Valid Extreme Cases and Data Errors

Why an Unusual Case Needs Investigation, Not Automatic Deletion

A regression diagnostic can identify a case that has an unusual predictor value, a large residual, or a strong effect on the fitted coefficients. That is a reason to look more carefully at the record; it is not evidence by itself that the record is wrong. The distinction matters because deleting a genuine observation can change what population the analysis describes and can conceal important variation.

The first task is therefore to establish whether the recorded values accurately represent an eligible observational unit. Only after checking the record should an analyst decide how it belongs in the analysis. Statistical influence describes what happens to a fitted model when a case is included; it does not determine whether the case is valid.

The Model and the Role of Influence

For context, consider simple linear regression with predictor \(X\) and response \(Y\). The working model is \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), where \(\beta_0\) is the intercept, \(\beta_1\) is the slope, and \(\varepsilon_i\) represents variation not explained by the linear mean. The ordinary least squares (OLS) fit chooses the intercept and slope to minimize the sum of squared residuals.

A regression interpretation relies on assumptions appropriate to the question. The conditional mean should be adequately represented by the chosen linear form, observations should correspond to distinct and appropriately sampled observational units, and the conditional errors should have mean zero. Standard-error-based inference additionally relies on appropriate error-variance assumptions; exact small-sample \(t\) procedures also use a normal-error assumption. These assumptions are not guaranteed by fitting a line or by checking whether a record is valid. In particular, a handful of points cannot establish that a linear model is adequate.

The calculations below compare OLS slopes to show how records affect a fit. The calculations themselves are algebraic; they do not establish a causal relationship or justify statistical inference. In each example, treat the line as a working summary and do not infer that its assumptions have been verified merely because a slope can be calculated.

Why Extreme Values Can Change a Regression Fit

In simple regression, the OLS slope can be written as the ratio of the cross-product of centered predictor and response values to the centered predictor sum of squares. This expression helps explain why an unusual case may matter: a point far from the predictor mean can contribute strongly to both the numerator and denominator.

This formula follows by minimizing the squared residuals. For any fixed slope \(b_1\), the minimizing intercept is \(b_0=\bar{y}-b_1\bar{x}\), so fitted residuals can be written as \((y_i-\bar{y})-b_1(x_i-\bar{x})\). Differentiating the sum of their squares with respect to \(b_1\), setting the derivative to zero, and solving gives the displayed expression, provided the predictor values are not all identical.

An extreme \(x_i\) can be far from \(\bar{x}\), giving it substantial leverage in the calculation. If its response also differs from what the rest of the data would suggest, the fitted line may move noticeably. This explains why influence prompts investigation, but the equation contains no information about whether the observation was recorded correctly.

How to Investigate a Flagged Observation

A useful investigation separates data quality, study eligibility, and statistical influence. Begin with the record and the collection process rather than with a preferred regression result. Depending on the study, checks may include the original form or instrument file, units and permitted ranges, the case identifier, coding rules, data-import steps, and whether values from different individuals were joined incorrectly.

A surprising value is not automatically impossible. A value outside a typical range might be a genuine rare event; a value inside a plausible range might still have been attached to the wrong person. Compare the observation with independent records or source material when possible. Do not use agreement with the fitted line as the standard for deciding whether the measurement is true.

If the source cannot resolve a questionable value, do not label it a confirmed error. Document the uncertainty and use a defensible analysis plan, such as reporting sensitivity to plausible alternatives when that is appropriate. A missing or uncertain value is not permission to substitute whatever value would make the fitted model more convenient.

Exclusion, Duplicates, and the Target Population

A valid observation can sometimes be excluded from a particular analysis for a reason fixed by the study design. For example, a study may target a defined age range, location, or treatment group. A verified case outside that scope is not a data error; it is simply not part of that analysis’s target population. That decision should follow the stated eligibility rule, not a post hoc wish to make the regression less sensitive.

A duplicate deserves similar care. Two records that look identical may be an accidental copy, but they may also represent two different people who happen to share the recorded values. Check identifiers and source records before removing either. A mistaken duplicate gives that case extra weight in OLS, while removing a genuine second case discards valid information.

Why Deleting Valid Cases Can Distort the Analysis

The fitted regression summarizes the cases and target population represented in the analysis. Removing a valid, eligible observation because it pulls the slope away from a preferred result can alter that summary and hide genuine behavior at the edge of the predictor range. It can also make the remaining cases look more consistent with the model than the full eligible data actually are.

Deleting cases selectively after seeing their influence is especially difficult to defend: the rule depends on the result it is being used to change. If the primary analysis excludes a valid case, the reason should come from a prespecified design rule or a clearly justified scientific scope, not from the fact that the case is inconvenient. A sensitivity analysis that compares the full-data fit with a fit omitting a verified case can be informative, provided it is labelled as a sensitivity analysis and the full-data result is not silently replaced.