Common Misreadings of Residual-Based Specification Checks

Residual Plots Provide Clues, Not Verdicts

A residual plot can reveal features that a fitted regression equation does not capture. Yet it is easy to read too much into one: a cloud that looks unstructured may be taken as proof that the model is adequate, while a curve or cluster may be treated as proof that a particular term is missing. Neither conclusion follows from the plot alone. Residuals are calculated from the fitted model, and their appearance depends on the data, the specification, and the way the diagnostic is displayed.

The aim of a residual-based specification check is more modest and more useful: look for evidence that challenges the current model, work out what explanations could produce that evidence, and decide what additional checks are appropriate. A residual pattern can motivate a candidate explanation; it cannot establish that explanation by itself.

What a Residual Check Is Assessing

For observation \(i\), let \(y_i\) be the observed response, \(\hat{y}_i\) its fitted value, and \(e_i=y_i-\hat{y}_i\) its residual. A residual is the observed response minus the response predicted by the fitted model. In a residual-versus-fitted plot, the horizontal coordinate is \(\hat{y}_i\) and the vertical coordinate is \(e_i\). Other displays can put a predictor, an observation order, or a group label on the horizontal axis.

For orientation, consider the usual regression model \(Y_i=\beta_0+\beta_1x_{i1}+\cdots+\beta_px_{ip}+\varepsilon_i\). The conditional mean is represented by the specified regression function; a correctly specified mean has \(E(\varepsilon_i\mid X)=0\), where \(X\) denotes the included predictor information. Classical regression inference also commonly assumes independent errors with constant variance, and normal errors for exact small-sample \(t\) and \(F\) procedures. These assumptions have different roles: residual patterns may alert us to a problem with the mean structure or spread, while a residual plot alone cannot verify independence or normality.

A plot may be inspected in several ways: Is the residual center near zero across the range? Is there a smooth trend, a change in spread, a cluster, or an unusual observation? These questions identify possible concerns. Determining what the concern means requires checking how the plot was formed, which observations contribute to it, and whether a plausible alternative explanation is supported.

A Constraint Created by Least Squares

Some appearances of residuals are partly mechanical. In ordinary least squares with an intercept, the fitted residuals sum to zero. More generally, the residuals are orthogonal to each included predictor. These facts describe the fitting calculation; they do not certify that the regression function is suitable.

To see why, write the fitted values as \(\hat{y}_i=\hat{\beta}_0+\sum_{j=1}^{p}\hat{\beta}_jx_{ij}\). Least squares chooses the coefficients to minimize \(SSE=\sum_i(y_i-\hat{y}_i)^2=\sum_i e_i^2\). Differentiating \(SSE\) with respect to the intercept gives

At the least-squares minimum, this derivative is zero, so \(\sum_i e_i=0\). Differentiating with respect to an included coefficient \(\hat{\beta}_j\) similarly gives \(\sum_i x_{ij}e_i=0\). Thus a residual plot against a predictor already in the model can show a zero overall linear association as a consequence of fitting. It may still reveal curvature, clustering, changing spread, or other structure, but a zero average or zero linear association is not a general adequacy check.

A Visible Pattern Is Not a Diagnosis

A curve in residuals may suggest that the mean function needs more flexibility. A fan-shaped spread may suggest nonconstant variance. Clusters may suggest an omitted subgroup or an unrecorded feature. Each is a useful lead, but more than one mechanism can produce a similar picture. A few observations, a data or coding problem, or random variation can also create a pattern that looks striking in a small sample.

The correct next step depends on the proposed explanation. For a possible mean-pattern error, inspect residuals against the relevant predictor and consider whether an added term has a defensible interpretation. For a possible spread difference, examine the scale of residual variation across the range rather than treating a change in the average as the same problem. When groups appear, compare their residual behavior by group and check whether the grouping has a substantive or data-recording explanation. The earlier material on partial \(F\) tests and lack-of-fit tests provides formal comparisons for suitable cases; a visual impression is not a substitute for those procedures.

Why One Plot Can Hide What Another Reveals

A pooled plot can conceal structure when observations with different characteristics are combined. Conversely, adding many views creates more opportunities to notice chance features. A careful diagnostic review is therefore not “keep plotting until something looks convincing.” It is to use displays that answer distinct questions and to treat a feature as more persuasive when it is coherent with the data structure and supported by other evidence.

Common Misreadings and a More Careful Practice

The interpretation of a diagnostic should match what was actually checked. In particular, do not confuse “I did not see a pattern” with “the model assumptions hold,” or confuse “I saw a pattern” with “I know its cause.” The following distinctions help keep conclusions proportional to the evidence.

A useful workflow is to name the concern before choosing the next check. If the concern is a mean pattern, use a display tied to the relevant predictor and consider a model comparison. If it is unequal spread, examine residual variability across fitted values or relevant groups. If it is clustering, compare the groups rather than relying only on a pooled cloud. Also verify that the variables, units, and records used to fit the model are correct; a striking residual pattern need not have a statistical-model explanation.