How to Calculate Variability in Regression
Enter observed values and predicted values to compute total variability, explained variability, unexplained variability, R-squared, residual variance, and RMSE. This tool is designed for students, analysts, and researchers who want a clear breakdown of regression fit.
Interactive Calculator
For simple linear regression, use 1. The intercept is handled automatically.
Choose output precision for the metrics and chart labels.
Enter numbers separated by commas, spaces, or line breaks.
The list must contain the same number of values as the observed series.
Results
Click “Calculate Variability” to generate the decomposition of variation in regression.
Expert Guide: How to Calculate Variability in Regression
Variability is one of the central ideas in regression analysis. When analysts ask whether a regression model is “good,” they are usually asking how much of the outcome’s variation can be explained by the predictors and how much remains unexplained. Learning how to calculate variability in regression gives you a practical way to judge model quality, compare competing models, and communicate findings with statistical precision.
At a high level, regression breaks the variation in a response variable into two major pieces. The first piece is the part the model explains. The second piece is the part left over after fitting the model, which is often called error or residual variation. This decomposition is the foundation of R-squared, analysis of variance for regression, residual standard error, and many model comparison methods.
Why variability matters in regression
If your outcome variable barely changes from one observation to the next, there is not much variation to explain. But in most real data, outcomes vary due to measurable factors, random noise, omitted variables, and natural heterogeneity. Regression helps you estimate the systematic component of that variation. By computing variability correctly, you can answer several important questions:
- How spread out are the observed outcome values around their mean?
- How much of that spread is explained by the regression line or regression surface?
- How much variation remains in the residuals?
- Is the fit strong enough to support prediction or explanation?
- Does adding predictors meaningfully reduce unexplained variability?
In standard least squares regression, the total variation in the response variable can be decomposed into explained variation and unexplained variation. That is why terms such as SST, SSR, and SSE appear so often in statistics courses and software output.
The three core sums of squares
To calculate variability in regression, you usually start with three sums of squares:
- Total Sum of Squares (SST): measures the total variability of the observed outcomes around their mean.
- Error Sum of Squares (SSE): measures the unexplained variability, or the residual variation around the fitted values.
- Regression Sum of Squares (SSR): measures the explained variability due to the regression model.
Key identity: In ordinary least squares regression with an intercept, the decomposition is SST = SSR + SSE. This identity is one of the most important relationships in applied regression.
Here is what each quantity means mathematically:
- SST = Σ(Yi – Ȳ)²
- SSE = Σ(Yi – Ŷi)²
- SSR = Σ(Ŷi – Ȳ)²
Where:
- Yi is the observed value of the response variable
- Ŷi is the predicted value from the regression model
- Ȳ is the sample mean of the observed response values
Step-by-step process to calculate variability in regression
If you are working from raw observed and predicted values, the process is straightforward:
- Compute the sample mean of the observed response values, Ȳ.
- For each observation, subtract the mean from the observed value and square the result. Add those squared differences to get SST.
- For each observation, subtract the predicted value from the observed value and square the result. Add those squared residuals to get SSE.
- Compute SSR either directly with Σ(Ŷi – Ȳ)² or indirectly as SST – SSE.
- Calculate R² = SSR / SST or equivalently R² = 1 – SSE / SST.
- If needed, compute residual variance as SSE / (n – p – 1), where n is sample size and p is the number of predictors.
- Take the square root of residual variance to get residual standard error.
This is exactly what the calculator above does. Once you paste observed values and fitted values, it computes the decomposition and visualizes how total variability splits into explained and unexplained parts.
Worked example with actual numbers
Suppose you have eight observed outcomes and a set of fitted values from a regression model:
- Observed Y: 3, 5, 4, 7, 9, 10, 12, 11
- Predicted Ŷ: 2.8, 4.7, 5.1, 6.8, 8.6, 10.3, 11.5, 11.2
The mean of the observed values is 7.625. If you compute the squared deviations from the mean and add them, the total sum of squares is 71.875. If you then compute the squared residuals and add them, the error sum of squares is 1.360. That means the explained variation is:
SSR = SST – SSE = 71.875 – 1.360 = 70.515
Now compute R-squared:
R² = 70.515 / 71.875 ≈ 0.981
This tells you the regression explains about 98.1% of the variability in the response variable. That is a very strong fit, at least in terms of variance explained. If the model has one predictor, the residual variance is 1.360 / (8 – 1 – 1) = 0.227, and the residual standard error is about 0.476.
| Metric | Formula | Value in worked example | Interpretation |
|---|---|---|---|
| Total variability | SST = Σ(Yi – Ȳ)² | 71.875 | Total spread of observed outcomes around the mean |
| Unexplained variability | SSE = Σ(Yi – Ŷi)² | 1.360 | Residual variation left after fitting the regression |
| Explained variability | SSR = SST – SSE | 70.515 | Variation captured by the regression model |
| Coefficient of determination | R² = SSR / SST | 0.981 | About 98.1% of variance explained |
How R-squared connects to variability
R-squared is a normalized measure of explained variability. Because it divides the explained variation by the total variation, it produces an easy-to-interpret proportion. If R² = 0.60, then 60% of the observed variability in the response is explained by the model and 40% remains unexplained. This makes R-squared especially useful for comparing models fit to the same response variable on the same dataset.
However, you should not use R-squared by itself as a complete measure of model quality. A high R-squared does not guarantee that the model is appropriate, causal, unbiased, or stable out of sample. It also does not prove that the residuals satisfy the assumptions of linear regression. That is why experienced analysts pair variability metrics with residual plots, subject-matter knowledge, and inferential checks.
Residual variance, MSE, and residual standard error
After you calculate SSE, you can transform it into a per-degree-of-freedom measure of unexplained variance. In multiple regression with p predictors and an intercept, the residual degrees of freedom are n – p – 1. The estimated residual variance is:
s² = SSE / (n – p – 1)
The square root of this quantity is the residual standard error, sometimes called the standard error of the regression. It is especially valuable because it is reported in the same units as the response variable. If your outcome is measured in dollars, the residual standard error is also in dollars.
In machine learning contexts, you will often see RMSE, the root mean squared error. Depending on the convention, RMSE may use SSE / n or SSE / (n – p – 1) inside the square root. In classical regression inference, the latter is more closely tied to estimated error variance. In predictive model reporting, the former is common. The calculator above reports both residual standard error and RMSE to make the distinction clear.
Comparison of models using variability metrics
One of the most practical uses of regression variability is model comparison. Imagine you are evaluating three models on the same dataset with the same response variable. The table below shows how explained and unexplained variability may differ.
| Model | SST | SSE | SSR | R² | Residual Std. Error |
|---|---|---|---|---|---|
| Simple linear model | 500.0 | 220.0 | 280.0 | 0.560 | 4.83 |
| Two-predictor model | 500.0 | 145.0 | 355.0 | 0.710 | 3.91 |
| Four-predictor model | 500.0 | 110.0 | 390.0 | 0.780 | 3.45 |
This table shows a common pattern: as predictors are added, SSE tends to fall and SSR tends to rise, so R-squared increases. But that does not automatically mean the largest model is best. You should also consider adjusted R-squared, prediction error on holdout data, collinearity, interpretability, and whether the extra variables are theoretically justified.
Common mistakes when calculating variability in regression
- Using the wrong mean: SST must be based on the mean of the observed response values, not the mean of predictions.
- Mixing up SSE and SSR: SSE is residual variation, while SSR is explained variation.
- Forgetting the intercept condition: The clean identity SST = SSR + SSE depends on fitting a model with an intercept in ordinary least squares.
- Interpreting R-squared too broadly: A large R-squared does not prove causation or guarantee good forecasts outside the sample.
- Ignoring scale: SSE depends on the units of the response variable and the sample size, so comparisons across different datasets can be misleading.
- Confusing variance with standard deviation: Residual variance is in squared units, while residual standard error is in the original units.
How this relates to ANOVA in regression
Analysis of variance for regression is built directly on the variability decomposition. The ANOVA table usually contains sums of squares, degrees of freedom, mean squares, and an F-statistic. The regression mean square is MSR = SSR / p. The error mean square is MSE = SSE / (n – p – 1). Then the model F-statistic is F = MSR / MSE. This provides a formal test of whether the predictors explain a statistically significant amount of variability in the response.
In practical terms, the ANOVA view and the R-squared view are two ways of looking at the same underlying partition of variation. One is inferential, the other is descriptive. Both start with correctly calculating variability in regression.
How to interpret high and low variability explained
A high percentage of explained variability can indicate that your predictors are closely associated with the outcome, but the context matters. In physical sciences and engineered systems, very high R-squared values are not unusual. In economics, healthcare, education, or social science, outcomes often depend on many unobserved factors, so even moderate R-squared values can still be useful. There is no universal cutoff that defines a “good” model.
Likewise, a low R-squared is not always bad. If your goal is causal inference or estimating a specific coefficient rather than maximizing prediction accuracy, you can still have a meaningful model with modest explained variance. The critical point is to interpret variability metrics in conjunction with the study design, residual diagnostics, and the consequences of prediction error in the real world.
Recommended authoritative references
If you want to go deeper into regression variability, sums of squares, and residual analysis, these resources are strong starting points:
- NIST/SEMATECH e-Handbook of Statistical Methods
- Penn State STAT 462: Applied Regression Analysis
- UCLA Institute for Digital Research and Education: Statistical Resources
Final takeaway
To calculate variability in regression, you are really doing a decomposition of the response variable’s total spread. The total variation, SST, is partitioned into explained variation, SSR, and unexplained variation, SSE. From those quantities, you can compute R-squared, residual variance, residual standard error, and additional ANOVA-based statistics. Once you understand these relationships, regression output becomes much easier to read and much more useful for decision-making.
The calculator on this page gives you a direct way to perform that decomposition from observed and predicted values. Use it to verify homework steps, audit regression output, compare competing models, or quickly illustrate how explained and unexplained variation change when model quality improves.