Why Axis Limits Are Part of the Diagnostic
Axis limits determine which values are visible and how much visual space a given difference occupies. In automated diagnostic reports, that choice is repeated across many panels, so a rule that seems harmless on one plot can make comparisons misleading when applied inconsistently. A narrow panel-specific range can make small fluctuations appear large; a broad range chosen because of one extreme value can compress patterns in every other panel.
The goal is not to make every plot look smooth or to remove unusual observations from view. It is to choose limits that make comparable panels visually comparable, while ensuring that observations outside a displayed range remain explicitly visible as out-of-range values. An extreme point is a reason to check the diagnostic and the plotting rule, not by itself a reason to discard the point.
A limit-selection rule is defensible when it is specified consistently for the panels being compared, uses the intended plotted data, and does not silently hide values. The rule should also respect the meaning and domain of the plotted quantity. For example, a plot of a nonnegative quantity should not imply negative values are possible merely to make the scale symmetric. The appropriate limits depend on the diagnostic question and the plotted variables; the key is to make that choice explicit and reproducible.
Shared Limits and a Transparent Calculation
For panels that are intended to be compared, first identify the group of panels and the data range that the rule will use. A simple shared-range rule uses the most extreme value across all included panels and adds a stated margin. The margin may be a fixed amount or a documented proportion of the range, but it should be selected for a reason and applied uniformly. It is not a substitute for checking whether the resulting range is useful.
Worked Example: Choosing Shared Limits for Residual Panels
Suppose two hypothetical panels show standardized residuals by occasion. Panel A contains \(-1.2,\ 0.3,\ 2.4,\ -0.8\), and panel B contains \(-2.1,\ 1.7,\ 0.4\). The panels will be compared directly, so choose symmetric shared limits using a fixed padding of 0.5 residual units beyond the greatest absolute observed value.
The greatest absolute value across the panels is \(M=\max(1.2,0.3,2.4,0.8,2.1,1.7,0.4)=2.4\). With the stated padding rule, the positive limit is \(B=M+0.5\). Written as the calculation for these data, \((2.4/0.5+1)\times 0.5=5.8\times 0.5=2.9\).
Therefore the shared limits are \([-2.9,2.9]\). All listed observations lie within these limits: the minimum is \(-2.1\), and the maximum is \(2.4\). The range is symmetric around zero, uses the same scale for both panels, and leaves 0.8 units below the minimum and 0.5 units above the maximum. The 0.5-unit padding is a plotting policy, not a statistical guarantee; another margin could be appropriate if it were justified and applied consistently.
For a non-symmetric quantity, limits need not be symmetric. If the plotted values are nonnegative and the meaningful lower boundary is zero, a common range might begin at zero and extend far enough to include the pooled maximum plus a stated margin. In either case, record the rule and the values used to compute the limits so that the plot can be regenerated and checked.
Why Separate Autoscaling Can Change the Story
A shared vertical range supports visual comparison because the same change in data occupies the same fraction of the panel height. With panel-specific ranges, that relationship changes from panel to panel. The following simple derivation shows the effect for a linear axis with lower limit \(L\), upper limit \(U\), and panel height \(H\). A value \(y\) is mapped to vertical position \(H(y-L)/(U-L)\), measured upward from the bottom. A data change of \(\Delta y\) therefore occupies the fraction \(\Delta y/(U-L)\) of the panel height. If \(U-L\) changes between panels, the same data change has a different visual size.
For example, a one-unit change occupies \(1/4=0.25\), or 25% of a panel whose range is four units, but \(1/10=0.10\), or 10% of a panel whose range is ten units. Separate autoscaling can therefore amplify a modest fluctuation in one panel relative to another, even though the underlying change is the same size.
Worked Example: Detecting a Scale-Driven Visual Difference
Two hypothetical residual-by-time panels each contain a change from \(-0.5\) to \(0.5\), a one-unit change. The first panel uses limits \([-2,2]\); the second uses limits \([-5,5]\). Assume both panels have the same physical height.
For the first panel, the axis span is \(2-(-2)=4\). The one-unit change occupies \(1/4=0.25\), or 25% of the panel height. For the second, the span is \(5-(-5)=10\), and the same change occupies \(1/10=0.10\), or 10% of the height.
The data changes are identical, but the first appears 2.5 times as large in relative panel height: \(0.25/0.10=2.5\). A viewer comparing these panels could wrongly conclude that the first has a more pronounced trend. Shared limits would avoid this particular scale difference; they would not, by themselves, establish whether either trend is meaningful.
Keeping Extreme Values Visible
A shared range based on pooled extrema is often a straightforward choice, but it can become so broad that central patterns are hard to inspect. A fixed display range may be useful when the same diagnostic range is required across many analyses or reports. In either case, a value beyond the displayed limits must not disappear without explanation. Common safeguards include an out-of-range marker at the boundary, a label giving the actual value, and a count or listing of affected observations. When central detail and the full range cannot both be read clearly in one panel, a separate full-range view can supplement—not silently replace—the primary plot.
A boundary marker alone can be ambiguous: a point drawn at the upper limit might be an actual value at that limit or a value beyond it. Indicate which is the case, and make the actual value retrievable in the plot or its associated diagnostic output. Do not move an out-of-range value to the boundary and then present that boundary coordinate as the observation itself.
Worked Example: A Fixed Range With an Above-Limit Observation
A hypothetical diagnostic report uses fixed standardized-residual limits of \([-3,3]\) for every panel. One panel contains residuals \(-1.1,\ 0.2,\ 2.7,\ 4.6\). The fixed range supports comparison across reports, but the value \(4.6\) lies above the upper limit.
The amount above the upper limit is \(4.6-3=1.6\) residual units. Three of the four observations lie within the displayed range, and one of four, or \(1/4=0.25\) (25%), lies above it. The point at \(4.6\) must not be omitted or shown as if its value were exactly 3.
A suitable plot can place a distinct upward boundary marker at 3 and label it “4.6; 1.6 above limit,” while reporting that one of four observations exceeds the range. The marker tells the viewer where the value lies relative to the display; the label gives its actual value. If the point and central pattern cannot be assessed adequately together, include a full-range companion view as well. The fixed limits remain the same across panels, and the extreme value remains identifiable.
Comparing Practical Limit-Selection Rules
| Rule | Useful when | Main risk and safeguard |
|---|---|---|
| Panel-specific autoscaling | Inspecting the local shape of one panel in isolation | Visual magnitudes are not directly comparable across panels. Label the rule and avoid using independently scaled panels to compare effect size or pattern strength. |
| Shared limits from pooled extrema | Panels should be compared and all values can fit on a readable scale | A single extreme may compress the central pattern. Check readability and retain the extreme in the view. |
| Pre-specified fixed limits with explicit out-of-range display | Consistent comparison across runs, reports, or analyses is important | Values may fall outside the display. Show boundary indicators and actual values or provide a clearly linked full-range view. |
There is no universally best rule. Panel-specific limits can help reveal local structure, but are a poor default for direct visual comparisons. Pooled-extrema limits preserve all plotted values in the main panel, but an unusually large value can dominate the scale. Fixed limits make repeated displays consistent, but require an explicit policy for observations beyond those limits. A broken axis or a truncated plot without clear annotation is generally a poor solution because it can make distances and values difficult to interpret.
Checks for Automated Plotting
Automated plotting is reliable only when the limit rule is tested against the data and the intended comparison. The checks below are particularly useful when the same code generates many panels.
- Confirm the comparison group. Verify that every panel meant for direct comparison uses the same limits. Do not assume that a plotting library’s automatic defaults will do so.
- Check extrema against limits. Count values below and above the displayed range. A zero count means the range includes the plotted values; nonzero counts require an explicit display safeguard.
- Inspect the central pattern. A range that includes every extreme can still make the main data pattern too compressed to assess. If so, add a clearly identified supplementary view rather than concealing the extreme.
- Check the axis domain and units. Limits should make sense for the quantity being plotted. A nonnegative quantity, for example, should not be given a negative lower limit without a reason.
- Recheck after data changes. Missing or newly added values can change pooled extrema, counts, and the usefulness of a fixed range. Report generation should make these changes detectable.
These checks diagnose presentation problems, not model adequacy. A plot with well-chosen limits can still show sparse data, unusual residual structure, or other patterns requiring investigation. Conversely, changing limits can make a pattern easier or harder to see, but does not change the underlying observations.
Common Mistake and Limitations
Axis-limit choices also have limits as a diagnostic tool. Shared scales improve the fairness of visual comparisons but cannot remove differences in data density or uncertainty, and they do not determine whether a pattern is scientifically important. Very broad ranges may obscure central structure; very narrow ranges may hide extremes. A secondary view can help, but it should be visibly distinguished from the primary, comparable display. State the limit rule in the plotting workflow so reviewers can understand what is shown and what lies beyond it.
Check Your Understanding
For each question, consider both the visual comparison and whether any observations could be hidden.
- Two panels are intended for direct comparison, but each uses its own automatic limits. What is the main interpretive risk?
- A panel has values \(-1.0,\ 0.5,\ 2.2\). A symmetric limit rule adds 0.4 units to the greatest absolute value. What are the limits?
- A fixed display range ends at 5, and an observation equals 6.3. How far beyond the range is it, and what should the plot communicate?
- Why can a shared range based on all observed extrema make the central pattern harder to inspect?
Solutions. (1) The same data change can occupy different fractions of each panel’s height, so visual differences in pattern strength may be caused by scaling rather than the data. Use shared limits for direct comparison, or clearly state that the panels are independently scaled and do not compare their apparent magnitudes. (2) The greatest absolute value is \(2.2\), so the bound is \(2.2+0.4=2.6\), giving limits \([-2.6,2.6]\). (3) The observation is \(6.3-5=1.3\) units above the upper limit. Show it with an explicit above-limit marker and its actual value, and report the number of observations beyond the range; do not represent it as a value of 5. (4) A large extreme can expand the axis span, making ordinary values occupy a smaller portion of the panel. A supplementary view may make the central pattern easier to inspect, but the extreme must remain visible and identifiable.