Asking a specific question of your groups
Today let's learn how to ask a fertilizer trial one specific question, instead of settling for the one blunt question an ANOVA already answers.
A field trial gives forty corn plots one of four nitrogen rates: 0, 60, 120 or 180 pounds per acre, ten plots at each rate. At harvest, every plot's yield is recorded in bushels per acre. Here is what all forty plots look like, grouped by the rate each one received.
Yield climbs steadily from 0 to 120 pounds per acre, then barely moves by 180. "Do these four groups differ" is too blunt a question for a pattern like that. Sharper questions are available: does any nitrogen help at all, does yield keep climbing as the rate climbs, and does one particular jump in rate matter.
What makes a contrast "planned"
A contrast is a weighted sum of the group means, where the weights add up to zero. Each contrast encodes one specific question about the groups, and the weights are how you write that question down in numbers.
Here are three such questions about the corn trial, in plain words.
- Does any nitrogen help at all, compared to no nitrogen?
- Does yield keep rising as the nitrogen rate rises, across all four rates?
- Does going from 60 to 120 pounds per acre actually matter, on its own?
Each of those is a contrast waiting to be written down. What makes a contrast "planned" is that you decide the question before you run any test, not after looking at which gaps in the data happen to look biggest. That order matters, because picking the biggest-looking gap after seeing the data and then testing it is a kind of search, and a search costs more than one planned question does.