Do AP Voters Favor the Blue Bloods?

The blue bloods

Alabama, Ohio State, Michigan, Notre Dame, USC, Texas, Oklahoma, Georgia, LSU, Penn State, Florida State, Tennessee, Florida, Ole Miss, Oregon. The list was fixed before any result was looked at.

Against equally strong teams

Most blue bloods were also among the best teams of these seasons, and the AP poll rewards the very best far more than a straight line would: the points for #1 dwarf those for #25. So every question compares a blue blood with teams in the same strength tier by the v4-state-space ratings, not with the average team. One AP spot is about 61 points.

The second bar in each card is a check: the same question for the 15 strongest teams outside the list, who have the strength but not the name. If the method is fair, it finds nothing for them. Lines are 95% intervals; a bar is highlighted when its interval clears zero.

Bias, or do voters know something?

A higher ranking is only bias if the team doesn't live up to it. Our ratings can't settle that - every group of teams that turned out good beat them, name or not - so this asks the betting market, the best estimate there is. Bars are points of a game's margin.

Team by team

Each blue blood's average AP points a week beyond what its rating, tier, record, scoring and schedule explain. A description of who got the credit, not a test - no intervals here.

Does it hold up?

The same questions with only the classic 7 (Alabama, Ohio State, Michigan, Notre Dame, USC, Texas, Oklahoma), and with a different rating model measuring strength. Highlighted when the interval clears zero.

How this was done

Enough to recreate every number on this page. Each result's own method line above lists its rows, outcome and controls.

• Data: AP Top 25 ballots, game results and Vegas closing lines from CollegeFootballData (CFBD) for the 2022-2025 regular seasons, played games between two FBS teams only (pulled with games through 2026-10-10). Team strength is NovaModel's v4-state-space rating, built only from games before each week and carrying part of last season in, so remembering last season isn't counted as bias.
• Effects: every number is the blue-blood coefficient from an ordinary least squares (OLS) regression with an intercept - AP points for the three poll questions, points of a game's margin for the two Vegas ones. Each card's method line lists its outcome and every control.
• 95% confidence intervals: percentile cluster bootstrap. Teams (for games, the home team) are drawn with replacement 2,000 times, keeping every row of each drawn team; the regression is refitted on each draw, and the interval runs from the 2.5th to the 97.5th percentile of the coefficient.
• p-values: permutation test. The blue-blood label is reshuffled among the teams in the data 2,000 times, keeping how many teams carry it, and the regression refitted each time; p = (shuffles whose coefficient is at least as far from zero as the real one, plus 1) / (shuffles plus 1), two-sided. A random set of teams isn't an equally strong one, so for an elite group this runs optimistic - judge a result by its interval first.
• Multiple comparisons: Benjamini-Hochberg false-discovery-rate adjustment (q-values) across the four main questions - preseason, weekly, after a loss, against the line. A result is called real only when its 95% interval excludes zero and its q is under 0.05.
• Checks, reported but never part of the verdict or the adjustment: the placebo; the classic 7 (Alabama, Ohio State, Michigan, Notre Dame, USC, Texas, Oklahoma) with the other blue bloods left out; and v2-iterative-adjusted ratings in place of v4-state-space for the three poll questions.
• Poll spots: the average AP-points gap between #1 and #25, divided by 24, over every AP Top 25 release in the data (60.6 points a spot).
• Random seed 42, for both the bootstrap and the shuffles. The limit: brand and strength overlap almost completely in these seasons, and the tiers and the placebo are as far as the data goes in separating them.