual",
        "external_studentized_residual",
        "leverage",
        "cooks_distance"
    ]
].describe()

print(summary_table)

flag_counts = diagnostics[
    [
        "flag_abs_standardized_gt_2",
        "flag_abs_external_studentized_gt_3",
        "flag_cooks_gt_4_over_n"
    ]
].sum()

print(flag_counts)

top_cases = diagnostics.sort_values("abs_external_studentized", ascending=False).head(20)
print(top_cases[["G1", "G2", "G3", "predicted_G3", "raw_residual",
                 "external_studentized_residual", "leverage", "cooks_distance"]])

R Code for Studentized Residual

# Studentized Residual analysis in R

df <- read.csv("dataset.csv")

vars_needed <- c("G3", "G1", "G2", "studytime", "failures", "absences", "age")
df_model <- df[vars_needed]
df_model[] <- lapply(df_model, as.numeric)
df_model <- na.omit(df_model)

model <- lm(G3 ~ G1 + G2 + studytime + failures + absences + age, data = df_model)

summary(model)

diagnostics <- data.frame(
  predicted_G3 = fitted(model),
  raw_residual = resid(model),
  standardized_residual = rstandard(model),
  external_studentized_residual = rstudent(model),
  leverage = hatvalues(model),
  cooks_distance = cooks.distance(model),
  dffits = dffits(model)
)

diagnostics$abs_external_studentized <- abs(diagnostics$external_studentized_residual)

n <- nrow(diagnostics)

diagnostics$flag_abs_standardized_gt_2 <- abs(diagnostics$standardized_residual) > 2
diagnostics$flag_abs_external_studentized_gt_3 <- abs(diagnostics$external_studentized_residual) > 3
diagnostics$flag_cooks_gt_4_over_n <- diagnostics$cooks_distance > (4 / n)

summary(diagnostics)

colSums(diagnostics[c(
  "flag_abs_standardized_gt_2",
  "flag_abs_external_studentized_gt_3",
  "flag_cooks_gt_4_over_n"
)])

top_cases <- diagnostics[order(-diagnostics$abs_external_studentized), ][1:20, ]
print(top_cases)

Excel Formulas for Studentized Residual

Step 1:
Run multiple regression and save predicted values.

Step 2:
Calculate raw residual:
=Observed_G3 - Predicted_G3

Step 3:
Calculate residual standard error:
=SQRT(SSE / residual_df)

Step 4:
Calculate leverage for each row:
hii = diagonal value from X(X'X)^-1X'

Step 5:
Calculate internally Studentized Residual:
=Residual / (Residual_Standard_Error * SQRT(1 - Leverage))

Step 6:
Flag review cases:
=IF(ABS(Studentized_Residual)>3,"Strong review","OK")

Step 7:
Cook's distance support:
CookD = (Studentized_Residual^2 / number_of_predictors) * (Leverage / (1-Leverage))

Step 8:
Flag Cook's distance:
=IF(CookD>4/n,"Influence review","OK")

APA Reporting Wording for Studentized Residual

When reporting a Studentized Residual diagnostic, describe the model, residual rule, number of flagged cases, largest residual range, and whether influence was supported by Cook’s distance. Do not report only “outliers were removed” unless removal was justified and the analysis was repeated transparently.

APA-Style Full Report

Short APA-Style Version

Case Review Wording

Common Mistakes in Studentized Residual Interpretation

When to Use Studentized Residuals

Use Studentized Residuals when you run regression and need to identify cases that the model predicts poorly. They are especially useful when the goal is assumption checking, outlier review, influence screening, model improvement or transparent reporting.

Studentized Residuals should be used with other diagnostics. For variance and group assumption checks, see Levene’s test, Brown-Forsythe test, and Cochran’s C test. For normality checks, see the Kolmogorov-Smirnov test, Lilliefors test, D’Agostino-Pearson test, and Cramer-von Mises test.

Downloads and Resources for Studentized Residual

The SPSS output PDF below verifies the regression model, residual statistics, Studentized Residuals, standardized residuals, Cook’s distance, leverage, diagnostic flag counts, largest residual cases and model-diagnostic interpretation used in this guide.

Related Statistical Guides

Use these related guides to connect Studentized Residual with regression diagnostics, model assumptions, normality checks, descriptive analysis and reporting.