The model conditions every pick on its position. A first-overall RB should produce more career starts than a first-overall WR, by construction; the baseline knows that. So when career-starts residuals drift across decades for a single position group, the drift can't be explained by changes in pick slot or aging; those are already baked in.
Plot mean residuals against draft-year buckets, faceted by detailed position group:
R/audit/te_rb_position_evolution_plot.R.What the four panels say
Tight ends are the cleanest of the four. They sit near zero in the 1980s and 2000s, dip slightly in the 2000s, and jump to +0.15 in the 2020-bucket, the highest TE residual in the forty-five-year span. The post-2017 model usage of the position (Kelce, Kittle, Andrews, McBride) is showing up as systematic over-baseline production from drafted TEs.
Wide receivers are similar but louder. They held slightly negative through the 1980s and 1990s, dipped to roughly −0.16 in the 2000s, recovered toward zero in the 2010s, and pushed to +0.09 in the 2020s. The 2000s trough is interesting: it's the run-heavy era between the receiver booms, and the residual reflects that the era's drafted receivers under-delivered relative to position-and-pick expectations.
Offensive linemen are the inverse story. The 1980s and 1990s were OL's golden era of residual: +0.21 and +0.33 mean career-starts residual respectively. From the 2000s onward the line flattens: to +0.06, then 0.00, then a modest +0.10 in the 2020s. One read is that drafting good OL in 1990 was easier than drafting good OL now, because positional commodification (interchangeable five-star OG/OT prospects on every Power 5 roster) compressed the gap between expected and observed. Another read: the y_cgs outcome we're measuring (starts on drafting team) penalizes 2010s OL more because they're the position most subject to free-agency turnover.
Running backs and fullbacks stay negative across every decade except 2020 (−0.02). The post-2010 RB devaluation story isn't a 2010s phenomenon in the residuals; RBs have been under-delivering on career-starts expectations since 1980. What shifted in the 2010s was the consensus pick value, not the residual: GMs stopped paying for them, but the picks they did take performed about as RBs always have. The 2020-bucket's near-zero residual is partly the consensus-correction (low pick capital → low expectations → easier to beat them).
Why we preserve this
The PRD draws an explicit line: position × draft_year residuals are allowed to trend. The s(draft_year) smoother in the baseline GAMs absorbs league-wide era-economic shocks (rookie wage scale, 17-game season, AV inflation) without fitting position- specific structural shifts. The shifts are real football evolution, and they're the kind of finding the project is supposed to surface rather than launder away.
The closing methodological note: there is a confounder we can't fully strip. The 2020-bucket numbers are heavily influenced by 2020- and 2021-class picks who are still in their early career; the yos covariate handles this in the model, but the residual aggregation across a partial career is sensitive to which players have started cumulating. Expect the 2020-bucket numbers to settle slightly toward the mean over the next two seasons as the cohort matures.