Statistics Glossary → Polynomial (trend) contrasts
Polynomial (trend) contrastsconcept
Mutually orthogonal contrasts that test whether the means follow a linear, curved (quadratic) or more complex trend across the levels of an ordered factor.
When the levels of a factor have a natural order (doses of 10-20-30-40 mg, or measurements at months 1-2-3), the question 'how do the means change across levels?' is more informative than 'do the groups differ?'. Polynomial contrasts split this change into independent components. For a four-level factor the linear contrast has coefficients (-3, -1, 1, 3), the quadratic (1, -1, -1, 1) and the cubic (-1, 3, -3, 1). With k levels there are at most k - 1 components, and their sums of squares exactly partition the between-groups sum of squares.
A significant linear component indicates a steady rise or fall; a significant quadratic component indicates a curve with a peak or a trough. The standard coefficients assume equally spaced levels; if the spacing is unequal (say 0, 10 and 100 mg) the coefficients must be recomputed from the actual values. Higher-order components (4th and 5th degree) are rarely interpretable, and without a theoretical reason the linear and quadratic components are usually enough.
An everyday example
A gardener gives tomato seedlings four amounts of fertilizer: 0, 50, 100 and 150 grams. Average yields come out at 2, 4, 5 and 4.5 kg. The linear contrast shows that yield generally rises with fertilizer, while the quadratic contrast shows that the rise levels off and dips slightly, meaning that too much fertilizer does not help.
In MerQur
It is offered as 'Polynomial trend' (from linear up to 5th degree) in One-Way ANOVA and as 'Polynomial contrasts' (Linear, Quadratic, Cubic) in Repeated Measures ANOVA and Mixed-Design ANOVA; the mixed design also tests the group x trend interaction.
Analyses where this concept appears
One-Way ANOVARepeated Measures ANOVAMixed-Design ANOVA, Split-Plot