Why Assess Residual Normality?
Regression diagnostics ask whether the assumptions supporting a fitted model and its inference are reasonable. A normal quantile-quantile plot, usually called a normal Q-Q plot, compares the ordered regression residuals with the quantiles that would be expected from a normal distribution. The plot is useful because departures from a straight-line pattern can reveal skewness, unusually heavy tails, or an isolated extreme residual.
The plot does not establish that errors are normal, and it is not a test of whether the regression mean is linear or whether the error variance is constant. Those questions require other diagnostics. Here the focus is the shape of the residual distribution and the consequences that shape may have for inference.
Constructing the Plot
A common plotting convention assigns the \(i\)th ordered residual the probability \(p_i=(i-0.5)/n\), where \(n\) is the number of observations. Its standard normal quantile is \(z_i=\Phi^{-1}(p_i)\), where \(\Phi\) is the cumulative distribution function of \(N(0,1)\). Plot \(z_i\) horizontally and \(e_{(i)}\) vertically. Software may use a slightly different plotting-position convention, so small numerical differences between plots are normal.
The theoretical quantiles can be used directly when the reference distribution is standard normal. Alternatively, if the reference is \(N(\mu,\sigma^2)\), the corresponding theoretical residual values are \(\mu+\sigma z_i\). A reference line describes the location and scale of the residuals; it does not require the residuals to have variance exactly one. In practice, software may estimate the line, or an analyst may use a line based on the middle of the data. A line based on middle quantiles is less affected by extreme tail points than a line fitted to every point.
The normal quantiles are spaced increasingly far apart toward the ends of the plot. Consequently, a departure at an endpoint can be more informative than a similar vertical distance near the middle: the endpoint compares the most extreme observed residuals with what the normal model predicts for the tails.
What the Shape Shows
| Q-Q plot pattern | Typical interpretation | What to check |
|---|---|---|
| Points lie close to a straight line | No pronounced departure from the reference normal shape is apparent | Small irregularities are expected, especially in modest samples |
| Both ends bend away from the line in opposite vertical directions | Observed tails may be heavier than normal | Check whether the extremes are generally more pronounced than the reference predicts |
| One end departs more strongly than the other, producing a curved or asymmetric pattern | Skewness may be present | Identify which tail has the more extreme residuals |
| Most points follow the line, but one point is far from it | An isolated extreme residual may be present | Investigate the observation and other diagnostics; the plot alone does not explain its cause |
The exact appearance of a curve depends on which axis contains the observed residuals and on the direction of the departure. A safe interpretation is to describe the pattern, then relate it to the ordered residuals: a longer or more extreme upper tail indicates unusually large positive residuals, while a longer or more extreme lower tail indicates unusually large negative residuals. Avoid naming a direction of skewness solely from a memorized sketch without checking the axes.
A Justification for the Straight-Line Benchmark
The benchmark comes from the normal quantile function. If an error \(E\) has distribution \(N(\mu,\sigma^2)\), then the probability integral transform gives \(F_E(E)\sim\text{Uniform}(0,1)\). Applying the inverse normal distribution function to a probability \(p\) gives the corresponding value \(\mu+\sigma\Phi^{-1}(p)\). Thus, the population quantile at probability \(p\) is a linear function of the standard normal quantile.
The ordered sample values are not exactly equal to population quantiles at fixed probabilities, so a normal sample will not make a perfectly straight plot. The plotting positions \(p_i\) provide a convenient set of approximate probabilities for the ordered observations. Random variation produces small bends; systematic curvature or a conspicuous tail departure is more concerning than minor point-to-point noise.
Worked Examples
Worked Example: Symmetric Heavy-Tail Departures
A hypothetical regression analysis has seven ordered residuals, in model units, of \(-0.48,-0.20,-0.08,0,0.08,0.20,0.48\). Compare them with a reference normal distribution \(N(0,0.20^2)\). Use \(p_i=(i-0.5)/7\), round standard normal quantiles to three decimals, and calculate the residual departures from the reference values.
Method. Calculate the seven plotting probabilities, obtain the standard normal quantiles, multiply by the reference standard deviation \(0.20\), and subtract each reference value from its ordered residual.
The plotting probabilities are \(1/14,3/14,5/14,7/14,9/14,11/14,13/14\), or approximately \(0.071429,0.214286,0.357143,0.500000,0.642857,0.785714,0.928571\). The corresponding standard normal quantiles, rounded to three decimals, are \(-1.465,-0.792,-0.366,0,0.366,0.792,1.465\). In particular, the quantile for \(3/14\) is approximately \(-0.79164\), which rounds to \(-0.792\).
Multiplying each quantile by \(0.20\) gives the reference values:
Subtracting these in order from the observed residuals gives the departures:
Interpretation. The departures are symmetric: the two extreme residuals are farther from zero than the reference values by the same amount, and the interior residuals are close to the reference. This is consistent with heavier-than-normal tails in this small illustrative sample, not proof that the error distribution has heavy tails.
Check and common mistake avoided. The probabilities \(3/14\) and \(11/14\) are symmetric around \(0.5\), so their normal quantiles must be equal in magnitude and opposite in sign. Using \(-0.791\) and \(0.791\) would give incorrect reference values \(-0.1582\) and \(0.1582\), and incorrect departures \(-0.0418\) and \(0.0418\). The corrected rounded quantiles are \(-0.792\) and \(0.792\), giving \(-0.1584\), \(0.1584\), \(-0.0416\), and \(0.0416\).
Worked Example: Recognizing an Upper-Tail Departure
A hypothetical fitted model has ten ordered residuals \(-0.156,-0.116,-0.086,-0.056,-0.036,-0.016,0.014,0.044,0.094,0.314\). Use a reference distribution \(N(0,0.15^2)\) and \(p_i=(i-0.5)/10\). Find the departures from the reference values and describe the pattern.
Method. The plotting probabilities are \(0.05,0.15,0.25,0.35,0.45,0.55,0.65,0.75,0.85,0.95\). Their standard normal quantiles, rounded to three decimals, are \(-1.645,-1.036,-0.674,-0.385,-0.126,0.126,0.385,0.674,1.036,1.645\).
Multiply each quantile by \(0.15\). The reference values are \(-0.24675,-0.15540,-0.10110,-0.05775,-0.01890,0.01890,0.05775,0.10110,0.15540,0.24675\). Subtracting those values from the ordered residuals gives:
Interpretation. The positive extreme residual \(0.314\) is above the reference value \(0.24675\) by \(0.06725\). The departures also change direction across the ordered values, rather than staying close to zero. This is an asymmetric, curved pattern with a pronounced upper-tail departure, consistent with right skewness or an unusually long upper tail. A Q-Q plot suggests a distributional shape; it does not distinguish with certainty between a general skewness pattern and a single unusual observation.
Check and common mistake avoided. The residuals sum to zero, as expected for a least-squares regression with an intercept: the negative residuals sum to \(-0.466\), and the positive residuals sum to \(0.466\). A zero residual mean does not rule out skewness. It would be a mistake to call the data symmetric merely because their sum is zero.
Worked Example: Separating an Isolated Departure from a General Tail Pattern
A third hypothetical regression has nine ordered residuals \(-0.20,-0.16,-0.12,-0.08,-0.04,0,0.04,0.08,0.48\). Compare them with \(N(0,0.15^2)\), using \(p_i=(i-0.5)/9\). Does the plot suggest a broad heavy-tail pattern or mainly an isolated departure?
Method. The probabilities are \(0.055556,0.166667,0.277778,0.388889,0.500000,0.611111,0.722222,0.833333,0.944444\). The corresponding standard normal quantiles, rounded to three decimals, are \(-1.593,-0.967,-0.589,-0.282,0,0.282,0.589,0.967,1.593\).
Multiplying by \(0.15\) gives reference values \(-0.23895,-0.14505,-0.08835,-0.04230,0,0.04230,0.08835,0.14505,0.23895\). The residual minus reference departures are:
Interpretation. The final residual \(0.48\) is far above its reference value \(0.23895\), while the remaining points do not show matching extreme departures at both ends. The main visual feature is therefore an isolated upper-tail departure, not clear evidence of a broad, symmetric heavy-tail pattern. The other residuals also sit somewhat below the chosen reference line, which is a reminder that the conclusion depends on the line and scale used.
Check and common mistake avoided. The residuals sum to zero: the first five sum to \(-0.60\), and the last four sum to \(0.60\). The isolated high residual does not make the plot a formal outlier test, nor does it establish why that response occurred. Investigate the observation and use other diagnostics before drawing a substantive conclusion.
What Nonnormality Means for Regression Inference
A standard regression model for this diagnostic discussion can be written \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), with fixed predictor values \(x_i\). The usual normal-error assumption is that the errors are independent and each has distribution \(N(0,\sigma^2)\), with the same variance \(\sigma^2\). Under the model, normality supports exact small-sample \(t\)-based inference for regression coefficients and mean responses, when the other required model assumptions also hold.
The plotted quantities are fitted residuals, not the unobserved errors. Because the coefficients are estimated from the same observations, residuals are related to one another and their variances need not be equal. With an intercept, their sum is zero. Thus, even when the model errors are normal, a residual Q-Q plot should not be expected to look like a perfect sample of independent normal values. It is an approximate diagnostic, not a direct observation of the errors.
A marked departure matters most when it undermines the reliability of the normal-theory inference being used, particularly in a small sample or when extreme tails are important. A nearly straight plot is reassuring but does not confirm the other assumptions. Conversely, a visible departure does not by itself prove that a fitted slope is wrong or that every inferential conclusion is invalid. Consider the sample size, the severity and pattern of the departure, and the purpose of the analysis.
Common Mistakes
- Calling every bend skewness. Symmetric departures at both ends can indicate heavy tails; an asymmetric pattern is more suggestive of skewness. Describe the observed shape before naming it.
- Confusing one extreme point with a tail pattern. A single point far from the line may be an isolated departure. It is not the same evidence as a systematic separation across several points in a tail.
- Assuming residuals are independent normal observations. They are calculated using fitted values from the same data, so their dependence and unequal variances affect how literally the plot should be read.
- Treating the plot as a pass-or-fail test. Small deviations are expected from random variation. Consider whether the departure is substantial and relevant to the inference.
- Ignoring the plotting positions and rounding. Symmetric probabilities must have symmetric normal quantiles. Keep enough precision in intermediate calculations and round only when reporting the requested values.
Check Your Understanding
For each problem, use the stated plotting convention and reference distribution. Distinguish a visual suggestion from a conclusive test of the error distribution.
- A regression has five ordered residuals \(-0.70,-0.25,0,0.25,0.70\). Use \(p_i=(i-0.5)/5\) and a reference distribution \(N(0,0.50^2)\). The standard normal quantiles are approximately \(-1.282,-0.524,0,0.524,1.282\). Calculate the reference values and residual departures, then describe the tail pattern. (Answer: Reference values are \(-0.641,-0.262,0,0.262,0.641\); departures are \(-0.059,0.012,0,-0.012,0.059\), suggesting symmetric heavier tails in this small sample.)
- Six ordered residuals are \(-0.18,-0.11,-0.06,-0.02,0.04,0.33\). A Q-Q plot shows the first five points near the line and the last point clearly above it. Identify the main feature and state one conclusion the plot alone cannot establish. (Answer: An isolated upper-tail departure; the plot alone cannot establish its cause or prove that the observation is an outlier under a formal test.)
- A plot has several points at both ends bending away from a straight reference line in opposite vertical directions, while its middle points remain close to the line. What distributional feature is suggested, and why is this different from one isolated extreme point? (Answer: Heavier-than-normal tails are suggested; the systematic departures at both ends involve multiple tail points rather than a single isolated point.)
- A residual Q-Q plot is nearly linear for a regression based on a modest sample. The analyst concludes that all regression assumptions have been verified. Assess this conclusion and explain what the plot does and does not address. (Answer: The conclusion is too strong; the plot gives limited evidence about residual normality, but does not verify linearity, constant variance, independence, or the other model assumptions.)
- In a small regression sample, residuals show a pronounced asymmetric curve and a much larger upper-tail departure than lower-tail departure. Explain why this deserves attention before relying on small-sample \(t\)-based inference, while avoiding an absolute claim about validity. (Answer: The pattern suggests skewness or an unusually long upper tail, which can affect normal-theory small-sample inference; its impact depends on the severity, sample size, and other model conditions, so the plot does not by itself prove that inference is invalid.)