Every Claim, and What Backs It

By Justin Pietsch · August 10, 2026

Generated by code. Do not edit by hand.

Every claim below is one a published page makes and one a test asserts. There are 110 of them, grouped by the page that says each one first, in reading order. A claim said on more than one page appears once, under the first, with the others named beneath it.

The analysis computes 115 guarded claims in all. The 5 not listed here are cited only by working documents that do not publish, or by nothing at all; a row for a claim no reader can meet is disclosure with no audit purpose.

A test re-checks every row on this page against the number behind it on every test run, so a claim whose support moves out from under it fails the commit that moved it.

Each of these claims could have come out the other way. All 115 of this question’s guarded claims are observed failing at least once against a constructed input, somewhere in its test suite, so none of them is a cut that nothing could fail. A check re-verifies this on every commit that touches the guard set, so it is a live property of the code rather than a measurement someone took once. It establishes that each cut can fail, not that the input which failed it was a realistic one.

A method is described for 110 of the 110 claims. That field is written beside the computation it describes and fills in claim by claim, so this count is here rather than a promise that it is complete.

Every claim here names the tests it rests on, declared beside the computation that runs them rather than read off the prose around them. 14 of the 110 rest on at least one, and the row names the correction each of those tests entered, or says that none is declared for it. Everything named on a row is set out under “The test families” at the foot of this page.

93 name no p-value, so no multiple-comparison family applies to them. That is an answer rather than a gap, and it does not mean the claim is a mere description: a count and a direct comparison land there, and so does a judgement read off an interval instead of a test statistic. Which of those a given claim is, the method line beneath it says.

3 have not been attributed yet. That is the number this page is waiting on, and it is printed rather than smoothed over because an unattributed claim is one you cannot tell a corrected test from an uncorrected one behind.

Why NBA Home Court Advantage Is Disappearing

The weaker team at home matched or exceeded the stronger team’s home win rate.

• Evidence: 1984–94: weaker 65% vs stronger 71%; 1995–01: weaker 66% vs stronger 60%
• Also said in: Home Court Advantage Is Fading, The Playoffs Are Different (and the Same), The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions
• Method: per-era comparison of the lower-seed and higher-seed playoff home win rates across the two earliest eras
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Rebounding is the largest single driver of the decline.

• Evidence: largest is rebounding (30%), largest in 76% of season-block resamples
• Also said in: Home Court Advantage Is Fading, The Three-Point Suspect, The Mystery on the Glass, How I Know This Isn’t Made Up, The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: ranking the four box-score channels by their share of the regular-season decline in the linear-probability mediation (each channel’s win coefficient times its own year-trend, over the total trend), with the season-block bootstrap frequency of that ranking carried alongside
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The three-point shift reaches close to half the decline overall.

• Evidence: 3PA reach ~ 45% of the decline
• Also said in: Home Court Advantage Is Fading, The Whistle, The Three-Point Suspect, The Mystery on the Glass, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: the shooting share of the decline plus the three-point-absorbed fractions of the foul and turnover shares, composited from the same 3PA-controlled channel trends and thresholded to a 40-to-50 percent band
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The regular-season net-rating advantage shrank by more than a third.

• Evidence: 3.13 → 1.97 = 37% drop
• Also said in: Home Court Advantage: Frequently Asked Questions
• Method: mean home net rating per 100 possessions in the 2005-17 era against the 2023-26 era, read as a proportional drop from the earlier level
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

After removing the year trend, only the 1994-95 (1995-01) boundary shows a significant one-time level shift.

• Evidence: 1995–01:p=0.010; 2002–04:p=0.988; 2005–17:p=0.722; 2018–22:p=0.990; 2023–26:p=0.836
• Also said in: Home Court Advantage Is Fading, The Usual Suspects, How I Know This Isn’t Made Up
• Method: a game-level logistic on the year trend plus rule-change era dummies, counting which era boundaries clear p < 0.05 once the trend is partialled out; the 1995-01 dummy is the published one
• Family: standalone (pre-specified level-shift test at the known 1994-95 hand-checking rule change)

The last game of a series is nearly as friendly to the host as the first: Game 7 sits closer to Game 1 than to the games the lower seed hosts, and the across-game trend is not significant, so road teams show no sign of adapting as a series wears on.

• Evidence: G1 69%, G7 64%, lower-host avg 55%; trend p=0.52
• Also said in: The Playoffs Are Different (and the Same), Home Court Advantage: Frequently Asked Questions
• Method: distance comparison of the game-7 home win rate against the game-1 rate and against the mean of the three lower-seed-hosted games, paired with a count-weighted least-squares trend across game numbers failing to reach significance
• Family: series_game_number (m=2, BH FDR q=0.05)

The top seed’s home win rate fell far less than the weaker seed’s: the playoff collapse is the underdog’s, not the favorite’s.

• Evidence: higher-seed drop 1 pp vs lower-seed drop 17 pp
• Also said in: Home Court Advantage Is Fading, The Playoffs Are Different (and the Same)
• Method: drop in each seed tier playoff home win rate from the mean of the two earliest eras to the most recent one, comparing the higher-seed drop against half the lower-seed drop
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The weaker seed’s home win rate collapsed from the mid-60s in the earliest two eras to below 50% today.

• Evidence: 1984–94: 65%; 1995–01: 66%; 2023–26: 49%
• Also said in: Home Court Advantage Is Fading, The Playoffs Are Different (and the Same), The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: level comparison of the lower-seed playoff home win rate in the two earliest eras against the most recent era
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Barely better than a coin flip.

• Evidence: 51.9%
• Also said in: The Playoffs Are Different (and the Same), Home Court Advantage: Frequently Asked Questions
• Method: monte carlo of best-of-seven 2-2-1-1-1 series between two otherwise-equal teams, driven by the latest era observed single-game home win rate, read against a band just above an even split
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Shooting is the largest piece of the home advantage, more than 40%.

• Evidence: largest is shooting (43%)
• Also said in: Home Court Advantage Is Fading, The Mystery on the Glass
• Method: ranking the four channels by their share of the regular-season home-advantage LEVEL in the linear-probability decomposition (each channel’s win coefficient times its mean home-minus-away differential), and thresholding the top share at 40 percent
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The turnover fade nearly matches rebounding’s share of the decline.

• Evidence: TOV 27% vs REB 30%
• Also said in: The Mystery on the Glass, How I Know This Isn’t Made Up, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: the gap between two of the four decline shares in the linear-probability mediation, thresholded at 8 percentage points
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

With the game’s 3PA rate in the model, the home shooting (eFG%) downward trend vanishes entirely (fully absorbed by the three-point shift).

• Evidence: eFG trend 210% absorbed by the 3PA control
• Also said in: The Three-Point Suspect, How I Know This Isn’t Made Up, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: the shooting differential’s year-trend fit with and without the game’s three-point-attempt rate as a control, checking the control absorbs the whole slope (an absorbed share at or above 100 percent, meaning the trend reverses sign)
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Of the decline’s two traceable real-world forces, the three-point shift reaches further than the genuine officiating change.

• Evidence: 3PA reach 45% vs whistle share 9% of the decline
• Also said in: The Three-Point Suspect, Home Court Advantage: Frequently Asked Questions
• Method: the composited three-point reach compared with the un-absorbed share of the foul fade, both built from the same 3PA-controlled channel trends; a comparison of two magnitudes, with no test of whether the difference is real
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The three-point shift explains about half of both the foul and turnover declines.

• Evidence: foul 51% / turnover 54% absorbed by 3PA
• Also said in: Home Court Advantage Is Fading, The Whistle, The Three-Point Suspect, The Mystery on the Glass, The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: each channel’s home-differential year-trend re-fit with the game’s three-point-attempt rate added as a control, cluster-robust by season; the absorbed share is one minus the controlled slope over the raw slope, thresholded to a 40-to-65 percent band for the foul and turnover channels
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Rebounding is the surest of the four channels: the least absorbed by the three-point control.

• Evidence: rebounding 8% absorbed, the minimum of the four channels
• Also said in: The Three-Point Suspect, The Mystery on the Glass, Home Court Advantage: Frequently Asked Questions
• Method: ranking the four channels by how much of their year-trend the three-point-rate control absorbs, and thresholding the rebounding figure at 20 percent. A ranking of absorption, not a test of the controlled rebounding trend
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

In the playoffs the frozen four-factor model beats both a flat guess and the extrapolated trend.

• Evidence: playoff RMSE: channel 3.87 vs trend 7.30 vs flat 8.11 pp
• Also said in: How I Know This Isn’t Made Up
• Method: held-out root-mean-square error of the frozen four-channel playoff model against two baselines fit on the same training window, an extrapolated trend line and a flat training mean; three error numbers ordered, with no test on the differences
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Predicts each later season’s home win rate to within about a point.

• Evidence: largest held-out miss 1.6 pp
• Also said in: How I Know This Isn’t Made Up
• Method: a four-channel linear-probability win model frozen on the training seasons and used to predict each held-out season’s home win rate from that season’s own box-score edges, thresholding the largest miss at 1.6 points
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The raw season-level parity link is null, and a small detrended one appears that runs the opposite way to the usual hunch.

• Evidence: raw r=-0.092 (p=0.556); detrended r=-0.345 (p=0.025)
• Also said in: The Usual Suspects, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: season-level pearson correlation of the league win-percentage spread against the home win rate, taken raw and again after removing a year trend from each series, with the sign of the detrended correlation read as well as its p-value; see “Competitive Balance” in questions/home_court/docs/home_court_investigation.md
• Family: parity_hca_specs (m=3, BH FDR q=0.05)

The travel effect is far too small to bend the trend, and it leans slightly against the home team rather than for it.

• Evidence: -0.075 pp per 100 miles = 1.87 pp over a 2,500-mile trip
• Also said in: The Three-Point Suspect, The Usual Suspects, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: game-level logistic of the home win on the visiting team great-circle travel distance, season-clustered standard errors, converted to percentage points of home win rate per 100 miles and scaled to a 2,500-mile trip; see “Travel and Time Zones” in questions/home_court/docs/home_court_investigation.md
• Family: primary_tests (m=14, BH FDR q=0.05)

The rest advantage has not shifted across eras, so it cannot drive the decline.

• Evidence: rest × era interaction LR p = 0.474 (regular season)
• Also said in: The Usual Suspects, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: likelihood-ratio test of a rest-differential by era interaction against the additive rest plus era logistic, regular season; see “Rest and Altitude” in questions/home_court/docs/home_court_investigation.md
• Family: rest_era_interaction (m=2, BH FDR q=0.05)

In the recent seasons when home court sat lowest, arenas stayed near their fullest on record.

• Evidence: recent-3 mean 18,249 vs series max 18,384/game
• Also said in: Home Court Advantage Is Fading, The Usual Suspects, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: mean league attendance per game over the three most recent seasons with gate data, against the maximum of the whole season series; see “Crowd Size” in questions/home_court/docs/home_court_investigation.md
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Beyond the presence of a crowd at all, a bigger crowd shows no detectable effect on the home win rate in the one season where crowd size varied.

• Evidence: per-1,000-fans effect +0.51 pp, p = 0.18, median crowd 3280
• Also said in: The Three-Point Suspect, The Usual Suspects, Home Court Advantage: Frequently Asked Questions
• Method: game-level logistic of the home win on attendance in thousands, over the 2020-21 season, the one season in which local health rules made crowd size vary within a single year
• Family: standalone (the single test of the one natural experiment this question has: 2020-21 is the only season in which crowd size varied within a year, and this is a game-level dose-response fitted on that season alone. The four season-level specifications corrected together in attendance_hca_specs ask a different question, on different data and at a different grain, so this test joins no family and has no companions of its own to be corrected against)

Denver and Utah still hold the two largest home-court advantages in the recent era, not just across the whole record.

• Evidence: Denver Nuggets (+22.5), Utah Jazz (+22.9)
• Also said in: Home Court Advantage Is Fading, The Playoffs Are Different (and the Same), The Altitude Holdouts, The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions
• Method: ranking of the qualifying franchises by their recent-era home-minus-road win rate alone, taking the top two, so no era mixing enters the ordering
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Playoffs: the win-margin spread (Q90−Q10) widens over time, with the whole 95% CI above zero.

• Evidence: Q90−Q10 spread slope +0.35 pts/yr, 95% CI [+0.23, +0.47] (season-cluster bootstrap)
• Also said in: The Altitude Holdouts, The Disappearing Home-Court Advantage: A Series, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: season-cluster bootstrap of the gap between the dither-averaged q90 and q10 unconditional quantile slopes of home margin on year, resampling whole seasons with replacement. The claim is inferential and rests on the whole 95 percent percentile interval sitting above zero; no p-value is computed for this quantity, which is why the tests list is empty
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Regular season: the win-margin spread (Q90−Q10) widens over time, with the whole 95% CI above zero.

• Evidence: Q90−Q10 spread slope +0.23 pts/yr, 95% CI [+0.15, +0.31] (season-cluster bootstrap)
• Also said in: The Altitude Holdouts, The Disappearing Home-Court Advantage: A Series, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: season-cluster bootstrap of the gap between the dither-averaged q90 and q10 unconditional quantile slopes of home margin on year, resampling whole seasons with replacement. The claim is inferential and rests on the whole 95 percent percentile interval sitting above zero; no p-value is computed for this quantity, which is why the tests list is empty
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Home Court Advantage Is Fading

The playoff home-court decline is genuine home-court weakening, not the top seeds and their opponents bunching closer in quality.

• Evidence: year trend 102% retained after quality control (-2% absorbed by quality)
• Also said in: The Playoffs Are Different (and the Same)
• Method: logistic fits of the playoff home win on year alone and on year plus the same-season seed-quality gap, comparing how much of the year coefficient survives the control
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The regular-season quality effect on who wins at home shrank over time (p<0.001), the opposite direction from the playoff seeding gap, which widened.

• Evidence: omnibus era LR p=1.304e-12; direction coef=-0.02810, p=5.909e-12; gap (stronger/weaker hosts) 1984–94=81.6%/46.7%, 2023–26=69.5%/41.3%
• Also said in: The Playoffs Are Different (and the Same), The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions
• Method: logistic of the regular-season home win on the prior-season quality gap interacted with year, signing that continuous interaction and thresholding its p-value, alongside an omnibus likelihood-ratio test of the same gap interacted with era
• Family: rsquality_era_interaction (m=2, BH FDR q=0.05)

Home court advantage has not vanished: home teams still win clearly more than half their regular-season games, though far less often than at the start of the data.

• Evidence: last-5-season home win 55.3% vs trend start 65.6%
• Also said in: Home Court Advantage: Frequently Asked Questions
• Method: two levels of the season home-win series compared: the mean over the last five regular seasons against the fitted trend value at the first season, thresholded at 52 percent and a five-point fall
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The record favors more than one bend over a single one.

• Evidence: P(k=1) 19.3% vs P(k≥2) 79.3%
• Also said in: How I Know This Isn’t Made Up, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: a Bayesian change-point model over the season home-win series, comparing the posterior mass on two or three breaks against the mass on exactly one; posterior probabilities, not a p-value
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Prose: the decline settled to about a quarter of a point per year after the bend (article 1, investigation).

• Evidence: QLR post-break slope -0.255 pp/yr
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: break test on the season home-win series (supremum Chow F over every candidate year, outer 15 percent trimmed), reading the fitted subperiod slope AFTER the selected break year and thresholding it to a band around 0.25
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Prose: the decline ran at about 0.6 points per year before the late-1990s bend (article 1, investigation).

• Evidence: QLR pre-break slope -0.652 pp/yr
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: break test on the season home-win series (supremum Chow F over every candidate year, outer 15 percent trimmed), reading the fitted subperiod slope BEFORE the selected break year and thresholding it to a band around 0.6
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Every franchise with a long enough record shows a declining home-court edge.

• Evidence: 0/31 franchises have a rising raw slope
• Also said in: How I Know This Isn’t Made Up, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: a per-franchise ordinary-least-squares year slope of the regular-season home-minus-road win-rate gap, fit for every franchise with at least ten seasons; a count of how many of those raw slopes come out rising
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The decline is a uniform league-wide drift, not concentrated in particular franchises: every team faded at roughly the same rate.

• Evidence: true between-franchise SD 0.000 vs pooled slope 0.490 pp/yr, 100% noise
• Also said in: The Usual Suspects, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: a method-of-moments split of the observed spread in per-franchise slopes into true between-team variance and sampling noise; the claim fails only when the true spread is large relative to the pooled slope AND most of the observed spread is not noise
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Prose: every franchise faded at about half a point of the home-road gap per year (article 5).

• Evidence: pooled league-wide slope -0.490 pp/yr
• Also said in: The Usual Suspects
• Method: the pooled cluster-robust year slope of the team-season home-minus-road win-rate gap panel, thresholded to a band around half a point per year
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

In the most recent regular season, exactly one team won more on the road than at home.

• Evidence: 1 team(s) with road>home: Charlotte Hornets
• Method: count of teams in the most recent regular season whose road win rate exceeds their own home win rate
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The weakest teams pick up most of their wins at home, while the best teams win nearly as often on the road.

• Evidence: home share of total wins (2026 reg. season): bottom quartile 61% vs top quartile 53%
• Method: quartiles of the most recent season teams by overall win rate, comparing the home share of total wins in the bottom quartile against the top quartile; a share, not a per-team home-court difference
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The four box-score channels capture nearly all of the home advantage level in both contexts (at least 90%).

• Evidence: level share RS 95% / PO 93%
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: the share of the home-advantage level the four channels account for under the linear-probability level identity (everything except the model intercept), read in the regular season and the playoffs and thresholded at 90 percent in both
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

How much each factor counts toward winning stayed nearly constant decade to decade.

• Evidence: efg [+3.28, +3.61]; fouls [-2.20, -1.63]; tov [-3.57, -3.19]; reb [+1.57, +1.70]
• Method: the four-channel linear-probability model re-fit inside each rule-change era, checking that every channel keeps its sign in every era and that its era-to-era range stays within half its mean magnitude
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The home foul and free-throw-attempt bias collapsed from the 1980s to today.

• Evidence: foul gap 1.23 → 0.25, FTA gap +1.97 → +0.46
• Also said in: The Three-Point Suspect, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: era means of the home-minus-away foul and free-throw-attempt gaps, regular season; each recent-era gap compared against 40% of its earliest-era size
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Rebounding and turnovers together outweigh the more famous suspects (shooting and fouls) in the decline.

• Evidence: REB+TOV 57% vs eFG+fouls 39%
• Also said in: The Mystery on the Glass
• Method: the rebounding and turnover shares of the regular-season decline summed against the shooting and foul shares summed, both read off the same linear-probability trend decomposition
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The series-level edge and the per-game edge gave up the same share of themselves, so the format shrank the edge rather than sheltering it from the decline.

• Evidence: per-game kept 37.4% of its earliest-era edge, series kept 38.2%
• Also said in: The Playoffs Are Different (and the Same), The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: share of its earliest-era home edge that the latest era still holds, computed once at the per-game level and once at the simulated series level, compared as a difference of those two shares
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

A best-of-7 leaves the home-court team about a third of its per-game edge, in every era.

• Evidence: series-to-per-game edge ratio by era: 0.33, 0.33, 0.33, 0.33, 0.33, 0.34
• Also said in: The Playoffs Are Different (and the Same), The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: ratio of each era simulated series-level home edge to its per-game home edge, both measured from an even split, checked as a band across every era rather than at the endpoints
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The Whistle

Nearly every official with a qualifying playoff record called fewer fouls on the home team.

• Evidence: 45/47 home-favouring
• Method: count of playoff officials with at least 50 games whose career home-minus-away foul gap is negative, as a share of all such officials
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Referees called fewer fouls on the home team in every era, in both the regular season and the playoffs.

• Evidence: 12 era rows, all negative
• Also said in: The Disappearing Home-Court Advantage: A Series, Home Court Advantage: Frequently Asked Questions
• Method: sign check on every era-by-context cell of the home-minus-away foul differential table, regular season and playoffs
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

After adjusting for games worked, the most home-leaning officials sit roughly a full foul per game apart from the most even-handed.

• Evidence: shrunken spread 1.02 fouls/game
• Method: empirical-Bayes shrinkage of each official career mean foul gap toward the league mean, weighted by how noisy that official record is, then the range across the shrunken means
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

About 60% of the raw spread across officials is sampling noise (random bounce).

• Evidence: noise share 60%
• Method: method-of-moments split of the observed between-official variance in the career home-minus-away foul gap into a true component and a within-official sampling component, over playoff officials with at least 50 games
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The raw spread across officials is about the same size as the residual league-wide foul advantage that is left.

• Evidence: SD 0.64 vs residual gap 0.68 = 0.9x
• Method: observed between-official standard deviation of the career foul gap, divided by the recent-era playoff league foul gap the differentials section reports
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The 1995–01 per-official era row is built on a small fraction of the era’s playoff games, not on the era.

• Evidence: 36/496 games = 7%
• Method: share of the era playoff games whose box scores carry officials at all, from the per-era coverage counts, against a 25% ceiling
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The 1980s home free-throw edge was worth about a point and a half of scoring per game.

• Evidence: FTM diff, earliest era: +1.58 pts/game
• Method: era mean of the home-minus-away made free throws per game, earliest ERA_DEFS era, regular season
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The average home foul/FTA edge is small next to the night-to-night swing; on any single night random variation dwarfs it.

• Evidence: fta (early): swing 14x the edge; fta (late): swing 50x the edge; foul (early): swing 11x the edge; foul (late): swing 49x the edge
• Method: ratio of the P90-minus-P10 night-to-night width to the absolute mean edge, per era and per column, regular season
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The average foul/FTA edge faded far more than the night-to-night spread narrowed (the spread stays about as wide; only its center moved).

• Evidence: fta: edge fell 77%, spread narrowed 18%; foul: edge fell 80%, spread narrowed 14%
• Method: per-game P10/P90 spread of the foul and free-throw-attempt gaps by era, regular season; the percentage fall in the mean edge compared against the percentage narrowing of the P90-minus-P10 width
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

In close games (decided by <=4 points) the home team’s edge faded from a few points above a coin flip to about even.

• Evidence: <=4: 54.9% -> 50.1%; <=2: 53.2% -> 50.6%; all: 64.9% -> 55.6%
• Method: era means of the regular-season home win rate restricted to games decided by four points or fewer, earliest era against the most recent
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Most officials show a home-foul tilt that survives the correction for testing every official at once, so the differences between them are real and not only noise.

• Evidence: 29/47 significant after BH
• Method: one two-sided t-test per playoff official on that official career home-minus-away foul gap, using the per-game standard deviation, then Benjamini-Hochberg FDR at q=0.05 across all officials tested together; the survivor count and its share of officials are what the guard thresholds

The Three-Point Suspect

The home turnover edge shrank partly as a road-specific effect: the road turnover penalty fell from about 0.8 to 0.1 per 100 as road turnovers fell more than home.

• Evidence: gap 0.85→0.13; home Δ-2.81, away Δ-3.53
• Also said in: The Mystery on the Glass
• Method: league turnovers per 100 possessions split by venue and averaged over the first and the last five seasons of the turnover era, comparing the change in the away-minus-home penalty against the change on each venue separately
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

A season’s three-point rate does not predict the next season’s home court beyond home court’s own past.

• Evidence: 3PA→HCA Granger p = 0.23, 0.27
• Also said in: How I Know This Isn’t Made Up, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: granger causality F-tests at one and two lags, from the season three-point rate to the home win rate, on first differences when an augmented Dickey-Fuller test calls both series I(1)
• Family: granger_3pa_lead (m=4, BH FDR q=0.05)

No long-run tie between the 3PA rate and home win % is detectable, so the season-level correlation between them is not evidence on its own.

• Evidence: Engle-Granger p = 0.766
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: engle-granger cointegration test on the season-level three-point-rate and home-win-rate series, run only when an augmented Dickey-Fuller test calls both series I(1)
• Family: standalone (spuriousness diagnostic rather than a hypothesis about home court: it asks whether the season-level three-point rate and home win rate share a long-run relationship, which is what decides whether the correlation between them can be read at all, and it runs only when an augmented Dickey-Fuller test calls both series I(1))

Judged against expected pre-tipoff pace rather than the pace the game actually ran at, the within-era pace-HCA link no longer clears the bar in either specification, the within-era one only just short: part of the raw link is the game’s own script (reverse causation).

• Evidence: expected-pace p: bivariate 0.227, within-era 0.063
• Also said in: The Usual Suspects, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: game-level logistic of the home win on leave-one-out expected pre-tipoff pace, fitted bivariate and again with era fixed effects, season-clustered standard errors, regular season; see “Pace of Play” in questions/home_court/docs/home_court_investigation.md
• Families: primary_tests (m=14, BH FDR q=0.05), standalone (the second specification of a hypothesis whose first is corrected inside primary_tests, which is where the era-controlled expected-pace fit runs and is counted. A fact can enter only one family in this census, so the bivariate is declared here rather than counted twice. Both specifications ran, both are published, and neither reaches significance)

With arenas empty, home court advantage nearly vanished (within 2 pp of a coin flip).

• Evidence: empty-arena home win 51.0%
• Also said in: The Usual Suspects, Home Court Advantage: Frequently Asked Questions
• Method: home win rate over the 2020-21 games played in fully empty arenas, against a 50% coin flip
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The Mystery on the Glass

League raw field-goal percentage is no higher than it was in the first season of the record, so the efficiency gain the league did make is in points per shot rather than in the share of shots that go in.

• Evidence: FG% 49.20 → 47.10
• Also said in: The Disappearing Home-Court Advantage: A Series
• Method: league FGM / FGA over every regular-season team-game row in the first and last cached season, raw rather than effective or true-shooting
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

More field goals miss per team per game now than in the first season of the record, so the pool of rebounds available to anyone grew rather than shrank.

• Evidence: missed FG/team/game 45.02 → 47.13
• Also said in: The Disappearing Home-Court Advantage: A Series
• Method: league (FGA - FGM) divided by the team-game count, over every regular-season team-game row in the first and last cached season; free throws excluded (only the last of a set is rebounded)
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The league-wide OREB rate bottomed out mid-record and has risen over the last several seasons, rather than falling monotonically throughout.

• Evidence: trough 22.1% (season 38 of 43), recent 25.9%
• Method: trough location in the season-level league offensive-rebound rate series, searched only within the last twelve seasons and required to sit outside the final two, then the rise from that trough to the last season
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

No single tracking edge’s trend is individually established once the three tested together are corrected for; the OREB-conversion trend is the strongest of the three but not on its own conclusive.

• Evidence: OREB conversion edge (pp) p=0.024, Box-out edge (per game) p=0.775, 2nd-chance pts edge (per game) p=0.304 (0/3 survive BH)
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: ordinary least squares year trend per player-tracking rebounding edge over the tracking-era seasons, then Benjamini-Hochberg FDR at q=0.05 across the three trends tested together, counting how many survive
• Family: tracking_era_trends (m=3, BH FDR q=0.05)

The tracking offensive-rebound edge fell from ~1.2-1.3 in the mid-2010s to under 0.2 today.

• Evidence: 1.2 -> 0.2
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: fitted endpoints of an ordinary least squares year trend on the home offensive-rebound conversion edge, first tracking season against the last (fitted rather than raw, because a short series makes the raw endpoints noisy)
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The playoffs run on about a fifteenth as many games as the regular season.

• Evidence: 49,107 regular-season / 3,292 playoff = 14.9x
• Also said in: How I Know This Isn’t Made Up, Home Court Advantage: Frequently Asked Questions
• Method: ratio of the regular-season to the playoff game count in the assembled dataset, thresholded to a band around fifteen
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The playoff rebound share edge has fallen by roughly three-quarters.

• Evidence: +2.74 → +0.70 (74% drop)
• Method: era means of the pace-free rebound share edge, playoffs, as a fractional drop from the earliest era to the most recent
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The playoff turnover channel’s trend is too uncertain to call in either direction.

• Evidence: decline share CI [-13, 39]%
• Method: a season-block bootstrap 95 percent interval on the playoff turnover share of the decline, checked for straddling zero. The test is an interval rather than a p-value, so no p-value fact stands behind it
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The Usual Suspects

The home-away three-point-attempt gap has grown, and home teams now take slightly more threes than visitors.

• Evidence: 3PA-rate diff -0.35 → +0.44 pp of FGA
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: era means of the home-minus-away three-point attempt rate (share of field-goal attempts), earliest era against the most recent, regular season; see “Home vs. Away Three-Point Differential” in questions/home_court/docs/home_court_investigation.md
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

In the combined situational model, roughly half the explanatory power belongs to the situational factors combined and the other half to the era of the game.

• Evidence: Shapley: era 53% vs rest+altitude+tz+COVID 47%
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: a Shapley decomposition of the McFadden pseudo-R-squared over five predictor blocks (era, rest, altitude, time zone, COVID), averaging each block’s marginal gain across all 32 block-subset logits, then thresholding both the era share and the four situational blocks combined to a 40-to-60 percent band
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The move into the most recent playoff format period is the sharpest period-to-period fall in playoff home win % of any format change, which is what puts the 2014 change under suspicion in the first place.

• Evidence: into 2014–26: -6.8 pp; all transitions: 1985–02 +1.6, 2003–13 +0.1, 2014–26 -6.8
• Method: differences in raw playoff home win % between consecutive format periods, with the most recent transition checked against the minimum of all of them; the period whose transition must win is fixed as the last one rather than taken as an argmin
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

After controlling for the year trend, the 2014 format-period dummy adds no significant level shift; the post-2014 drop is the trend passing through.

• Evidence: 2014–26 dummy p = 0.298
• Also said in: The Playoffs Are Different (and the Same), Home Court Advantage: Frequently Asked Questions
• Method: logistic of the playoff home win on year plus playoff format-period dummies, thresholding the most recent format dummy p-value at 0.05

The four box-score channels account for nearly all of the regular-season decline, and for materially less of the playoff one, so a larger share of the playoff decline sits outside them.

• Evidence: decline share RS 96% / PO 67%
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: the share of the home-advantage era-to-era decline the four channels account for under the same linear-probability decomposition, read in each context and thresholded at 90 percent in the regular season and 80 percent in the playoffs, on the point estimates rather than on their bootstrap intervals
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The Playoffs Are Different (and the Same)

Through the 2005–17 era the weaker seed hosting won more often than not in every era; only the two most recent eras slip below the coin flip.

• Evidence: 1984–94: 65%; 1995–01: 66%; 2002–04: 52%; 2005–17: 51%; 2018–22: 47%; 2023–26: 49%
• Method: per-era comparison of the lower-seed home win rate against an even split, run over the first four eras and the two most recent ones separately
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The arena has mattered more to a typical playoff game’s outcome than the quality gap between the two teams.

• Evidence: arena +14.0 pp vs. typical quality swing +11.9 pp
• Method: bivariate logistic of the playoff home win on the seed-quality gap, comparing its fitted equal-quality home win edge against the same fit slope scaled by the observed mean absolute quality gap
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The home team wins more than half of the games at every game number in a series, the lower seed’s included.

• Evidence: lowest is G3 at 55.0%; G1 69.4%; G2 71.9%; G3 55.0%; G4 55.3%; G5 74.5%; G6 55.5%; G7 63.8%
• Method: minimum over the seven per-game home win rates, each a raw share of home wins among the playoff games played at that game number, read against an even split
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Game 5 is the most home-lopsided game of the series.

• Evidence: G1 69.4%; G2 71.9%; G3 55.0%; G4 55.3%; G5 74.5%; G6 55.5%; G7 63.8%
• Method: argmax over the seven per-game home win rates, each a raw share of home wins among the playoff games played at that game number
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

From the early 2000s onward the lower seed’s home rate sits at 47–52% while the higher seed’s held near 70–75% through 2022, easing to about 65% in recent seasons.

• Evidence: 1984–94: H 71%/L 65%; 1995–01: H 60%/L 66%; 2002–04: H 74%/L 52%; 2005–17: H 75%/L 51%; 2018–22: H 72%/L 47%; 2023–26: H 65%/L 49%
• Method: range check on the per-era higher-seed and lower-seed playoff home win rates, over the eras from 2002-04 onward and separately on the most recent one
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The playoff seeding gap (higher-minus-lower seed home win rate) widened sharply from the earliest era to its peak, and is still far wider today.

• Evidence: 1984–94 gap +5.6 pp, peak gap +24 pp, 2023–26 gap +16 pp
• Also said in: The Disappearing Home-Court Advantage: A Series
• Method: per-era subtraction of the lower-seed home win rate from the higher-seed rate, comparing the earliest era against the largest era gap and against the most recent era
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The playoff and regular-season series advantages are not far enough apart to separate on this many games: their 95% intervals overlap, so the playoffs’ higher point estimate is not a measured difference.

• Evidence: playoff 52.5% [50.9, 54.2] vs regular 51.9% [51.4, 52.3]
• Also said in: Home Court Advantage: Frequently Asked Questions
• Method: overlap of the two 95% bootstrap bands for the latest-era simulated series home win rate, one driven by observed playoff per-game rates and one by regular-season rates
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The genuine between-franchise spread in home-court advantage seen in the regular season shrinks to essentially zero in the playoffs.

• Evidence: true between-franchise SD: RS 4.1 pp vs PO 0.0 pp
• Also said in: The Altitude Holdouts, Home Court Advantage: Frequently Asked Questions, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: method-of-moments split of the observed franchise-to-franchise home-court variance into binomial sampling noise and true between-franchise variance, run separately in each context and compared as standard deviations
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Most franchises have fewer than 150 playoff home games on record.

• Evidence: 29/32 franchises under 150
• Also said in: The Altitude Holdouts
• Method: count of franchises whose playoff home-game total falls under 150, against half the franchise table
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Franchise HCA, averaged across the franchises with both records, is worth about 20 pp in the regular season and about 27 in the playoffs, roughly a 7-point premium.

• Evidence: RS +19.6 pp, PO +26.9 pp, premium +7.2 pp
• Also said in: The Altitude Holdouts
• Method: mean home-minus-road win rate across the franchises carrying both a regular-season and a playoff record, taken in each context and as their difference, with band checks on all three
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Playoffs: both margin tails move apart, Q10 (big home losses) falls and Q90 (big home wins) rises, each surviving BH correction across all ten quantile fits.

• Evidence: Q10 -0.121 (BH p=0.020), Q90 +0.228 (BH p=0.000)
• Also said in: The Altitude Holdouts, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: dither-averaged unconditional quantile slopes of home margin on year at q10 and q90, each with a season-cluster bootstrap standard error, signed and then judged against a benjamini-hochberg correction over all ten quantile fits in both contexts
• Family: margin_quantile_trends (m=10, BH FDR q=0.05)

Regular season: both margin tails move apart, Q10 (big home losses) falls and Q90 (big home wins) rises, each surviving BH correction across all ten quantile fits.

• Evidence: Q10 -0.171 (BH p=0.000), Q90 +0.055 (BH p=0.009)
• Also said in: The Altitude Holdouts, The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: dither-averaged unconditional quantile slopes of home margin on year at q10 and q90, each with a season-cluster bootstrap standard error, signed and then judged against a benjamini-hochberg correction over all ten quantile fits in both contexts

The Altitude Holdouts

Denver and Utah hold the two largest regular-season home-court advantages in the league.

• Evidence: top-2 by shrunken HCA: Denver Nuggets (+26.8), Utah Jazz (+25.7)
• Also said in: Home Court Advantage: Frequently Asked Questions
• Method: ranking of franchises by empirical-bayes shrunken regular-season home-minus-road win rate, taking the top two
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Every franchise with a long enough record in both eras saw its home-court advantage fall, the altitude teams included.

• Evidence: 26/26 declined
• Also said in: Home Court Advantage: Frequently Asked Questions
• Method: count of qualifying franchises whose era-to-era change in home-minus-road win rate is negative, against the full qualifying set
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The four franchises just below Denver and Utah on the all-time ranking whose name or city later changed (Bullets, SuperSonics, Kansas City Kings, New Jersey Nets) each rank near the top, and the three whose continuation the article names (Wizards, Thunder, Brooklyn Nets) each sit below their old name and in the bottom half.

• Evidence: Washington Bullets #3 vs Washington Wizards #28 of 39; Seattle SuperSonics #4 vs Oklahoma City Thunder #37 of 39; New Jersey Nets #6 vs Brooklyn Nets #39 of 39; Kansas City Kings #5 of 39 (rank only)
• Method: rank positions of four relocated franchise names, and of the three modern continuations the article names, on the all-time shrunken home-court ranking, checked against the top third and the bottom half of that table
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Sacramento and Phoenix now sit in the bottom third of recent-era home court advantage.

• Evidence: Charlotte Hornets (+9), Boston Celtics (+13), New York Knicks (+14), Philadelphia 76ers (+14), Minnesota Timberwolves (+14), Phoenix Suns (+14), Chicago Bulls (+15), Sacramento Kings (+15)
• Method: ranking of the qualifying franchises by their recent-era home-minus-road win rate, checking membership of the bottom third
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Sacramento and Phoenix fell the most.

• Evidence: Sacramento Kings (-16), Phoenix Suns (-15)
• Method: per-franchise change in home-minus-road win rate between the pre-2002 and post-2002 eras, ranked ascending, taking the two largest falls
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The Los Angeles Lakers barely moved, a drop of under a point.

• Evidence: Lakers HCA drop 0.8 pp
• Method: one named franchise magnitude of change in home-minus-road win rate across the same pre-2002 against post-2002 era split
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Playoffs: the widening is mostly the Q90 tail (big home wins), which moves further than the other end does.

• Evidence: |Q90| 0.228 vs |Q10| 0.121 pts/yr (65% of the total movement)
• Method: magnitudes of the dither-averaged q10 and q90 unconditional quantile slopes of home margin on year, compared against each other within one context, with the tail expected to dominate fixed per context rather than taken as an argmax
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Regular season: the widening is mostly the Q10 tail (big home losses), which moves further than the other end does.

• Evidence: |Q10| 0.171 vs |Q90| 0.055 pts/yr (76% of the total movement)
• Method: magnitudes of the dither-averaged q10 and q90 unconditional quantile slopes of home margin on year, compared against each other within one context, with the tail expected to dominate fixed per context rather than taken as an argmax
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

How I Know This Isn’t Made Up

Every specification shows a decline.

• Evidence: 12/12 negative
• Method: specification curve over the twelve regular-season decline fits: three start years by two estimators (binomial GLM and ordinary least squares on the season win rate) by keeping or dropping the COVID seasons, counting how many slopes land below zero
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Starting the clock in 1995 gives the shallowest slopes in the specification set.

• Evidence: 4 shallowest specs start in from 1995
• Method: rank of the twelve decline specifications by slope, checking whether the shallowest block is exactly the rows whose start year is 1995
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The start year is the only one of the three analytic choices that moves the slope much; the estimator and the COVID handling barely move it.

• Evidence: level-mean swing (pp/yr): COVID 0.013, estimator 0.004, start year 0.069
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: level-mean swing per factor across the same twelve-specification decline curve: for each analytic choice, the spread between the mean slope of its levels, compared across the three choices
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The two- and three-bend fits place their most recent bend in the same season.

• Evidence: k=2 last 2020, k=3 last 2020
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: the maximum-a-posteriori break years of the two-break and the three-break change-point fits, compared for agreement on the most recent one; an equality check between two fits
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The flexible model puts rebounding, turnovers and fouls within a few points of their straight-line shares.

• Evidence: rebounding 28% vs 30%; turnovers 23% vs 27%; fouls 15% vs 18%
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: each channel’s SHAP share of the regular-season decline compared against its straight-line share from the linear mediation, thresholded at 6 percentage points of agreement for rebounding, turnovers and fouls; a distance between two point estimates, with no interval on either
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Shooting tops the flexible (SHAP) model’s own ranking of the decline.

• Evidence: top channel Shooting at 34%
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: ranking the four channels by their mean SHAP contribution to the early-minus-late regular-season decline in a gradient-boosted win model fit on the same four box-score edges
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

In the playoff flexible (SHAP) model’s own split, shooting and rebounding are the two largest channels, ahead of fouls and turnovers.

• Evidence: Shooting +3.5, Rebounding +3.3, Fouls +1.3, Turnovers +1.2 (pp)
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: ranking the four channels by signed SHAP contribution within the playoff gradient-boosted win model, checking that the smaller of shooting and rebounding still beats the larger of fouls and turnovers; this model’s own internal ranking, not agreement with the linear split
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Rebounding is the largest single driver of the playoff decline too, though on far fewer games the four shares are not separated from one another.

• Evidence: largest is rebounding (28%); shooting 12%, fouls 18%, turnovers 10%, rebounding 28%
• Also said in: Home Court Advantage: Frequently Asked Questions
• Method: ranking the four box-score channels by their share of the PLAYOFF decline in the same linear-probability mediation the regular-season ranking uses, taking the argmax only; no separation between the shares is asserted
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

A hidden cause would have to explain a large share of everything left unexplained in both the channel and who wins to overturn the shooting or the foul link.

• Evidence: eFG RV 60.5%, foul RV 28.9%
• Method: robustness values in the sense of Cinelli and Hazlett, computed from the classical t-statistic and residual degrees of freedom of the full linear-probability model, thresholded at 40 percent for the shooting channel and 20 percent for the foul channel
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The era-by-era decline barely moves once home/away team identity is controlled for.

• Evidence: largest era-coefficient shift 0.5 pp
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: largest change in any era coefficient of the regular-season home win logit when home-team and away-team fixed effects are added, both fits clustered by season
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Home Court Advantage: Frequently Asked Questions

The regular-season pace-free rebound share edge collapsed roughly tenfold.

• Evidence: +2.14 → +0.21
• Also said in: The Investigation: What Drives Home Court Advantage, and What Doesn’t
• Method: era means of the pace-free rebound share edge (home share of available offensive boards minus the away share), earliest era divided by the most recent, regular season
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The central forecast keeps sliding in both the regular season and the playoffs, and the whole plausible range for the final regular-season forecast year stays below the mid-1980s home win rate.

• Evidence: RS 54.6→53.5%, PO 58.4→57.1%, RS upper 58.6% < early 66.7%
• Method: a local-linear-trend state-space model fit to each context’s season home-win series and extrapolated forward; the central path is checked for falling in every step of both contexts, and the final regular-season 95 percent upper bound against the early level
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The Investigation: What Drives Home Court Advantage, and What Doesn’t

The four-factor share of the regular-season decline stays within a few points of complete across every mediation specification: the whole curve spans under ten percentage points, and its median sits within ten points of a complete (100%) share.

• Evidence: spread 5.9 pp; median 98.7%, range [96.0, 101.9]%
• Method: spread (largest minus smallest) of the same eight-specification mediation-share curve, and the distance of its median from a 100% (complete) share, each thresholded at ten percentage points. It asserts that the curve is TIGHT and sits near complete mediation; it asserts nothing about any individual specification’s precision, and the eight shares are point estimates with no interval attached here (the bootstrap bands on the published share are a separate computation)
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The era-by-era rest estimates do not shrink as home court fades; if anything they drift upward, with the most recent era the largest, so the stability result is a null and not a flat line.

• Evidence: pp per day of rest by era: +1.1, +1.7, +1.1, +1.3, +1.6, +2.8
• Method: bivariate rest-differential logistic fitted inside each era, regular season, its per-day coefficients converted to percentage points and read for direction only: last era against first, and last era against the maximum. The interaction test next door (rest_gap_stable_across_eras) is what establishes that this drift is not distinguishable from noise; see “Rest and Altitude” in questions/home_court/docs/home_court_investigation.md
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

Home teams retained a substantial share of their altitude advantage after 2014 even as the overall home-win level fell.

• Evidence: post-2014 altitude log-odds +0.209 (54% of the pre-2014 +0.389)
• Method: the altitude main effect plus its post-2014 interaction in the same pooled stability logit, checked for staying positive and keeping at least 40 percent of the pre-2014 slope. A coefficient magnitude: the p-value on that interaction backs the neighbouring claim, not this one
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

No situational effect’s post-2014 change survives correction for the three interactions tested together; the altitude one is the largest of the three but is not individually established.

• Evidence: rest_diff × post2014 p=0.142, altitude_home × post2014 p=0.026, tz_diff × post2014 p=0.917 (0/3 survive BH)
• Method: a pooled regular-season logit carrying post-2014 interactions on rest, altitude and time zone at once, with Benjamini-Hochberg FDR at q=0.05 across exactly the interaction terms the fit retained; the guard counts survivors and refuses a zero-of-zero verdict
• Family: stability_interactions (m=3, BH FDR q=0.05)

Separating out which team was the better seed costs the playoff rest effect roughly a third of its size and leaves a band that includes no effect at all.

• Evidence: 2.4 -> 1.6 pp per day of rest (67% retained), p = 0.113
• Method: playoff logistic of the home win on the rest differential, fitted bivariate and again with the two teams regular-season quality gap in the model, read as the share of the bivariate per-day effect the controlled fit retains and as whether that controlled effect still clears the bar; see “Rest and Altitude” in questions/home_court/docs/home_court_investigation.md
• Family: standalone (the second specification of the playoff rest hypothesis, whose bivariate first is corrected inside primary_tests; primary_tests’ own note already records that the reader meets playoff rest through this quality-controlled fit rather than through the bivariate corrected there, and a fact can enter only one family in this census. Both specifications ran and both are published; this one reads null, and a correction can only make a null reading safer.)

Season to season, pace and home court advantage barely move together, and the two contexts point opposite ways: a weak positive correlation in the regular season against a weak negative one in the playoffs.

• Evidence: season-level r: regular season +0.241 (p=0.120), playoffs -0.142 (p=0.370)
• Method: season-level pearson correlation of mean pace against the home win rate, taken separately in the regular season and the playoffs, thresholding each correlation at |r| < 0.30 and requiring the regular-season one to be positive and the playoff one negative. It asserts SIZE and DIRECTION only. It does NOT assert that either correlation is statistically insignificant, even though both p-values are far from any threshold and are carried in the value beside the correlations; and it says nothing about the game-level link, which is a different test and is guarded separately by pace_link_fades_expected; see “Pace of Play” in questions/home_court/docs/home_court_investigation.md
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The combined situational model contains exactly era, rest, altitude, time zone, and the COVID seasons; travel is tested separately and is not one of its variables.

• Evidence: blocks = era, rest, altitude, tz, covid
• Method: an equality check on the block list the combined situational logit is actually built from, so that adding or dropping a predictor fails here rather than leaving the prose describing a model that no longer exists; a check on the model, not on data
• Names no p-value, so no multiple-comparison family applies. That covers a count or a direct comparison, and also a judgement made from an interval rather than from a test statistic.

The test families

Significance tests run together are corrected together, and this is every family this question declares. The first number is what was RUN, not what is reported: where more tests ran than reached a claim above, the difference is stated with the reason for it, because a count of tests that quietly matched the count of findings would be the more suspicious page. A claim resting on tests from two families is counted under both, so these claim counts can add up to more than the number of claims above.

Each entry also says how many of the p-values a family publishes clear the family’s own correction. That is a statement about a threshold and not a retraction: every estimate, interval and p-value behind this page is reported as it came out, and this line says where each one falls once the family it was run in is taken into account. Where a family was run to rule something out, a p-value that does not clear is the finding rather than a problem with it.

Ranks are taken among the members whose p-values this question publishes, which is the strict reading rather than a flattering one. A test whose p-value is not published could only push a published member to a higher rank, and a BH threshold rises with the rank, so a member that clears here clears whatever the rest of its family did. The converse does not follow, and the wording is chosen for it: a member listed as not clearing is one this page cannot show clears, not one the correction is known to reject.

attendance_hca_specs

• Tests run together: 4, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 1 of this family’s 4 published p-values clears its threshold. Those that do not clear: attendance.residual_corr_p (p = 0.098, threshold 0.025); attendance.raw_corr_p (p = 0.212, threshold 0.038); attendance.firstdiff_corr_p (p = 0.724, threshold 0.050).
• Why none of the 4 is reported here: all four specifications ran and all four are published with their p-values. What no claim rests on is any one of them: crowd size is ruled out by the attendance level staying near capacity while the advantage fell, not by a correlation failing to land. Only the Spearman specification reaches significance, and it does so with a negative sign, which argues against the crowd explanation rather than for it.

differential_trends

• Tests run together: 14, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 1 of this family’s 2 published p-values clears its threshold. The one that does not clear: trend_po.fouls_p (p < 0.010, threshold 0.007).
• Why none of the 14 is reported here: every tracked box-score differential is fit against year in both the regular season and the playoffs, in one loop, before any of them is looked at: the column list is fixed in run_differential_analysis and nothing is dropped for how it came out. Two of those fits are quoted, in the methods companion and nowhere a reader goes, and they are the two the surrounding prose is about rather than the two that landed best. The slopes of the others are printed in the results file with their era tables.

format_2014_change

• Tests run together: 3, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 1 of this family’s 2 published p-values clears its threshold. The one that does not clear: format.lr_test_p (p = 0.197, threshold 0.033).
• Why none of the 3 is reported here: the third test is the most recent format-period dummy’s coefficient p in the trend-controlled logit, which no FACTS.set publishes under any name. It is the one the section’s verdict actually turns on, and it is reported in the results file and in the console output rather than as a fact; see the comment above format_2014_not_significant for why that guard names no test at all. All three point the same way, and the raw drop is the only one of the three that reaches significance on its own.

franchise_hca_consistency

• Tests run together: 2, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 0 of this family’s 2 published p-values clear their threshold. Those that do not clear: franchise.consistency_pearson_p (p = 0.042, threshold 0.025); franchise.consistency_spearman_p (p = 0.128, threshold 0.050).
• Why none of the 2 is reported here: both correlations ran and both are published with their p-values. What no claim on this page rests on is either of them: the articles say the franchises with a home-court reputation are not the ones who keep it in the playoffs, which is argued from the mean gap and its spread rather than from the correlation’s significance.

granger_3pa_lead

• Tests run together: 4, corrected BH FDR q=0.05
• Reported here: 2 of them, behind 1 claim on this page
• Against this family’s own correction: 0 of this family’s 4 published p-values clear their threshold. Those that do not clear: granger.tpa_to_hw_lag1_p (p = 0.230, threshold 0.013); granger.tpa_to_hw_lag2_p (p = 0.268, threshold 0.025); granger.hw_to_tpa_lag1_p (p = 0.421, threshold 0.038); granger.hw_to_tpa_lag2_p (p = 0.743, threshold 0.050).
• Why the other 2 are not reported here: all four tests ran and all four are published with their p-values: two directions, the three-point rate leading home court and the reverse, each at one lag and at two. The lag count is fixed in the code before any fit, at two, because the differenced series leaves too few degrees of freedom for a third, so nothing was added or dropped after a result was seen. The two augmented Dickey-Fuller tests in the same computation are not counted here: they decide whether the series are differenced before the Granger tests run, and they test no home-court claim. All four came out null, and the prose reads them together as one finding rather than one each.

margin_quantile_trends

• Tests run together: 10, corrected BH FDR q=0.05
• Reported here: 2 of them, behind 1 claim on this page
• Against this family’s own correction: 2 of this family’s 2 published p-values clear their threshold.
• Why the other 8 are not reported here: the five quantiles (0.10, 0.25, 0.50, 0.75, 0.90) and the two contexts are fixed in the code before any fit runs, so every slope in the sweep is estimated and every one enters the correction. Nothing was dropped for failing to reach significance. What decides which of them appear above is which document quotes them, not how they came out: the general-reader articles print no p-value at all by house rule, and the investigation quotes the playoff tails, where the claim that both tails moved turns on per-quantile significance. The regular-season tails back a claim of their own on exactly the same footing and are simply not quoted in a document that prints p-values, and the interior quantiles back no claim in either context and are published as slopes only. The full sweep, adjusted p-values included, is printed in the results file.

net_rating_trends

• Tests run together: 2, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 1 of this family’s 1 published p-value clears its threshold.
• Why none of the 2 is reported here: the home net-rating year trend is fit in both contexts and both slopes are published; only the regular-season p-value is, because the playoff series is short enough that the articles quote its slope as a direction rather than resting anything on its significance. No claim on this page rests on either p-value.

parity_hca_specs

• Tests run together: 3, corrected BH FDR q=0.05
• Reported here: 2 of them, behind 1 claim on this page
• Against this family’s own correction: 0 of this family’s 2 published p-values clear their threshold. Those that do not clear: parity.residual_corr_p (p = 0.025, threshold 0.017); parity.raw_corr_p (p = 0.556, threshold 0.033).
• Why the other 1 is not reported here: all three specifications ran and all three are published with their p-values. The third, the first-differenced correlation, is corrected inside primary_tests rather than here, because that is where the fit it is reported beside runs and a fact can enter only one family in this census. It still counts toward this denominator, which is why m is one more than the members listed.

placebo_step_scan

• Tests run together: 24, corrected none
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: it declares none, so there is no threshold here for its published p-values to fall short of.
• Why none of the 24 is reported here: the scan fits a step dummy at every candidate year from 1987 to 2010, so that 1994-95 can be read against the rest of the distribution rather than on its own. Only the 1994-95 step is published, because the question the scan answers is whether that boundary stands out, not whether any single year is significant. The other years are not withheld: each is printed with its coefficient and p-value in the results file, and 1994-95 is not the strongest step in the scan.

playoff_seed_quality

• Tests run together: 2, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 2 of this family’s 2 published p-values clear their threshold.
• Why none of the 2 is reported here: both tests ran and both are published with their p-values. Neither backs a claim on this page, and that is the point of the section: what it publishes is the share of the playoff year-trend still standing after the quality control, a magnitude rather than a test, and these two are the diagnostics that say the control was worth applying.

primary_tests

• Tests run together: 14, corrected BH FDR q=0.05
• Reported here: 2 of them, behind 2 claims on this page
• Against this family’s own correction: 6 of this family’s 9 published p-values clear their threshold. Those that do not clear: tpa.within_po_p (p = 0.027, threshold 0.025); parity.firstdiff_corr_p (p = 0.033, threshold 0.029); pace.expected_era_p (p = 0.063, threshold 0.032).
• Why the other 12 are not reported here: the tests in this family split three ways. Most are published with their p-values, and the investigation reports them as evidence a reader can check rather than as the number a claim is thresholded against, which is what the count above measures. Four are game-level bivariate factor tests the table re-runs so that the correction has an honest denominator: rest in the regular season, rest in the playoffs, altitude, and the time-zone gap. Neither rest test is published, and the reader meets rest through the combined situational model and a separate quality-controlled playoff fit, neither of which is the bivariate corrected here; altitude and the time-zone gap do reach the reader, but from the plain-SE bivariate rather than the cluster-robust fit corrected here, which is a real difference between two estimators and not a formality. One test is published and not enrolled: the playoff era-dummy likelihood-ratio test comes out of this same computation and reaches the reader as era.lr_p_po, which reads null and is reported as the absence of any rule-change fingerprint in the playoffs, a reading a correction can only make safer. its regular-season twin is enrolled, as is the within-era expected-pace test, which is the same cluster-robust fit the mechanisms section publishes. None of them was dropped for being inconvenient: the table is fixed before it is run, and every test in it is printed in the results file with its p-value, its BH threshold and its verdict.

reb_share_trend

• Tests run together: 2, corrected OLS share-edge slope per context; no cross-context correction
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: not derived. The correction is recorded above as written prose rather than as a rule this page can read as a threshold, so it does not say how many of the 2 p-values this family publishes would clear one.
• Why none of the 2 is reported here: both slopes ran and both are published with their p-values. What no claim on this page does is rest on either of them: what the section asserts is the size of the collapse in the era levels of the share edge, which is a magnitude rather than a significance test, and the two slopes are reported beside it as corroboration.

rebound_trends

• Tests run together: 6, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 3 of this family’s 3 published p-values clear their threshold.
• Why none of the 6 is reported here: the offensive, defensive and total rebounding differentials are each fit against year in both contexts, and the three regular-season p-values are published because the regular season is where the rebounding story is argued. The playoff fits ran on the same columns and are printed as era tables in the results file. The fourth column of the same table, the share edge, is corrected in reb_share_trend instead, which is why this denominator is six and not eight.

referee_bias

• Tests run together: 47, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: this family publishes no p-value of its own, so there is nothing here to hold against its threshold.
• Why none of the 47 is reported here: every one of these tests ran and all of them are corrected together, but the section publishes no per-official p-value. It reports the survivor count and the split between real spread and sampling noise instead, because a table of named officials each carrying its own p-value invites exactly the per-official reading the empirical-Bayes shrinkage exists to discourage: most of the raw spread between officials is noise. Nothing was screened out, and the denominator is every official with a qualifying playoff record rather than a selection from them.

rest_era_interaction

• Tests run together: 2, corrected BH FDR q=0.05
• Reported here: 1 of them, behind 1 claim on this page
• Against this family’s own correction: 0 of this family’s 2 published p-values clear their threshold. Those that do not clear: rest.era_interaction_p_rs (p = 0.474, threshold 0.025); rest.era_interaction_p_po (p = 0.727, threshold 0.050).
• Why the other 1 is not reported here: both tests ran and both are published with their p-values: the same rest-by-era interaction, once in the regular season and once in the playoffs. The denominator is two rather than something wider, and that is a judgement worth stating. The same computation also fits a bivariate rest logistic inside each era, one per era in each context, and every one of those is printed in the results file with its coefficient and its p-value while none is published as a fact. They are the picture the interaction test summarises rather than separate hypotheses, so they are not corrected against it. Both interactions came out null.

rsquality_era_interaction

• Tests run together: 2, corrected BH FDR q=0.05
• Reported here: 1 of them, behind 1 claim on this page
• Against this family’s own correction: 2 of this family’s 2 published p-values clear their threshold.
• Why the other 1 is not reported here: both fits ran and both are published with their p-values. they are corrected together because they ask the same question of the same regular-season games, whether the effect of a prior-season quality gap on who wins at home depends on when the game was played, once across era dummies and once as a continuous interaction with year. only the year interaction carries a direction, so it is the one the section’s claim is thresholded on, but the omnibus ran on the same data in the same block and so belongs in the denominator whatever any claim rests on. both clear the correction: the smallest threshold a family of two imposes at q=0.05 is 0.025, and the year interaction sits many orders of magnitude below it, so nothing in the section turns on the size of the denominator.

series_game_number

• Tests run together: 2, corrected BH FDR q=0.05
• Reported here: 1 of them, behind 1 claim on this page
• Against this family’s own correction: 1 of this family’s 2 published p-values clears its threshold. The one that does not clear: sawtooth.trend_p (p = 0.516, threshold 0.050).
• Why the other 1 is not reported here: both tests ran on the same seven-point table of home win % by game number, and both are published with their p-values. they are corrected together because the section reads them together: the chi-square asks whether the seven rates differ at all, the count-weighted slope asks whether they drift in one direction as a series wears on, and the reading the section draws is that they differ sharply without trending, which is what a hosting pattern looks like rather than a fatigue one. nothing else in the section is a test: the game-5 superlative is an argmax over the same seven rates and the game-7 comparison is a distance between three of them, which is why the denominator is two and not four. correcting the pair leaves both readings where they were. the smallest threshold a family of two imposes at q=0.05 is 0.025 and the uniformity test sits far below it, while the trend stays well above 0.05, so the correction moves the trend further from significance rather than nearer, which is the direction that matters for a claim resting on it being flat.

shot_zone_trends

• Tests run together: 8, corrected BH FDR q=0.05
• Reported here: none of them is behind a claim on this page
• Against this family’s own correction: 2 of this family’s 2 published p-values clear their threshold.
• Why none of the 8 is reported here: all four shot zones (paint, mid-range, corner three, above the break) are fit against year in both contexts, in one comprehension over a fixed zone list, so nothing here was selected after the fact. The two published p-values are the two zones the three-point story is about, and the playoff paint trend is published as a slope with its p-value deliberately dropped. Every zone’s era table and slope is printed in the results file.

stability_interactions

• Tests run together: 3, corrected BH FDR q=0.05
• Reported here: 3 of them, behind 1 claim on this page
• Against this family’s own correction: 0 of this family’s 3 published p-values clear their threshold. Those that do not clear: stability.altitude_interaction_p (p = 0.026, threshold 0.017); stability.rest_interaction_p (p = 0.142, threshold 0.033); stability.tz_interaction_p (p = 0.917, threshold 0.050).

tracking_era_trends

• Tests run together: 3, corrected BH FDR q=0.05
• Reported here: 3 of them, behind 1 claim on this page
• Against this family’s own correction: 0 of this family’s 3 published p-values clear their threshold. Those that do not clear: track.oreb_conv_trend_p (p = 0.024, threshold 0.017); track.secondchance_trend_p (p = 0.304, threshold 0.033); track.boxout_p (p = 0.775, threshold 0.050).