Seeing a residual plot with “no obvious pattern” is not the same as confirming homoscedasticity, and a residual plot that looks scattered at a glance can still hide a funnel, a curve, or a single influential point doing all the work. The plot itself doesn’t tell you anything — how you read it does. This guide walks through the four patterns you’ll actually encounter, what each one means for your regression, and what a researcher writes in a methods or results section once they’ve made the call.
What a Residual Plot Actually Shows
A residual is the vertical distance between an observed value and the value your model predicted: ei = yi − ŷi. A residual plot puts these residuals on the y-axis against the fitted (predicted) values ŷi, or sometimes against a specific predictor, on the x-axis. If ordinary least squares (OLS) regression is doing its job, the residuals should scatter randomly around the horizontal zero line with no discernible pattern and roughly constant vertical spread across the whole range of fitted values.
That constant spread is homoscedasticity — “same scatter.” Its opposite, heteroscedasticity, is when the spread of residuals changes systematically as the fitted values change. It’s easy to conflate homoscedasticity with “the residuals are small” or “the model has a high R².” Neither is true: a model can have small residuals and still be heteroscedastic, and R² says nothing about whether the error variance is constant. Homoscedasticity is specifically about the shape of the spread, not its size.
The Four Patterns You Need to Recognize
1. Well-Behaved: Random Scatter, Constant Spread
Points fall above and below the zero line in a band of roughly even width across the full range of fitted values, with no funnel, no curve, and no single point sitting far outside the rest. This is what you want to see. It supports the homoscedasticity assumption and, combined with a roughly straight-line pattern on a Q–Q plot of the residuals, supports the normality assumption used for OLS inference. Nothing to fix here — report it as confirmed, if you report diagnostics at all.
2. Funnel (or Megaphone) Shape: Heteroscedasticity
The band of residuals narrows at one end and widens at the other — classically, tight near low fitted values and fanning out at high fitted values. This is the textbook signature of heteroscedasticity: the error variance is not constant, it grows (or shrinks) with the level of the outcome. It shows up constantly with count data, income/spending data, reaction times, and anything else where larger predicted values naturally come with larger absolute variability.
Heteroscedasticity does not bias your coefficient estimates — OLS is still unbiased. What it breaks is the standard errors, which means your p-values and confidence intervals are no longer trustworthy even though the point estimates themselves are fine.
3. Curved or U-Shaped Pattern: Model Misspecification
The residuals trace a visible curve — a U, an inverted U, or a systematic bow — rather than scattering randomly. This isn’t primarily a variance problem; it’s a sign that the linear functional form is wrong for this relationship. A straight line is being asked to fit a curve, and the leftover curvature shows up in the residuals because the model has nowhere else to put it. Common causes: a missing quadratic or polynomial term, a predictor that needs a log or square-root transform, or an omitted variable that’s correlated with both the predictor and the outcome (see our guide on confounding variables).
4. Outlier- or Leverage-Driven Pattern: A Few Points Doing the Work
Most of the plot looks fine, but one or a handful of points sit noticeably far from the zero line, or far out along the x-axis with unusual leverage. A single high-leverage point can distort a fitted line enough to make the rest of the plot look worse than it is, or mask a real pattern in the bulk of the data. The conventional (not formal) screening rule is to flag standardized residuals beyond ±3 (some sources use ±2 as a looser flag) and to check Cook’s distance, where values above roughly 4/n are commonly treated as worth a closer look. These are flagging conventions for further investigation, not automatic exclusion rules — the next step is understanding why the point is unusual, not deleting it.
Worked Example: Reading a Real-Looking Residual Plot
Illustrative example, not drawn from any specific published study. Suppose a simple regression models reaction time (ms) as a function of drug dosage (mg) across six dosage levels:
| Dosage (mg) | Fitted value (ms) | Residual (ms) |
|---|---|---|
| 5 | 210 | −3 |
| 10 | 225 | +5 |
| 20 | 250 | −8 |
| 40 | 300 | +15 |
| 80 | 400 | −28 |
| 160 | 600 | +52 |
Plotted against fitted values, the residuals widen from single digits at the lowest fitted values to roughly ±50 ms at the highest — a funnel, not random noise. Here is the interpretation sentence a researcher would actually write:
“A plot of residuals against fitted values showed a funnel-shaped pattern of increasing spread at higher fitted values, indicating heteroscedasticity. The Breusch–Pagan test supported this reading. We therefore report heteroscedasticity-consistent (HC3) standard errors rather than the conventional OLS standard errors.”
Notice what that sentence does not do: it doesn’t discard the model, it doesn’t claim the coefficients are wrong, and it doesn’t imply the relationship between dosage and reaction time is invalid. It just correctly flags that the uncertainty around that relationship needs a different standard-error estimator.
Formal Tests: Confirming What Your Eye Sees
A visual read is a good first pass but is inherently subjective, especially with small samples where “some” funneling is hard to distinguish from noise. Two standard formal tests back it up:
- Breusch–Pagan test — regresses the squared residuals on the predictors and tests whether they explain a significant share of variance; a significant result indicates heteroscedasticity.
- White test — a more general version that also allows for nonlinear relationships between the predictors and the residual variance, at the cost of being less powerful in small samples.
Use the plot to build intuition about where the variance is changing and the test to confirm whether it’s changing enough to matter for your sample size.
Fixing Heteroscedasticity
Once a funnel pattern is confirmed, the common remedies are, roughly in order of how often they’re used:
- Transform the outcome variable — a log or square-root transform often stabilizes variance when the spread grows with the level of the outcome, which is the most common shape in practice.
- Heteroscedasticity-consistent (robust) standard errors — White/Huber-type estimators (commonly HC0 through HC3) leave the coefficient estimates and functional form untouched and correct only the standard errors. This is the most widely used fix in applied social-science and biomedical regression because it doesn’t require you to change the model.
- Weighted least squares (WLS) — down-weights observations with higher variance if you can model how the variance changes across observations.
- Bootstrap standard errors — a resampling-based alternative when the variance structure is irregular or hard to specify parametrically.
When This Assumption Doesn’t Apply
Homoscedasticity is one of the classical assumptions of ordinary least squares regression, alongside linearity, independence of errors, and (for inference) approximately normal residuals — see our guide on checking the normality assumption. It is not a universal requirement across all regression methods:
- Generalized linear models (GLMs) such as Poisson or logistic regression explicitly model a mean–variance relationship rather than assuming constant variance, so a standard OLS-style residual plot is the wrong diagnostic; deviance or Pearson residuals are used instead.
- Quantile regression doesn’t require homoscedastic errors at all, since it estimates conditional quantiles rather than a conditional mean.
- Robust or WLS regression is specifically built to tolerate heteroscedasticity rather than require its absence.
Check what your specific model actually assumes before treating a funnel shape as automatically disqualifying — the diagnostic that matters is the one that matches your estimator.
Frequently Asked Questions
What does a “good” residual plot look like?
A horizontal band of points scattered randomly above and below zero, with roughly constant width across the full range of fitted values and no visible curve, funnel, or isolated outlier.
Does heteroscedasticity mean my regression is invalid?
No. It means your standard errors, p-values, and confidence intervals are unreliable under standard OLS assumptions — the coefficient estimates themselves remain unbiased. Correcting with robust standard errors or a transform is usually sufficient; it doesn’t require abandoning the model.
What’s the difference between a residual plot and a Q–Q plot?
A residual-vs-fitted plot diagnoses homoscedasticity and functional-form misspecification. A Q–Q plot compares the distribution of residuals to a theoretical normal distribution and diagnoses the normality assumption. They answer different questions and are typically produced together.
Can I just look at R² instead of the residual plot?
No — R² measures how much variance in the outcome the model explains overall. It doesn’t reveal whether that unexplained variance is spread evenly across the range of the data, which is exactly what a residual plot is for.
Is a curved pattern always fixed by adding a squared term?
Not always, but it’s the most common first attempt. A log or other transform of the predictor, or reconsidering whether an important variable is missing from the model (see confounding variables), are the other standard responses. Which one is correct depends on the shape of the curve and the substantive relationship being modeled.
Related Reading
Once you’ve confirmed your regression’s variance and functional-form assumptions, the diagnostics above feed directly into how you report the model’s results: see our guides on interpreting a p-value, how to report p-values, confidence intervals, and correlation coefficients. If your residuals also look skewed rather than symmetric, our guide on left- vs. right-skewed distributions covers that pattern specifically.
