Standard error in regression measures how far your model’s predictions typically fall from the actual observed values. It’s expressed in the same units as the thing you’re trying to predict, which makes it one of the most intuitive ways to judge how well a regression model performs. If you’re predicting home prices and your standard error is $15,000, your predictions are off by roughly $15,000 on average.
The term “standard error” actually refers to two related but distinct things in regression output, and confusing them is one of the most common stumbling blocks. One is the standard error of the overall model (how accurate the predictions are). The other is the standard error of each coefficient (how precisely you’ve estimated the effect of each variable). Both matter, and they show up in different parts of your output.
Standard Error of the Regression Model
The standard error of the regression, sometimes called the residual standard error or root mean square error, tells you the typical size of your model’s prediction errors. Technically, it’s the standard deviation of the residuals: the differences between each actual value and the value the model predicted. A smaller number means the model’s predictions cluster tightly around the real data. A larger number means wider misses.
The calculation works like this: take each prediction error (actual minus predicted), square them all, add them up, divide by the degrees of freedom, then take the square root. Degrees of freedom here equals the number of data points minus the number of coefficients in your model (including the intercept). So in a simple regression with one predictor and 50 observations, you’d divide by 48. In a multiple regression with three predictors and 50 observations, you’d divide by 46. Dividing by degrees of freedom instead of the raw sample size corrects for a slight bias that would otherwise make the estimate too small.
What makes this number so useful is that it’s measured in the same units as your outcome variable. If you’re modeling test scores, the standard error is in points. If you’re modeling weight, it’s in pounds or kilograms. This means you can immediately judge whether the model is precise enough for your purposes. A standard error of 3 points on an exam where scores range from 0 to 100 is excellent. A standard error of 30 points is nearly useless.
Standard Error of the Coefficients
Each coefficient in your regression output also has its own standard error, and this answers a different question: how much would this coefficient change if you collected a new sample and ran the regression again? A small standard error means the estimate is stable and precise. A large one means the estimate is wobbly, and you shouldn’t trust it too much.
The standard error of a coefficient is measured in the same units as the coefficient itself. For a slope coefficient, that’s “units of the outcome per unit of the predictor.” If you’re predicting salary based on years of experience, and the slope is $2,500 per year with a standard error of $400, you can be fairly confident the true effect is somewhere near $2,500. If the standard error were $2,200 instead, that same $2,500 estimate would be far less convincing because the uncertainty nearly swallows the estimate.
This is where significance testing enters the picture. Statistical software divides each coefficient by its standard error to produce a t-statistic. A coefficient of $2,500 divided by a standard error of $400 gives a t-value of 6.25, which is large enough to be highly significant. The same coefficient divided by $2,200 gives a t-value of about 1.14, which would not reach significance. The standard error is doing the heavy lifting in determining whether a predictor genuinely contributes to the model or just appears to by chance.
How It Differs From R-Squared
R-squared and standard error both describe model fit, but they do it in fundamentally different ways. R-squared is a proportion: it tells you what percentage of the variation in the outcome your model explains. It ranges from 0 to 1 and has no units. Standard error is a distance: it tells you how far off your predictions typically are, in the actual units of the data.
The standard error is often more informative for practical decisions. An R-squared of 0.85 sounds impressive, but it doesn’t tell you whether your predictions are off by $500 or $50,000. The standard error gives you that answer directly. If you need predictions accurate to within a certain threshold, the standard error is the number to check. R-squared tells you the model is working in relative terms. Standard error tells you whether it’s working well enough in absolute terms.
Building Prediction Intervals
One of the most practical uses of the standard error is constructing prediction intervals, which give you a range that a new observation will likely fall within. The basic structure is straightforward: take your predicted value, then add and subtract a t-multiplier times the standard error of the prediction.
For a 95% prediction interval, this roughly means your predicted value plus or minus about two standard errors (the exact multiplier depends on sample size and comes from the t-distribution). So if your model predicts a home will sell for $300,000 and the standard error of the prediction is $12,000, the 95% interval runs from approximately $276,000 to $324,000. The prediction interval is always wider than a confidence interval for the average response, because it accounts for both the uncertainty in estimating the line and the natural scatter of individual data points around that line.
The standard error of a prediction also grows as you move further from the center of your data. Predictions near the average value of your predictor are more precise. Predictions at the extremes carry more uncertainty. This is why extrapolating far beyond your data range produces prediction intervals so wide they become meaningless.
What Makes Standard Errors Unreliable
Standard errors are only trustworthy when the assumptions behind them hold. The most important assumption is homoscedasticity, which means the spread of the residuals stays roughly constant across all levels of the predictor. If prediction errors are small for low values and large for high values (or vice versa), the standard errors will be distorted. They may look reassuringly small when the model is actually performing poorly in certain ranges.
Non-normality of the residuals is less critical for large samples, but in small datasets it can throw off the t-statistics and p-values that depend on standard errors. Outliers pose a similar threat: a few extreme observations can inflate the standard error and mask real effects, or in some cases deflate it in misleading ways.
When homoscedasticity is violated, the fix is typically to use robust standard errors (sometimes called heteroscedasticity-consistent standard errors), which adjust for uneven variance without changing the coefficient estimates themselves. Most statistical software can calculate these with a simple option change. The coefficients stay the same; only the standard errors, t-statistics, and p-values are recalculated.
Reading Standard Error in Software Output
In most regression output, you’ll see the standard error in two places. First, near the top or bottom of the summary, you’ll find the residual standard error (or standard error of the estimate) for the overall model. This is the single number summarizing prediction accuracy. Second, in the coefficient table, each row will have a “Std. Error” column showing the standard error for that specific coefficient.
To quickly assess a coefficient, look at the ratio of the estimate to its standard error. If the coefficient is at least twice as large as its standard error, it’s typically significant at the 0.05 level (for reasonably sized samples). This is a useful mental shortcut when scanning output. You can also use the standard error to build a rough 95% confidence interval for the coefficient: the estimate plus or minus about two standard errors gives you the plausible range for the true value.
A common mistake is focusing only on whether coefficients are significant and ignoring the overall standard error. A model can have all significant predictors and still produce predictions that are too imprecise to be useful. The residual standard error keeps you honest about what the model can actually deliver.