What Residual Standard Deviation Measures
A student or analyst usually meets residual standard deviation after fitting a regression line and asking a practical question: "How far are my predictions off in the original units?" Residual standard deviation answers that question by summarizing the typical size of the model's residuals.
A residual is the observed value minus the model's fitted value. If a delivery-time model predicts 41 minutes and the actual delivery takes 44 minutes, the residual is +3 minutes. If the next order is predicted at 37 minutes and arrives in 35 minutes, the residual is -2 minutes.
Plain-English Meaning
Use this article when your fitted model already exists and you need to evaluate error size. If you are still reviewing raw spread before modeling, start with What Is Standard Deviation? or use the Descriptive Statistics Calculator.
Formula and Degrees of Freedom
For a regression model, calculate each residual, square it, add the squared residuals, divide by the residual degrees of freedom, then take the square root.
Residual Standard Deviation
Here, SSE is the sum of squared residuals, n is the number of observations, and p is the number of fitted parameters in the model. In simple linear regression with an intercept and one slope, p = 2, so the denominator is n - 2.
| Quantity | Meaning | Example |
|---|---|---|
| Residual | Observed value minus fitted value | 44 - 41 = +3 minutes |
| SSE | Sum of squared residuals | 9 + 4 + 1 + ... across all rows |
| n | Number of observations used to fit the model | 8 deliveries |
| p | Number of estimated model parameters | 2 for intercept plus slope |
| n - p | Residual degrees of freedom | 8 - 2 = 6 |
Why the Denominator Is Not n
Worked Example with Residuals
Here is the kind of worksheet check we use when teaching regression diagnostics: take the model's fitted values, compute residuals row by row, and verify that the reported residual standard deviation matches the arithmetic. The dataset below uses delivery distance to predict delivery time in minutes.
| Delivery | Actual minutes | Fitted minutes | Residual | Squared residual |
|---|---|---|---|---|
| 1 | 32 | 31 | +1 | 1 |
| 2 | 35 | 37 | -2 | 4 |
| 3 | 39 | 38 | +1 | 1 |
| 4 | 41 | 43 | -2 | 4 |
| 5 | 48 | 46 | +2 | 4 |
| 6 | 50 | 53 | -3 | 9 |
| 7 | 58 | 55 | +3 | 9 |
| 8 | 61 | 62 | -1 | 1 |
Add the squared residuals
Find the residual degrees of freedom
Divide and take the square root
Decision from the Example
To check the spread of the residual values themselves, paste `1, -2, 1, -2, 2, -3, 3, -1` into the Standard Deviation Calculator or compare the residual squares with the Variance Calculator. Remember that regression output uses n - p, not the plain sample denominator n - 1.
How to Interpret the Result
Residual standard deviation is expressed in the response variable's units. For delivery time, the result is minutes. For test scores, it is points. For house-price models, it is dollars. That unit match makes it easier to discuss model error with people who do not read regression tables.
Small residual SD
Large residual SD
Do not interpret residual standard deviation alone. Compare it with the business tolerance, measurement resolution, and the natural spread of the response variable. A 2.35-minute error may be excellent for city delivery routing and unusable for a timed lab process.
For approximately normal residuals, a quick rule is that many prediction errors should fall within about 2 residual standard deviations of zero. That is a diagnostic shortcut, not a guarantee. For standardized error interpretation, see Z-Score Explained and Detecting Outliers with Standard Deviation.
Use and Reporting Checklist
- Use:Use residual standard deviation after fitting a regression model when you need typical error in the response variable's units.
- Formula:Use sqrt(SSE / (n - p)), where p counts all fitted parameters, including the intercept when present.
- Report:Report the value with units, sample size, model form, and residual degrees of freedom.
- Compare:Compare the error size with a real decision threshold, not with zero.
- Diagnose:Check residual plots before trusting a single summary number.
| Question | Good sign | Action if not |
|---|---|---|
| Are residuals centered near zero? | No clear positive or negative bias | Review intercept, missing predictors, or data coding |
| Is the spread roughly constant? | Residual width looks similar across fitted values | Consider transformation, weighted regression, or separate groups |
| Are there extreme residuals? | No point dominates SSE | Investigate data errors, special cases, or robust methods |
| Is the error acceptable? | Residual SD is below the decision tolerance | Improve predictors or reconsider the model's use |
Common Mistakes
- Using n - 1 instead of n - p for regression residuals.
- Calling residual standard deviation the same thing as standard error. They are related, but standard error usually describes uncertainty in an estimate, while residual standard deviation describes model error around fitted values.
- Comparing residual SD across models with different response units or transformed response variables without converting the interpretation.
- Ignoring residual plots. A low summary value can still hide curvature, nonconstant variance, or one high-leverage observation.
- Treating the value as a percentage when the response variable is measured in raw units.
Weakest-Section Rewrite Check
Further Reading
Sources
References and further authoritative reading used in preparing this article.
- NIST/SEMATECH e-Handbook of Statistical Methods: Linear Least Squares Regression — NIST
- STAT 501: What Is the Common Error Variance? — Penn State Eberly College of Science