What this answers

This calculator answers "which individual points does this model fit poorly, and which points have unusual predictor values that give them outsized influence over the fit?" two related but distinct questions a single R-squared value cannot answer.

How it is calculated

The residual is simply observed minus fitted; the standardized residual divides that by the model's residual standard error adjusted for each point's own leverage, putting residuals of different points on a comparable scale. Leverage measures how far a point's predictor values sit from the average, using the diagonal of the hat matrix from the underlying regression fit; a point with high leverage has the potential to pull the fitted line toward itself, whether or not it actually does so.

Worked example

For four ordinary points and one point with an x value of 50, far beyond the others, that last point's leverage far exceeds the standard 2p-over-n threshold, correctly flagging it as a high-leverage observation, whether or not its outcome value turns out to fit the overall pattern well.

Assumption audit

Calculated from your data: every point's fitted value, residual, standardized residual, and leverage, plus a flag for leverage above the conventional 2p-over-n threshold.

Evidence to review: whether flagged points share something in common (a data-entry error, a genuinely unusual case, or a sign the model's functional form is wrong for part of the data range).

You must verify: the real-world validity of any flagged point before deciding to exclude or investigate it further; a statistical flag is a starting point, not justification on its own for removing a data point.

What this result does not mean

A flagged residual or leverage value is evidence to investigate, not proof of a bad model or a bad data point; high leverage only means a point has the potential for influence, which the Cook's Distance Calculator measures more directly by combining leverage with how far that point's residual actually pulls the fit.

Limitations

This calculator supports a single predictor for its current input shape; the underlying math generalizes to multiple predictors, but this specific page's data-entry format is built around one x and one y column.