14b_growth_net_revised.md

14b — Growth NET of mean-reversion (R3 revision of F14, red-team T1)

14b — Growth NET of mean-reversion (R3 revision of F14, red-team T1)

Date: 2026-05-30 Supersedes the development half of 14_growth_aging.md (the Gideon roster hypotheses there are unchanged). Addresses causal red-team T1 (91_causal_redteam.md): the headline "+0.43 z/edition growth" is partly regression-to-the-mean of a negatively-selected returning cohort (returners start below the field mean), so growth and reversion are the same selected cohort seen twice and must not be double-counted. Inputs: data/processed/combine_ability.csv (255 combine rows, theta is z-scored within season), data/processed/player_season_master.csv (207 game player-seasons). Script: scripts/growth_net_revised.py. Artifact: data/processed/growth_net_revised.csv (41 debiked-theta transitions).

LEAD CAVEATS (read first)

1. Tiny, survivorship-selected samples. Development can only be measured on players who returned: 33 of 214 combine players have multi-season combine data (29×2 seasons, 4×3); 25 of 181 players have multi-season game data (24×2, 1×3). Everything below is 22–41 transitions. CIs are wide.
2. Theta is standardized WITHIN season (each season re-centered to mean ≈ 0, std ≈ 1 — confirmed: nsba1..4 means −0.00/0.00/+0.06/+0.15). So "+0.43 z/edition" is a relative-rank climb within that year's field, not an absolute skill gain on a fixed scale. There is no cross-season level trend to regress on — a per-edition trend term (edn_from) is ill-posed here and was the wrong test.
3. The development metric is combine-heavy and nsba4-heavy. 24 of 30 consecutive combine transitions are nsba3→nsba4, and nsba4 has no game data — so most of the "growth" is a combine→combine relative climb into the draft-target season, not demonstrated on-floor improvement.
4. The combine is gameable (F10). A combine-theta "rise" can be a player tanking less, not getting better.

The T1 charge, tested

Returners are negatively selected on their first combine — confirmed:

So on the combine side the red-team is right: returners start ~0.18 z below the field. With a reversion slope near −0.45, a below-mean cohort drifts upward mechanically. (On the game side there is no such selection — game returners start exactly at the field mean, p=0.91 — so the game-PPTF growth was never a reversion artifact. This matches the red-team's own note that PPTF survivorship is not a serious threat.)

The fix: net development = reversion-purged intercept

The right decomposition is not "trend net of a separately-reported reversion slope" (the two are entangled and theta has no cross-season level), but a single regression of the consecutive-edition within-player change on the prior level:

Δtheta  =  β0  +  β1 · (prior theta)

• β1 is the mean-reversion slope (negative).
• β0 is the net development — the expected gain for a returner who started at the field mean (prior = 0), i.e. with the reversion bump removed.

Results (consecutive-edition pairs, within-player)

*The naive game-PPTF mean is small because the game cohort isn't negatively selected (no reversion lift to subtract); its net is larger than its naive pooled-Δ because the pooled-Δ mixed in long 2-edition gaps.

Cluster-bootstrap 95% CI on net development (debiked theta, resample players): [+0.13, +0.68] z, mean +0.40.

Reconciliation with the "100% reversion" decomposition

Evaluating the reversion-only model at the cohort's mean starting prior (−0.16) recovers ~100% of the naive mean Δ by construction — that is just the regression passing through the centroid and says nothing about development. The decision-relevant quantity is the intercept at prior = 0, which is +0.41 and survives. So: reversion inflates the pooled number modestly (raw +0.46 → net +0.41 for debiked theta), but a real, positive net development of ≈ +0.4 z (combine) / +0.2 PPTF (game) per edition remains after purging reversion.

Within-player fixed-effects cross-check

Demeaning theta and edition within each player and regressing gives a positive trend (debiked +0.38/edition, t=4.88; raw +0.25, t=3.59; PPTF +0.07, t=1.98, p=0.054). FE agrees with the intercept method in sign and rough magnitude, confirming the net signal is not a pooling artifact. (FE on a within-season-z metric measures the same "climb the field" quantity, so do not over-read its tiny p-value — it shares the 33-player sample.)

Where the signal lives (heterogeneity, all thin)

• Game-era transitions only (nsba1/2 starts, n=6): net +0.95 (p=0.06).
• nsba3→nsba4 only (n=24): net +0.29 (p=0.06), reversion slope −0.46. Both subsets positive; neither alone clears p<0.05. The headline rests on the pooled 30.

What changed vs F14

The qualitative bottom line ("expect modest growth from real returners; don't overpay") survives — the net effect is real and positive — but the magnitude shrinks and the mechanism is now correctly separated from reversion.

Draft implications (the load-bearing part)

1. Do NOT add a growth bump on top of a reversion shrink in the buy-low board (F19) — that double-counts. The buy-low board already prices a returner's low combine as a buy by leaning on the −0.43..−0.49 reversion pull. The growth intercept (+0.41) is the gain for a returner starting at the mean; for a returner who posted a low prior, the reversion term already delivers most of their projected rise. Applying reversion-shrinkage AND a separate "+0.4 z growth bump" to the same low-combine returner counts the same climb twice. Use one model: projected = prior + β1·prior + β0 with β1≈−0.43, β0≈+0.41 — the reversion and growth terms come out of the same regression and are already net of each other.
2. Growth does NOT transfer to a brand-new nsba4 entrant with no prior. β0 is a returner effect estimated on players seen ≥2 editions. A first-time nsba4 player has no prior to revert and no within-player trajectory — you cannot credit them +0.4 z of "development." Price them off their combine theta (shrunk toward the pool mean per F10's EAP), full stop.
3. The reversion lever is the robust, actionable one (it survives in both the selection check and the slope): shrink the biggest combine scores hardest; treat a real returning player's low combine as a buy. The growth term is a smaller, second-order add that is already inside the reversion regression.
4. Trust the game-PPTF net (+0.22, p=0.028) over the combine net where a player has game tape — it is not reversion-driven (no returner selection) and is on the real-scoring scale. The combine net is on a gameable, within-season-relative metric and is nsba4-combine-heavy.

Limitations

• n = 22–41 transitions; 33/214 combine and 25/181 game multi-season players. Survivorship is severe and endogenous to returning — we cannot observe the development of the 181 players who did not return, and reweighting to the full first-season distribution (inverse-propensity on return) is not done here (the return-propensity model would itself be fit on 33 positives).
• Within-season standardization makes all combine "growth" a relative-rank quantity; if the whole field improved year-over-year, that common gain is invisible by construction. The net intercept measures relative climb only.
• nsba4 has no games, so 24/30 consecutive combine transitions cannot be validated against on-floor production; the game-PPTF leg (nsba1↔nsba2↔nsba3) is the only one anchored to real play.
• The combine is gameable; a theta rise can be reduced tanking, not skill.
• Net development is positive but its CI ([+0.13, +0.68] z) is wide enough that the practical instruction is "small positive, don't lean hard," not a precise number.

Reproduce

/home/david/code/nsba/.venv/bin/python scripts/growth_net_revised.py

Prints the survivorship counts, selection table, naive/reversion/net models for debiked theta, raw theta, and game PPTF, the within-player FE cross-checks, the decomposition, and writes data/processed/growth_net_revised.csv.

NSBA Draft Analytics · embargoed until after the SSB draft · ← hub