Interpreting Cook's Distance

What Cook’s Distance Summarizes

A regression case can matter to the fitted model in more than one way. A case may have an unusual response, an unusual predictor pattern, or both; what matters for influence is how much the fitted model changes when that case is left out. Cook’s distance compresses the changes in all fitted values into one number for each case. It is an overall diagnostic, not a test of whether an observation is valid or whether a model is correct.

To interpret the measure, keep the comparison fixed. Fit the chosen least-squares model to all \(n\) observations. Then, for case \(i\), fit the same model again after omitting that case. Use the two fits to predict the response at each of the original \(n\) predictor patterns. Thus, the deleted fit also gives a prediction at the predictor pattern of the omitted case, even though that case was not used to estimate the deleted fit.

In this definition, \(\hat{y}_j\) is the fitted value for the \(j\)th original predictor pattern from the full-data fit, and \(\hat{y}_{j(i)}\) is the prediction at that same pattern from the fit with case \(i\) deleted. The \(MSE\) is from the full-data fit. This is the deleted-fit definition established in the earlier material on Cook’s Distance from Deleted Fits; here the emphasis is on what its value says.

The numerator measures the total squared change in fitted responses across all \(n\) predictor patterns. The denominator scales that change by the model’s residual variation and by the number of fitted coefficients. Both numerator and denominator have squared-response units, so \(D_i\) has no units. A larger value means a larger overall change relative to this scale; it does not, by itself, identify which fitted values changed most or which coefficient changed most.

Reading the Size of a Distance

A Cook’s distance is a summary, not a distance in the original predictor space and not a residual measured in response units. For example, \(D_i=2\) does not mean that a prediction changed by two units. It means that the total squared change in the fitted values is twice the chosen scaling quantity \(p\,MSE\). To express a typical change in response units, it is useful to consider the root-mean-square change across the \(n\) original predictor patterns.

This relationship follows directly from the definition: multiply both sides of \(D_i=\sum_j(\hat{y}_j-\hat{y}_{j(i)})^2/(p\,MSE)\) by \(p\,MSE\), divide by \(n\), and take the nonnegative square root. Since \(\sqrt{MSE}\) is the residual standard deviation estimate, the formula expresses the root-mean-square fitted-value change in response units. It is an average across the original predictor patterns: some individual predictions may change more and others less.

Comparing Cases in One Dataset

Cook’s distances are most directly compared when cases come from the same fitted model, because they share the same \(p\) and \(MSE\). The case with the larger distance produced the greater overall fitted-value change on this common scale. The comparison ranks sensitivity to case deletion; it does not establish why a case is influential or whether its influence is problematic.

Cutoffs Are Screening Heuristics

There is no single Cook’s-distance cutoff that identifies every influential case in every regression setting. Two familiar rules of thumb flag cases with \(D_i>1\), or with \(D_i>4/n\). Some references use a closely related sample-size adjustment such as \(4/(n-p)\). These are conventions for drawing attention to cases for further examination, not formal decision rules and not guarantees that cases above a threshold are harmful.

The rules can flag different cases. In the hypothetical dataset above, \(4/n=4/12\approx0.3333\). Using the strict rule \(D_i>4/n\), Cases B, C, D, and E would be flagged, while A would not. Using \(D_i>1\), Cases C and E would be flagged; B is exactly 1 and so does not exceed that strict threshold. The disagreement is informative: a “flag” depends partly on the chosen screening convention.

What to Do With a Flagged Case

A large Cook’s distance signals sensitivity: the fitted values depend noticeably on that case relative to the scale used in the diagnostic. That is worth understanding, but influence is not automatically a defect. A valid observation at an unusual predictor pattern may be important to the population or range of interest. Conversely, an influential data-entry error deserves correction. The diagnostic alone cannot distinguish those situations.

A sensible interpretation therefore separates the numerical result from the follow-up judgment. First report which case or cases have comparatively large Cook’s distances and, if useful, the screening convention used. Then check records and measurement procedures for possible errors, consider whether the observation belongs to the target population, and examine whether the model form is adequate for the data. Any decision to retain, correct, or exclude a case needs a reason grounded in the data and analysis—not merely a threshold crossing.

Cook’s distance also depends on the fitted model. Changing the included predictors, the functional form, or the observations can change fitted values and therefore change the distances. It describes sensitivity within the specified least-squares fit; it does not certify that the regression assumptions are satisfied or that the model’s mean structure is appropriate.

Common Mistake

Wrong reasoning: “This case has \(D_i>1\), so it is an outlier and must be removed.” This confuses an influence screen with a data-quality rule. Cook’s distance describes how the whole set of fitted values changes under deletion. It does not, by itself, identify an unusual response, demonstrate a recording error, or justify excluding a valid observation.

Correction: State what the diagnostic actually shows: omitting the case produces a comparatively large overall change in fitted values on the Cook’s-distance scale. Then investigate the case and the model in context. A flagged case may warrant careful reporting or sensitivity analysis, but the threshold alone does not determine the action.