Curve of Best Fit
Fit the pattern, then inspect what the curve leaves behind
A curve of best fit is a model of the systematic relationship between variables. It should capture the broad pattern without being forced through every observation, because observed points may contain natural variation or measurement error. The residual is observed response minus predicted response; its magnitude and pattern reveal what the curve misses.
Compare candidate curves on a shared criterion such as sums of squared residuals, which prevent positive and negative deviations from cancelling. Then inspect a residual plot: a remaining curve, trend, or long run suggests the model form is missing structure. Greater flexibility is not automatically better if it captures noise rather than a repeatable relationship.
A fitted curve supports a limited descriptive statement about the observations. Interpolation stays within the observed predictor range; extrapolation assumes an untested continuation. Even excellent fit does not establish causation, because a curve cannot by itself rule out confounding or alternative explanations.
For the same scatterplot, a line has residuals whose squares sum to 10, while a proposed smooth curve has residuals whose squares sum to 6. Which statement is justified by this information?
A nonlinear curve has lower squared residual error than a line for observed data. What conclusion is supported without further design information?