Why a Residual Plot Is Not a Complete Picture
A residual plot is a useful diagnostic display, but it is not a transparent window onto every feature of the data. Points can overlap, clusters can be pooled together, and some predictor regions may contain very few observations. In addition, an observation that strongly affects the fitted model can have a small residual itself while changing the residuals of many other observations.
These limitations matter because the plot shows only selected coordinates, usually a residual against one fitted value or predictor. It does not automatically show group membership, how many observations share a location, or how the fitted model would look under a different data configuration. A pattern that appears absent may be hidden by the display; a pattern that appears present may be driven by one unusual observation or by a feature of the sampling layout.
For context, suppose the fitted model is based on \(Y_i=\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}+\varepsilon_i\). Ordinary least squares chooses coefficients to minimize the sum of squared residuals. Residual plots are commonly read in relation to a model whose mean response is represented adequately by the included predictors, with errors having mean zero and, for the usual standard inference, independence and constant variance; normality is additionally used for exact small-sample inference. A plot can suggest that something deserves investigation, but a lack of visible pattern does not verify these conditions. Normal quantile plots and residual displays across categories address different features; here the concern is how the arrangement and visibility of observations can mislead.
Clusters Can Hide Opposing Patterns
A cluster is a group of observations concentrated in part of the predictor space. Clusters can arise because data were collected at different sites, in different periods, or from different categories. If the plot does not distinguish those groups, a pooled display may blend patterns that differ by group. In a plot against one predictor, observations with different values of other predictors can also be superimposed or appear as a broad cloud.
A simple cancellation explains one way pooling can hide structure. At a given predictor value, suppose one group has residuals averaging \(+1\) and another averages \(-1\). The pooled average is \((1+(-1))/2=0\), even though neither group has an average near zero. A pooled summary can therefore look unremarkable while group-specific residual patterns oppose one another. The points may still be visible in the plot, but overplotting, a large number of observations, or the absence of group labels can make the distinction difficult to see.
Worked Example: Opposing Trends Within Two Groups
In a hypothetical analysis, a model includes a numeric predictor \(x\) and a group indicator, but does not include an interaction. The following residuals are displayed against \(x\). The groups have observations at the same three values of \(x\).
| Group | Residual at \(x=1\) | Residual at \(x=2\) | Residual at \(x=3\) |
|---|---|---|---|
| A | -1 | 0 | 1 |
| B | 1 | 0 | -1 |
If color or symbols distinguish the groups, the display shows a rising sequence for A and a falling sequence for B. If the points are pooled and summarized by the mean residual at each \(x\), however, the calculation at each value is \(0\): at \(x=1\), \((-1+1)/2=0\); at \(x=2\), \((0+0)/2=0\); and at \(x=3\), \((1+(-1))/2=0\). A summary of pooled means would hide both opposing trends.
This example shows why it is useful to retain group labels or use separate panels when groups are relevant. It does not establish that any particular model should include an interaction: that conclusion requires considering the scientific question and the full pattern, not just the pooled means. The practical lesson is narrower: a pooled residual display can conceal group-specific structure.
Clustering can also make a plot look more structured than it is. If groups occupy different ranges of a predictor, their separation may produce a visible band or gap. That appearance might reflect group composition rather than a single smooth pattern shared across all observations. Check whether groups differ in other predictors or in how observations were collected before assigning one interpretation to the pooled shape.
Sparse Regions Give Weak Visual Evidence
A residual plot is most informative where observations are reasonably plentiful. In a sparse region, one point can determine the apparent direction of a local pattern, and a gap contains no direct evidence about what happens between the observed regions. A smooth curve drawn through a few points can look persuasive, but the curve between widely separated observations is largely an interpolation choice, not a feature established by the data.
Worked Example: A Gap in the Predictor Range
Suppose a hypothetical diagnostic plot contains six observations at \(x=0,1,2\) and \(x=8,9,10\), with no observations between \(x=2\) and \(x=8\). The residuals in the lower range are \(-0.2, 0.1, 0.1\); those in the upper range are \(0.1, -0.1, 0.0\).
The visible residuals are small and do not show an obvious pattern within either observed range. But the plot cannot tell us whether the mean response behaves linearly throughout the unobserved interval from \(x=2\) to \(x=8\). There are no residuals there to inspect. It would be too strong to conclude that the fitted relationship has been checked equally well over the entire predictor range.
Now suppose the residual at \(x=9\) were \(1.2\) rather than \(-0.1\). That single point would be a large part of the visual evidence about the upper range; it would not make the region well-sampled. The appropriate description is that the plot raises a question about that observation, while evidence across the gap remains limited. The data collection context and the model's intended use help determine whether the sparse region matters.
Sparse regions also complicate comparisons between groups. A group with many observations may reveal a stable-looking pattern, while a group represented by only one or two points cannot support the same visual judgment. Do not treat a few points as strong evidence that a pattern is absent merely because no consistent shape is apparent.
Overplotting Hides the Number of Observations
Overplotting occurs when multiple observations fall at the same displayed coordinates, or close enough that the plotting symbols cover one another. A plot may then show fewer visible marks than there are observations. This can make a heavily repeated location appear no more important than an isolated point, or make a large number of observations look like a small sample.
Worked Example: Three Marks Represent Six Observations
A hypothetical simple regression uses six observations. The table gives the predictor value, fitted value, residual, and number of observations at each displayed coordinate. Each pair of observations has the same predictor value and residual.
| \(x\) | Residual \(e\) | Count at this coordinate |
|---|---|---|
| 0 | 1 | 2 |
| 1 | -2 | 2 |
| 2 | 1 | 2 |
The plot can show just three visible marks: \((0,1)\), \((1,-2)\), and \((2,1)\). Each mark represents two observations, so there are six observations, not three. For a consistency check, if the fitted line is \(\hat{y}=10+2x\), the observed responses at these three predictor values are \(11,10,15\), respectively, repeated twice each. The mean of the six responses is \(12\), and the fitted line has mean fitted value \(12\). The slope calculation gives \((0-1)(11-12)\cdot2+(1-1)(10-12)\cdot2+(2-1)(15-12)\cdot2=8\) for the numerator and \(4\) for the sum of squared deviations in \(x\), so the slope is \(8/4=2\). Thus the stated line is the least-squares line, and the residual coordinates in the table are consistent with it.
The three marks alone conceal the multiplicity. A plot can make repeated coordinates visible by using counts, transparency, or symbols whose size reflects the number of observations. Jittering points slightly can also help reveal overlap, but it changes their displayed locations: it should be clearly identified as a visual aid, not mistaken for the recorded data or used to alter calculations.
Points may also overlap because a plot rounds predictor values or residuals before drawing them. When an apparent cluster matters, check the underlying records or use a display that reports counts. Do not infer that observations are identical in every respect just because they occupy the same pixel or rounded coordinate.
Influential Observations Can Change the Pattern
An observation is influential in a practical sense when including it changes the fitted model substantially. An observation far from the others in predictor space can have substantial leverage on the fitted line or surface. It may pull the fit toward itself, leaving a small residual for that observation while changing the residuals of many others. Conversely, a visually unusual residual does not by itself establish that the observation strongly affects the fit.
This creates two possible misreadings. A point can help make the model follow a feature that is not representative of most observations, so its own residual looks modest while the bulk of the data show a pattern. Or a single observation can leave a conspicuous residual and make a plot look patterned even though the rest of the data do not share that behavior. The residual plot alone does not resolve whether the observation is a recording error, a valid but unusual case, or evidence that the model needs reconsideration.
Worked Example: One Distant Predictor Value Changes the Fit
Consider a hypothetical data set with five points: \((x,y)=(0,0),(1,0),(2,0),(3,0),(10,10)\). The first four observations lie at response zero; the fifth is far from them in the predictor direction. Fit a simple least-squares line with an intercept.
The means are \(\bar{x}=16/5=3.2\) and \(\bar{y}=10/5=2\). The sum of squared deviations in \(x\) is \(S_{xx}=\sum (x_i-\bar{x})^2=62.8\). The cross-product sum is \(S_{xy}=\sum (x_i-\bar{x})(y_i-\bar{y})=68\). Therefore the fitted slope is \(b_1=S_{xy}/S_{xx}=68/62.8\approx1.083\), and the intercept is \(b_0=\bar{y}-b_1\bar{x}\approx-1.465\). The fitted line is approximately \(\hat{y}=-1.465+1.083x\).
The residuals for the first four observations are approximately \(1.465, 0.382, -0.701,\) and \(-1.783\). The distant point at \(x=10\) has fitted value about \(9.363\), so its residual is only \(10-9.363=0.637\). The observation farthest out in \(x\) does not have the largest residual. Yet its inclusion has tilted the fitted line across the range occupied by the other four observations.
If the distant observation were omitted, the remaining four responses are all zero, and their fitted line would be \(\hat{y}=0\). This comparison illustrates how one observation can change the residual pattern for the main cluster. It is not a rule for deleting observations: the point may be valid and important to the question. It is a reason to inspect the case and the context before treating the full-sample plot as an uncomplicated summary.
How to Read a Potentially Misleading Plot
When a residual display seems unusually empty, unusually patterned, or inconsistent with other evidence, first check how it was made and what observations it represents. A careful reading asks whether the plot is concealing group structure, gaps, or repeated coordinates before drawing a strong conclusion from its overall appearance.
Identify which predictor or fitted value is on the horizontal axis and whether rounding or the plotting scale could make distinct observations overlap.
Look for meaningful categories or clusters, then note predictor regions with few or no observations. A blank region supplies no direct evidence about behavior there.
Inspect the observation and its context. Ask whether the fitted relationship appears to depend heavily on it, without assuming that a large residual and a large effect on the fit are the same thing.
Describe what the displayed observations suggest, which groups or regions are poorly represented, and what cannot be concluded from this plot alone.
Common Mistakes
- Counting marks as observations. Several cases can share displayed coordinates. Use counts or the underlying data before judging how many observations support a pattern.
- Calling an empty gap evidence of a good fit. A region with no observations cannot show whether residuals there follow a pattern.
- Assuming pooled groups have one common pattern. Opposing group-specific patterns can cancel in a pooled summary. Preserve meaningful group labels while inspecting the plot.
- Equating a large residual with influence. A large residual describes a discrepancy for one observation; influence concerns how much the fitted model changes when that observation is included.
- Removing a point just because it changes the plot. Verify its accuracy and consider its relevance. A valid influential case may be essential to the population or predictor range of interest.
Check Your Understanding
Try these questions without looking back.
- At one predictor value, one group has average residual \(+2\) and another has average residual \(-2\). What pooled average would result, and what feature could the pooled summary hide?
- Why is a plot with no visible pattern in a predictor interval uninformative if that interval contains no observations?
- A residual plot contains three visible marks, each representing four observations. How many observations are represented, and what display feature should be added or used to communicate that fact?
- Explain how a distant predictor value could have a modest residual but still change the fitted line substantially.
- What is the difference between an observation having a large residual and being influential?