Why Examine Residual Quantiles?
A residual histogram summarizes the overall shape of a set of residuals, but it can conceal how the center and tails differ from a reference distribution. A quantile-quantile plot, or Q-Q plot, makes that comparison directly: it pairs ordered observed residuals with the quantiles expected from a specified reference distribution. In population pharmacokinetics, a normal-reference Q-Q plot can help assess whether residuals that are expected to be approximately standard normal show symmetric tail excess, asymmetry, or only a few extreme values.
The plot is a diagnostic, not a pass-or-fail test. Its shape depends on sample size, the construction of the residuals, and whether observations are dependent. It also describes the marginal distribution of the plotted residuals: it cannot by itself reveal a time trend, a concentration-dependent pattern, or within-subject serial dependence. Those features require diagnostics that retain the relevant ordering or covariate information.
Constructing and Reading the Plot
Let \(r_1,\ldots,r_n\) be the residuals being examined and write their ordered values as \(r_{(1)}\leq \cdots \leq r_{(n)}\). For each rank \(i\), choose a plotting probability \(p_i\) and calculate the standard normal quantile \(z_i=\Phi^{-1}(p_i)\), where \(\Phi\) is the cumulative distribution function of \(N(0,1)\). Plot \((z_i,r_{(i)})\). A common simple choice is \(p_i=(i-0.5)/n\). Other plotting-position conventions are also used; the choice has little practical effect on the broad shape, but can affect the most extreme plotted points in small samples.
The standard normal reference has mean zero and variance one. If the residuals are meant to be standard normal, the reference line is \(r=z\). A line fitted to the points, or a line through the sample and theoretical quartiles, can instead help assess distributional shape after allowing for a shift or scale difference. That adjustment changes the question: points following a fitted line do not establish that residuals are centered at zero or have unit spread. For a standard-normal diagnostic, inspect both shape and the location and scale relative to \(y=x\).
Why the Plot Has This Reference Pattern
The plotting positions approximate the probabilities associated with the ordered values. If \(R_1,\ldots,R_n\) are independent observations from a continuous distribution with cumulative distribution function \(F\), then \(F(R_j)\) are independent uniform random variables on \((0,1)\). Their \(i\)th ordered value has expected value \(i/(n+1)\). This gives a justification for assigning a probability near \(i/(n+1)\) to the \(i\)th ordered observation.
Under the standard normal model, applying the inverse normal distribution to those probabilities gives the expected reference quantiles. If instead the observations follow a normal distribution with mean \(\mu\) and standard deviation \(\sigma\), their population quantiles satisfy \(F^{-1}(p)=\mu+\sigma\Phi^{-1}(p)\). Thus normal data have a linear quantile relationship: a location shift changes the intercept, and a scale change changes the slope. For standard normal residuals, the target intercept is zero and the target slope is one.
This reasoning describes the reference under independent draws from a common distribution. Pharmacokinetic residuals can be grouped within subjects, can reflect fitted model parameters, and can have different precision or construction across observations. Therefore, the simple order-statistic argument explains the plot but does not guarantee that ordinary independent-sample variability applies to every residual set. If dependence or estimation effects are material, use the plot as a descriptive diagnostic and consider model-appropriate simulation or other diagnostics to judge how much curvature is unusual.
Worked Examples: Three Different Departure Patterns
Worked Example: Symmetric Heavy Tails in a Small Sample
Suppose nine hypothetical standardized residuals, already ordered, are \(-3.0,-1.8,-1.2,-0.7,0,0.7,1.2,1.8,3.0\). Use \(p_i=(i-0.5)/9\). The first plotting probability is \(0.5/9=0.0556\), and its standard normal quantile is approximately \(-1.59\). By symmetry, the ninth theoretical quantile is \(1.59\); the middle quantile is zero. The paired extremes are therefore approximately \((-1.59,-3.0)\) and \((1.59,3.0)\).
At the second rank, \(p_2=1.5/9=0.1667\), giving \(z_2\approx-0.97\), while the observed residual is \(-1.8\). Its symmetric partner is \((0.97,1.8)\). In contrast, the middle points are close to the reference: for example, \(z_4\approx-0.28\) is paired with \(-0.7\), and the fifth pair is \((0,0)\). The two tails extend farther from zero than their corresponding normal quantiles, on both sides.
This small sample is compatible with symmetric heavy tails, but it cannot establish them. With only nine points, the extreme order statistics naturally vary substantially, and a single observation has a large effect on the apparent shape. The appropriate conclusion is that both ends merit attention, not that the residual distribution has been proven heavy-tailed.
Worked Example: An Asymmetric Upper Tail
Consider 15 hypothetical ordered residuals: \(-1.7,-1.2,-0.9,-0.7,-0.5,-0.3,-0.1,0,0.1,0.2,0.4,0.7,1.0,1.5,2.2\). Using \(p_i=i/(n+1)=i/16\), the middle position \(i=8\) has \(p_8=0.5\), theoretical quantile zero, and observed residual zero. At \(i=4\), \(p_4=0.25\), so \(z_4\approx-0.674\), close to the observed \(-0.7\). At \(i=12\), \(p_{12}=0.75\), so \(z_{12}\approx0.674\), close to the observed \(0.7\).
The upper ranks depart progressively. At \(i=14\), \(p_{14}=0.875\) and \(z_{14}\approx1.15\), compared with an observed residual of \(1.5\). At \(i=15\), \(p_{15}=0.9375\) and \(z_{15}\approx1.53\), compared with \(2.2\). The lower tail does not show a comparable outward departure. The resulting Q-Q pattern would bend upward in the upper tail rather than departing symmetrically at both ends.
That one-sided, progressive pattern is consistent with positive skewness or an unusually extended upper tail. It is not enough to identify the cause, and a small sample leaves considerable uncertainty. A residual-versus-prediction or residual-versus-time diagnostic may help determine whether the apparent asymmetry is concentrated in a particular part of the data.
Worked Example: Tail Excess in a Large Sample
Suppose a hypothetical set contains \(n=1000\) residuals. Its empirical 1st, 50th, and 99th percentiles are \(-3.0\), \(0.02\), and \(3.1\), respectively. The corresponding standard normal quantiles are approximately \(-2.326\), \(0\), and \(2.326\). The two extreme percentiles extend beyond the normal reference in both directions, while the median is close to zero. This pattern is consistent with symmetric tail excess rather than a simple location shift.
A second check counts observations beyond three standard deviations. Under a standard normal reference, \(P(|Z|>3)\approx0.00270\). The expected count among 1000 independent observations is therefore \(1000(0.00270)=2.70\). Suppose 14 residuals exceed \(3\) or are below \(-3\). The observed proportion is \(14/1000=0.0140\), or 1.40%, compared with the reference probability of about 0.27%. That is a substantial descriptive excess and agrees with the percentile comparison.
The count is not, by itself, a formal test: its sampling distribution and interpretation depend on independence and on how the residuals were estimated. If several extremes come from one subject, for example, treating all 1000 residuals as independent may exaggerate the amount of independent evidence. The large sample makes the overall tail pattern easier to see, but dependence and model construction still matter.
Worked Example: One Isolated Extreme Point
In a hypothetical sample of 40 residuals, suppose the ordered values mostly track the standard normal reference, but the largest residual is \(4.1\). With \(p_i=(i-0.5)/40\), the largest plotting probability is \(39.5/40=0.9875\), whose standard normal quantile is approximately \(2.24\). The top plotted pair is thus approximately \((2.24,4.1)\), far above the line \(y=x\).
If the next-largest residual is \(2.0\), and the rest of the points follow the reference reasonably well, the plot shows one isolated upper-tail departure rather than a broad bend across many upper ranks. This differs from the progressive upper-tail pattern in the preceding example. It does not show that the observation is erroneous or that it should be removed. It signals a case to investigate in context, while retaining it in the analysis unless a defensible data or model reason supports a change.
Distinguishing Tail Shapes and Using Other Diagnostics
| Q-Q pattern | Possible interpretation | What to check |
|---|---|---|
| Both tails depart outward in a roughly balanced way | Symmetric tail excess relative to the normal reference | Whether departures span several ranks and whether residuals are dependent |
| One tail bends away more than the other | Skewness or one-sided tail excess | Whether the pattern is progressive and related to time, prediction, or a subgroup |
| Most points track the line, with one point far away | An isolated extreme observation | Data integrity and the observation's context; do not infer a cause from the plot alone |
| Points are roughly straight but shifted or have a different slope | Location or scale mismatch, or residuals not standardized as expected | Residual definition, centering, scaling, and the target reference |
A histogram is useful for seeing broad shape and multiple modes, but its appearance changes with bin width and small samples can look very different under different bins. A Q-Q plot orders the observations and makes tail departures easier to locate, though its extreme points are also noisy in small samples. Residual-versus-prediction and residual-versus-time plots answer different questions: they can show structure in the residuals that a Q-Q plot erases by sorting values. Formal goodness-of-fit tests provide a test statistic and a decision rule under stated assumptions, but a test may be sensitive to sample size and does not replace examination of the pattern or its context.
For residuals expected to be approximately standard normal, compare points with the zero-intercept, unit-slope reference as well as assessing general shape. If a fitted line is used to emphasize shape, report or inspect its location and scale separately. For small samples, avoid treating a few tail points as decisive; for large samples, distinguish a practically meaningful tail pattern from minor departures that become visible because many observations are available.
Common Pitfalls and Limits
- Calling one extreme point “heavy tails.” Heavy-tail evidence is stronger when several ordered observations in both tails depart systematically. One point may be an isolated extreme, and needs contextual investigation.
- Calling a one-sided bend symmetric tail excess. Compare the upper and lower tails separately. Unequal departures can indicate skewness or one-sided tail behavior.
- Assuming a straight fitted line proves standard normality. A shifted or rescaled normal distribution can still look linear. Check centering and scale against the intended reference.
- Treating residuals as independent without checking their structure. Residuals from the same subject may be related. A marginal Q-Q plot does not assess that dependence.
- Using the plot to diagnose every model problem. Sorting discards time, prediction, and covariate order. Pair the Q-Q plot with diagnostics that preserve those relationships.
A Q-Q plot cannot identify whether an apparent departure arose from the structural model, the residual error model, data issues, or dependence. It also cannot justify automatic deletion of an extreme observation. Its value is narrower and more reliable: it shows where the empirical residual quantiles differ from a specified reference, prompting targeted checks.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A Q-Q plot has several upper and lower tail points outside the reference line in a balanced pattern. What distributional feature might this suggest, and what limitation should qualify the conclusion?
- Most points follow the line, but the highest point is far above it. Why is “the residuals are heavy-tailed” too strong a conclusion?
- For 500 independent standard normal observations, the expected number with \(|Z|>3\) is about \(500(0.00270)=1.35\). If a set has 2 such observations, does that count alone establish tail excess?
- Why can a Q-Q plot not rule out a time trend in residuals?
Solutions. (1) A balanced, systematic departure in both tails is consistent with symmetric tail excess, but sample variability, residual dependence, and model-estimation effects limit the strength of that conclusion. (2) One extreme point may be isolated; heavy tails are more persuasive when multiple ranks depart progressively in both tails. Investigate the observation rather than assigning a cause from the plot. (3) No. The expected count is 1.35, so observing 2 is not, by itself, compelling evidence; compare the full pattern and account for the assumptions, especially independence. (4) Sorting residuals removes their original time order. A Q-Q plot assesses marginal quantiles, so a residual-versus-time plot is needed to look for a time trend.