The rookie's AFL model came out clean, but the write-up makes two classic MLR slips.
0 / 2 graded
person
Rookie's submission
Model: win% ? tackles + inside50s
tackles: 1.0
inside50s: 1.4
"inside50s matter 40% more than tackles,
and each tackle just adds 1% to win rate."
1 · What's wrong with 'inside50s matter 40% more than tackles'?
Each coefficient is in units of win% per unit of that predictor. A tackle and an inside-50 aren't the same 'size', so 1.4 vs 1.0 isn't a fair importance ranking — standardise first if you must compare.
2 · And 'each tackle just adds 1% to win rate' is missing what?
In MLR, the tackles coefficient is the effect among teams with the same inside50s. Drop the 'holding constant' qualifier and the claim is simply wrong.
balance
Graded — units and conditions matter.
Two habits that separate a real MLR read from a shallow one: don't rank raw coefficients, and always say 'holding the others constant'.
