Does studentized mean the observation was removed?

Only the external scale estimate uses deletion; the internal version uses the full fit.

Can the residual be negative?

Yes. Its sign indicates whether the observation lies below or above its fitted value.

Does a large value prove influence?

No. Influence also depends on leverage and how the fitted model changes.

AI Glossary / Statistics and probability / Intermediate

Studentized residual

Definition

A studentized residual is a regression residual divided by an estimated standard deviation appropriate to that residual. It expresses an observation's discrepancy from the fitted model on a comparable scale while accounting for the fact that residual variability differs with leverage. Internally studentized residuals use the full-data error estimate; externally studentized residuals use an estimate obtained with the observation omitted. Terminology varies across software, so the formula and version should be stated.

What it means

A raw residual is observed outcome minus fitted outcome. Its size depends on measurement units and on where the observation lies in the predictor space. In ordinary least squares, a high-leverage observation partly pulls the fitted relationship toward itself, affecting the residual's variance.

Studentization adjusts the scale of the residual rather than changing the underlying observation. The external version reduces the observation's influence on the error-scale estimate used to assess that same observation. These quantities help identify cases that deserve investigation, but they do not prove that a case is erroneous. They also measure something different from leverage or overall influence on fitted coefficients.

How it works

For an ordinary least-squares model with n observations and p fitted coefficients, including any intercept, let e_i be the residual and h_ii the leverage. Under the usual constant-variance model, residual variance is sigma squared times (1 − h_ii).

Estimate sigma by s = sqrt[SSE / (n − p)]. The internally studentized residual is r_i = e_i / [s sqrt(1 − h_ii)]. For the external version, replace s by s_(i), the residual standard deviation from fitting without observation i. The resulting t_i uses e_i from the full fit together with the corresponding leverage adjustment.

Under the classical independent normal-error model and appropriate full-rank conditions, an individual externally studentized residual has a t reference distribution with n − p − 1 degrees of freedom. The internal version does not have that same exact reference distribution.

Inspect residual plots, leverage, and influence measures together. If many observations are screened, account for multiplicity when making formal outlier claims. Verify data quality and consider model misspecification before excluding a flagged case. Document the reason for any change and assess sensitivity to it.

Three practical examples

1. An analyst compares residuals from observations with different leverages rather than ranking raw errors alone.

2. An AI team investigates an unusual prediction error in a linear baseline and checks whether it reflects a valid rare case or a recording problem.

3. A quality engineer combines studentized residuals with Cook's distance to distinguish unusual outcomes from observations that strongly affect the fitted relationship.

Think of it this way

Think of judging how unusual a measurement is using a ruler calibrated for that measurement's expected variability. The external version recalibrates the ruler without letting the case under examination set its own scale. The result helps focus attention, but it does not tell you automatically why the measurement is unusual.

Worked example

Consider an original ordinary-least-squares teaching calculation with n = 10, p = 2, and full-model SSE = 32. The full residual variance estimate is 32 / (10 − 2) = 4, so s = 2. Suppose one observation has residual e_i = 3 and leverage h_ii = 0.25.

Its internally studentized residual is 3 / [2 sqrt(0.75)], about 1.732. Merely dividing by s would give 1.5 and omit the leverage adjustment.

For this full-rank setting, the deletion identity gives SSE_(i) = SSE − e_i squared / (1 − h_ii) = 32 − 9 / 0.75 = 20. After deletion, residual degrees of freedom are 10 − 2 − 1 = 7. Thus s_(i) = sqrt(20 / 7), about 1.6903.

The externally studentized residual is 3 / [sqrt(20 / 7) sqrt(0.75)], about 2.0494. The different values arise because the scale estimates differ. These stipulated fit summaries illustrate the formulas; no universal error label follows from either number. Formal inference would require the model assumptions and a screening procedure appropriate to the number of cases examined.

Illustrative case study

In a fictional company, Pine Forecasting reviews a linear demand model and flags a large externally studentized residual. The original record is correct: a one-time promotion produced unusually high demand. Analysts retain the observation, investigate whether promotional information belongs in the model, and compare fits with transparent sensitivity checks. They also inspect leverage and influence because a large residual alone does not describe the case's effect on every coefficient. The diagnostic initiates a reasoned review rather than automatic data deletion.

Common misconceptions and mistakes

Studentized residuals are not simply raw residuals divided by the overall sample standard deviation of outcomes. The relevant denominator follows the fitted regression model and includes leverage.

Internally and externally studentized residuals are not interchangeable. Software may use standardized for the internal version and studentized for the external version, so labels alone can be misleading.

A large residual is not proof of bad data, and high leverage is not the same as an unusual response conditional on predictors. Screening many cases using a single-case threshold can generate false alarms. Multiple unusual points can also affect one another's diagnostics, so one-case deletion is not a complete model-checking strategy.

Business use

Studentized residuals help businesses review regression-based forecasts, calibration models, and analytical baselines. They make discrepancies easier to compare while preserving a link to the assumed error structure.

Use them alongside domain investigation and broader diagnostic plots. If errors are heteroscedastic, correlated, or generated by a different model family, use diagnostics appropriate to that setting. A valid rare case may be important for product reliability, particularly in AI evaluation. Removing it simply to improve a model summary can hide exactly the situations the business needs to understand.

Check your understanding

Answer: A residual is 2, s is 1, and leverage is 0.75.

What is its internally studentized value?

Answer: It is 2 / sqrt(0.25) = 4. The leverage adjustment matters; dividing only by s would give 2.

Related terms

Residual
Leverage
Outlier
hat matrix
externally studentized residual
standardized residual
Cook's distance
regression diagnostics

Learn next

Study leverage and deletion diagnostics, then compare residual patterns with influence measures and methods for nonconstant error variance.

Sources

NIST, Regression Diagnostics: https://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/regrdiag.htm . R primary documentation, Regression Deletion Diagnostics: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/influence.measures.html . Internal and external definitions are distinguished explicitly. The numerical fit-summary calculation and fictional case are original teaching examples.

Previous: Studentized range statistic ( q ) Next: Stupid backoff