Marathon world record progression, 1908 to 2026: does a break appear in the shoe era after 2016?
- Status
- SUCCEEDED
- Started
- Finished
- Sessions
- 1
Goal
Question: after 2016, did the men's and women's marathon world record fall faster than the earlier trend, and how large is any break compared with its interval? I care because shoe claims often rest on single races or photos. A dated public record list is a cleaner test. A reader gets a break estimate in minutes per kilometre with a confidence interval, plus a plain statement of what a record list cannot separate (shoes, pacing teams, course choice, more athletes). This is a cohort-style time series, not a trial, and the post will say so.
Plan
1. Download the Wikipedia pages for marathon world record progression (men and women) from en.wikipedia.org with the download tool. Parse date and time into a pandas table. Save to the workspace. Read files with python -I from a separate script folder. 2. Convert each record to pace in minutes per kilometre (42.195 km). Check the parse by comparing five rows to the page by hand and by count of rows. 3. Fit three models to 1980 to 2026 records by sex: a linear trend, a log trend, and a trend with a level and slope change at 2016 and at 2019 (ratified records). Use HAC or bootstrap intervals over record events. Compare models with AIC and leave-one-out error. 4. Run a placebo test: place the same break at every year from 1990 to 2010 and report how often the fitted break equals or exceeds the 2016 one. 5. Outputs: one figure of pace against date with fitted models, one table of break size with 95% interval, one placebo histogram. 6. Success: a break estimate with interval and a placebo rank, and the results reported even if the interval includes zero. Failure: fewer than 15 parsed records per sex or a parse that fails the hand check. Then report the failure and stop. 7. Write limits: record lists have few points, are autocorrelated, and cannot isolate shoes from other causes.
Summary
Wikipedia's marathon record tables parsed cleanly (53 men's rows, 47 women's rows, hand-checked). A linear model with a slope break at 2016 shows the men's record falling faster afterwards, by 0.0034 min/km per year (HAC 95% CI -0.0042 to -0.0025). That result depends on the model: a quadratic trend with no break fits as well or better. The women's result is weaker and its sign depends on the start year.
Outputs
Record pace (min per km) by year, 1980 to 2026, with a linear trend and a trend with a break at 2016. 
Slope change at placebo break years 1990 to 2010 against the 2016 estimate (red line), men and women. - Download Break estimates (min per km per year) with HAC and block-bootstrap 95% intervals, AIC and leave-one-out error.
Resulting post
The Men's Marathon Record Fell Faster After 2016. A Curve Explains Half.
I fitted a slope break at 2016 to 47 years of marathon world records. The men's break is real in one model and half gone in another. The women's break does not hold.
Step log
1. Download the Wikipedia pages for marathon world record progression (men and women) from en.wikipedia.org with the download tool. Parse date and time into a pandas table. Save to the workspace. Read files with python -I from a separate script folder. 2. Convert each record to pace in minutes per kilometre (42.195 km). Check the parse by comparing five rows to the page by hand and by count of rows. 3. Fit three models to 1980 to 2026 records by sex: a linear trend, a log trend, and a trend with a level and slope change at 2016 and at 2019 (ratified records). Use HAC or bootstrap intervals over record events. Compare models with AIC and leave-one-out error. 4. Run a placebo test: place the same break at every year from 1990 to 2010 and report how often the fitted break equals or exceeds the 2016 one. 5. Outputs: one figure of pace against date with fitted models, one table of break size with 95% interval, one placebo histogram. 6. Success: a break estimate with interval and a placebo rank, and the results reported even if the interval includes zero. Failure: fewer than 15 parsed records per sex or a parse that fails the hand check. Then report the failure and stop. 7. Write limits: record lists have few points, are autocorrelated, and cannot isolate shoes from other causes.
The command failed. The run is logged for the operator.
$ cat > /work/scripts/tp.py <<'E' from html.parser import HTMLParser import re class P(HTMLParser): def __init__(s): super().__init__(); s.tables=[]; s.stack=[]; s.cell=None; s.row=None; s.hdr=[] def handle_starttag(s,t,a): if t=='table': s.stack.append([]); s.hdr.append(dict(a).get('class','')) elif t=='tr' and s.stack: s.row=[] elif t in('td','th') and s.row is not None: s.cell='' elif t=='sup' and s.cell is not None: s.cell+='' ; s.sup=True def handle_endtag(s,t): if t in('td','th') and s.cell is not None and s.row is not None: s.row.append(re.sub(r'\[\d+\]|\[[a-z]\]','',s.cell).strip()); s.cell=None elif t=='tr' and s.stack and s.row is not None: s.stack[-1].append(s.row); s.row=None elif t=='table' and s.stack: s.tables.append((s.hdr.pop(),s.stack.pop())) def handle_data(s,d): if s.cell is not None: s.cell+=d def tables(path): p=P(); p.feed(open(path,encoding='utf8').read()); return p.tables E cat > /work/scripts/inspect.py <<'E' import sys; sys.path.insert(0,'/work/scripts') from tp import tables for i,(c,t) in enumerate(tables('/work/data/marathon_wr.html')): print(i,c,len(t)); for r in t[:3]: print(' ',r) EShow 9 more lines
cd /work && python -I scripts/inspect.py exit 0 after 0.4 s [...]ar-mini abbr{font-variant:small-caps;border-bottom:none;text-decoration:none;cursor:inherit}.mw-parser-output .navbar-ct-full{font-size:114%;margin:0 7em}.mw-parser-output .navbar-ct-mini{font-size:114%;margin:0 4em}html.skin-theme-clientpref-night .mw-parser-output .navbar li a abbr{color:var(--color-base)!important}@media(prefers-color-scheme:dark){html.skin-theme-clientpref-os .mw-parser-output .navbar li a abbr{color:var(--color-base)!important}}@media print{.mw-parser-output .navbar{display:none!important}}vteAthletics record progressions'] ['European', '100 metres\nmen\n200 metres\nmen\n400 metres\nmen\n800 metres\nmen\n1500 metres\nmen\n5000 metres\nmen'] ['List of world records in athletics'] 4 nowraplinks hlist mw-collapsible autocollapse navbox-inner 7 ['vteMarathon running'] ['List of marathon races', 'Africa\nAsia\nEurope\nNorth America\nOceania\nSouth America'] ['International competitions', 'African Games\nAsian Games\u200e\u200e\nCommonwealth Games\u200e\nEuropean Championships\nEuropean Marathon Cup\u200e\u200e\nOlympics\nPan American Games\u200e\nParalympics\nWorld Championships\nWorld Marathon Cup\u200e\u200e\nWorld Athletics Label Road Races']$ cat > /work/scripts/dump.py <<'E' import sys; sys.path.insert(0,'/work/scripts') from tp import tables T=tables('/work/data/marathon_wr.html') for k in (0,1): print('TABLE',k) for r in T[k][1][1:]: print(r[:5]) E cd /work && python -I scripts/dump.py exit 0 after 0.4 s [...]oroupe', 'Kenya', 'April 19, 1998', 'Rotterdam Marathon'] ['2:20:43', 'September 26, 1999', 'Berlin Marathon', ''] ['2:19:46', 'Naoko Takahashi', 'Japan', 'September 30, 2001', ''] ['2:18:47', 'Catherine Ndereba', 'Kenya', 'October 7, 2001', 'Chicago Marathon'] ['2:17:18', 'Paula Radcliffe', 'United Kingdom', 'October 13, 2002', 'Chicago Marathon'] ['2:15:25 Mx', 'April 13, 2003', 'London Marathon', 'IAAF, ARRS', "First world record for the women's marathon ratified by the International Association of Athletics Federations. The ARRS notes Radcliffe's extended time as 2:15:24.6"] ['2:17:42 Wo', 'Great Britain', 'April 17, 2005', 'IAAF', ''] ['2:17:01 Wo', 'Mary Jepkosgei Keitany', 'Kenya', 'April 23, 2017', 'IAAF'] ['2:14:04 Mx', 'Brigid Kosgei', 'October 13, 2019', 'Chicago Marathon', 'IAAF'] ['2:11:53 Mx', 'Tigst Assefa', 'Ethiopia', 'September 24, 2023', 'Berlin Marathon'] ['2:16:16 Wo', 'Peres Jepchirchir', 'Kenya', 'April 21, 2024', 'London Marathon'] ['2:09:56 Mx', "Ruth Chepng'etich", 'October 13, 2024', 'Chicago Marathon', 'World Athletics'] ['2:15:50 Wo', 'Tigst Assefa', 'Ethiopia', 'April 27, 2025', 'London Marathon'] ['2:15:41 Wo', 'April 26, 2026', 'World Athletics', '']The command failed. The run is logged for the operator.
$ cd /work && mv scripts/inspect.py scripts/look_tables.py && rm -rf scripts/__pycache__ && python -I scripts/parse.py exit 0 after 1.0 s n nodate disp wo mx sex M 53 0 2 0 0 W 47 0 8 5 4 min max sex M 1908-07-24 2026-04-26 W 1918-09-29 2026-04-26 {'sex': 'M', 'raw': '2:55:18.4', 'date': Timestamp('1908-07-24 00:00:00'), 'pace': 4.1546786744085} {'sex': 'M', 'raw': '2:26:14', 'date': Timestamp('1935-03-21 00:00:00'), 'pace': 3.4656554884069988} {'sex': 'M', 'raw': '2:06:05', 'date': Timestamp('1998-09-20 00:00:00'), 'pace': 2.988110755618754} {'sex': 'M', 'raw': '2:01:09', 'date': Timestamp('2022-09-25 00:00:00'), 'pace': 2.871193269344709} {'sex': 'W', 'raw': '3:02:53', 'date': Timestamp('1970-02-28 00:00:00'), 'pace': 4.334241813801003} {'sex': 'W', 'raw': '2:35:15.4', 'date': Timestamp('1977-05-01 00:00:00'), 'pace': 3.6795038906663504} {'sex': 'W', 'raw': '2:21:06', 'date': Timestamp('1985-04-21 00:00:00'), 'pace': 3.343998104040763} {'sex': 'W', 'raw': '2:11:53 Mx', 'date': Timestamp('2023-09-24 00:00:00'), 'pace': 3.1255678002922935}$ cat > /work/scripts/analyse.py <<'E' import numpy as np, pandas as pd, json rng=np.random.default_rng(20261011) df=pd.read_csv('/work/parsed_records.csv',parse_dates=['date']) print(df[(df.sex=='W')&df.disputed][['raw','date']]) def series(sex,women_only=False): d=df[(df.sex==sex)&~df.disputed&~df.claim].copy() if sex=='W': d=d[d.tag=='Wo'] if women_only else d[d.tag!='Wo'] d=d.sort_values('date'); d['rm']=d.pace.cummin() ev=d[d.pace<=d.rm+1e-12] # progressive records return ev def annual(ev,y0=1980,y1=2026): yrs=np.arange(y0,y1+1); v=[] for y in yrs: e=ev[ev.date<=pd.Timestamp(y,12,31)]; v.append(e.pace.iloc[-1]) return yrs,np.array(v) def nw_cov(X,res,L): n=len(res); XtXi=np.linalg.inv(X.T@X); u=X*res[:,None]; S=u.T@u for l in range(1,L+1): w=1-l/(L+1); G=u[l:].T@u[:-l]; S+=w*(G+G.T) return XtXi@S@XtXi def design(t,T,brk=True): c=[np.ones_like(t),t-2016.0] if brk: c+= [(t>=T).astype(float),(t>=T)*(t-T)] return np.column_stack(c) def fit(t,y,T,brk=True): X=design(t,T,brk); b=np.linalg.lstsq(X,y,rcond=None)[0]; r=y-X@b return X,b,r def loo(t,y,T,brk,logy=False): e=[]Show 91 more lines
for i in range(len(t)): m=np.ones(len(t),bool); m[i]=False X=design(t[m],T,brk); b=np.linalg.lstsq(X,y[m],rcond=None)[0] e.append(y[i]-design(t[i:i+1],T,brk)@b) return float(np.sqrt(np.mean(np.square(e)))) def aic(r,k): n=len(r); return n*np.log(np.sum(r**2)/n)+2*k def block_boot(t,y,T,B=2000,bl=5): X,b,r=fit(t,y,T); fitted=X@b; n=len(y); out=[] for _ in range(B): idx=[] while len(idx)<n: s=rng.integers(0,n-bl+1); idx+=list(range(s,s+bl)) rb=r[idx[:n]]; yb=fitted+rb out.append(fit(t,yb,T)[1]) return np.array(out) res={}; rows=[] for label,sex,wo in (('Men','M',False),('Women (mixed-race record)','W',False),('Women (women-only record)','W',True)): ev=series(sex,wo); y0=1980 if not wo else 2005 ev.to_csv(f'/work/events_{label.split()[0]}_{"wo" if wo else "main"}.csv',index=False) n_ev=(ev.date.dt.year>=1980).sum() t,y=annual(ev,y0 if sex=='M' or not wo else 2005) print(label,'events since 1980:',n_ev,'annual points',len(t)) if wo: continue for T in (2016,2019): X0,b0,r0=fit(t,y,T,False); X1,b1,r1=fit(t,y,T,True) cov=nw_cov(X1,r1,4); se=np.sqrt(np.diag(cov)) bb=block_boot(t,y,T) # slope-change d (index 3), level change c (index 2) # also post-T slope = b[1]+b[3] post=bb[:,1]+bb[:,3] r=dict(group=label,T=T,n=len(t), pre_slope=b1[1],post_slope=b1[1]+b1[3],slope_change=b1[3],slope_change_hac_ci=[b1[3]-1.96*se[3],b1[3]+1.96*se[3]], slope_change_boot_ci=list(np.percentile(bb[:,3],[2.5,97.5])),level_change=b1[2],level_boot_ci=list(np.percentile(bb[:,2],[2.5,97.5])), aic_linear=aic(r0,2),aic_break=aic(r1,4),loo_linear=loo(t,y,T,False),loo_break=loo(t,y,T,True)) # log model ly=np.log(y); X0l,b0l,r0l=fit(t,ly,T,False); X1l,b1l,r1l=fit(t,ly,T,True) r.update(aic_log=aic(r0l,2),aic_logbreak=aic(r1l,4),loo_log=loo(t,ly,T,False)); # placebo: slope change at years 1990..2010, same window pl={} for P in range(1990,2011): pl[P]=fit(t,y,P,True)[1][3] r['placebo']=pl rows.append(r) print(json.dumps({k:v for k,v in r.items() if k!='placebo'},default=float,indent=1)) vals=np.array(list(pl.values())) # rank: how many placebos have slope change at least as negative (faster fall) than at T print(' placebo slope changes min/med/max',vals.min(),np.median(vals),vals.max(),' n as negative or more:',(vals<=b1[3]).sum(),'of',len(vals)) res[(label,T)]=(t,y,b0,b1) json.dump(rows,open('/work/results.json','w'),default=float) import pickle; pickle.dump(res,open('/work/fits.pkl','wb')) E cd /work && python -I scripts/analyse.py exit 0 after 1.9 s [...]2.3876069089799, "loo_linear": 0.042710183728105665, "loo_break": 0.0400207191058948, "aic_log": -412.2519224512125, "aic_logbreak": -417.2761262746922, "loo_log": 0.012632601688024766 } placebo slope changes min/med/max 0.001098373395266586 0.003935359820767518 0.02005669383819598 n as negative or more: 0 of 21 { "group": "Women (mixed-race record)", "T": 2019, "n": 47, "pre_slope": -0.007702510306959488, "post_slope": -0.01810847971530577, "slope_change": -0.01040596940834628, "slope_change_hac_ci": [ -0.01476499739463838, -0.00604694142205418 ], "slope_change_boot_ci": [ -0.023679341506929623, 0.0037727061518384565 ], "level_change": 0.04832971630920952, "level_boot_ci": [ -0.01926011256139566, 0.1201267117323991 ], "aic_linear": -297.8446592009625, "aic_break": -297.19571464798446, "loo_linear": 0.042710183728105665, "loo_break": 0.042189452413238814, "aic_log": -412.2519224512125, "aic_logbreak": -412.1534016616874, "loo_log": 0.012632601688024766 } placebo slope changes min/med/max 0.001098373395266586 0.003935359820767518 0.02005669383819598 n as negative or more: 0 of 21 Women (women-only record) events since 1980: 5 annual points 22$ cat > /work/scripts/sens_fig.py <<'E' import sys; sys.path.insert(0,'/work/scripts') import numpy as np, pandas as pd, json, pickle import matplotlib; matplotlib.use('Agg'); import matplotlib.pyplot as plt exec(open('/work/scripts/analyse.py').read().split("res={}; rows=[]")[0]) # reuse functions rows=json.load(open('/work/results.json')) def quad(t,y): X=np.column_stack([np.ones_like(t),t-2016.,(t-2016.)**2]); b=np.linalg.lstsq(X,y,rcond=None)[0]; r=y-X@b e=[] for i in range(len(t)): m=np.ones(len(t),bool);m[i]=False; bb=np.linalg.lstsq(X[m],y[m],rcond=None)[0]; e.append(y[i]-X[i]@bb) return aic(r,3),float(np.sqrt(np.mean(np.square(e)))),b out=[] for label,sex in (('Men','M'),('Women','W')): ev=series(sex); t,y=annual(ev) a,l,b=quad(t,y); print(label,'quadratic no-break AIC',a,'LOO',l,'curvature/yr2',b[2]) # quadratic + 2016 break slope change X=np.column_stack([np.ones_like(t),t-2016.,(t-2016.)**2,(t>=2016)*1.,(t>=2016)*(t-2016.)]); bq=np.linalg.lstsq(X,y,rcond=None)[0]; r=y-X@bq cov=nw_cov(X,r,4); se=np.sqrt(np.diag(cov)) print(label,'quad+break slope change 2016',bq[4],'HAC CI',bq[4]-1.96*se[4],bq[4]+1.96*se[4],'AIC',aic(r,5)) # men without 2024-26 (data to 2023-12-31 since last men record before 2026 is Oct 2023: Kiptum) ev=series('M'); ev2=ev[ev.date<'2024-01-01']; t,y=annual(ev2,1980,2023) for T in (2016,): X,b,r=fit(t,y,T); cov=nw_cov(X,r,4); se=np.sqrt(np.diag(cov)); print('Men to 2023 slope change',b[3],b[3]-1.96*se[3],b[3]+1.96*se[3]) ev=series('M'); t,y=annual(ev,1980,2025) X,b,r=fit(t,y,2016); cov=nw_cov(X,r,4); se=np.sqrt(np.diag(cov)); print('Men to 2025 (no Sawe) slope change',b[3],b[3]-1.96*se[3],b[3]+1.96*se[3]) # women start 1970 ev=series('W'); t,y=annual(ev,1970,2026); X,b,r=fit(t,y,2016); cov=nw_cov(X,r,4); se=np.sqrt(np.diag(cov)); print('Women from 1970 slope change',b[3],b[3]-1.96*se[3],b[3]+1.96*se[3]) # exclusion of 2016-start: pre-trend slope only 1980-2015 men extrapolated: gap in 2026 # ---- figuresShow 44 more lines
res=pickle.load(open('/work/fits.pkl','rb')) fig,ax=plt.subplots(1,2,figsize=(12,4.8)) for a,(label,key) in zip(ax,(('Men','Men'),('Women (mixed-race record)','Women (mixed-race record)'))): t,y,b0,b1=res[(key,2016)] a.step(t,y,where='post',color='k',lw=1.5,label='Record pace (year end)') X0=design(t,2016,False); X1=design(t,2016,True) a.plot(t,X0@b0,'--',color='tab:blue',label='Linear trend, no break') a.plot(t,X1@b1,color='tab:red',label='Break at 2016') a.axvline(2016,color='grey',ls=':'); a.set_title(key+': record pace, 1980 to 2026'); a.set_xlabel('Year'); a.set_ylabel('Pace (min per km)'); a.legend(fontsize=8) plt.tight_layout(); plt.savefig('/work/fig_pace_models.png',dpi=150); plt.close() fig,ax=plt.subplots(1,2,figsize=(12,4.5)) for a,key in zip(ax,('Men','Women (mixed-race record)')): r=[x for x in rows if x['group']==key and x['T']==2016][0] pl=np.array(list(r['placebo'].values()))*1000 a.hist(pl,bins=10,color='lightgrey',edgecolor='k',label='Placebo breaks 1990 to 2010') a.axvline(r['slope_change']*1000,color='tab:red',lw=2,label='Break at 2016') a.set_xlabel('Slope change (0.001 min per km per year)'); a.set_ylabel('Count of placebo years'); a.set_title(key); a.legend(fontsize=8) plt.tight_layout(); plt.savefig('/work/fig_placebo.png',dpi=150); plt.close() # table tab=[] for r in rows: tab.append(dict(group=r['group'],break_year=r['T'],slope_change_min_per_km_per_yr=round(r['slope_change'],5),hac_lo=round(r['slope_change_hac_ci'][0],5),hac_hi=round(r['slope_change_hac_ci'][1],5),boot_lo=round(r['slope_change_boot_ci'][0],5),boot_hi=round(r['slope_change_boot_ci'][1],5),pre_slope=round(r['pre_slope'],5),post_slope=round(r['post_slope'],5),aic_linear=round(r['aic_linear'],1),aic_break=round(r['aic_break'],1),loo_rmse_linear=round(r['loo_linear'],4),loo_rmse_break=round(r['loo_break'],4))) pd.DataFrame(tab).to_csv('/work/break_table.csv',index=False); print(pd.DataFrame(tab).T) E cd /work && p [...] exit 0 after 2.7 s [...]e change -0.00275959358789048 -0.0037636659912275586 -0.0017555211845534018 Women from 1970 slope change 0.0020245300632750567 -0.004753887681289884 0.008802947807839996 0 ... 3 group Men ... Women (mixed-race record) break_year 2016 ... 2019 slope_change_min_per_km_per_yr -0.00335 ... -0.01041 hac_lo -0.00419 ... -0.01476 hac_hi -0.00252 ... -0.00605 boot_lo -0.00553 ... -0.02368 boot_hi -0.00128 ... 0.00377 pre_slope -0.00366 ... -0.0077 post_slope -0.00701 ... -0.01811 aic_linear -419.9 ... -297.8 aic_break -447.5 ... -297.2 loo_rmse_linear 0.0116 ... 0.0427 loo_rmse_break 0.0086 ... 0.0422 [13 rows x 4 columns]attached image /media/2026/10/943baf44f21e0689a3a91b3599a44ffb4c26d63b8220018c4e9e17997e28eb5c.png (116.7 KB): Record pace (min per km) by year, 1980 to 2026, with a linear trend and a trend with a break at 2016.
attached image /media/2026/10/43beeff0c29e8bf8de7e935c6aaa3d0660c2e63ae7e85c039e7399217cdebe62.png (60.6 KB): Slope change at placebo break years 1990 to 2010 against the 2016 estimate (red line), men and women.
attached data /media/2026/10/a2becfbdf3b483705656df16c1c48cc1cad7c883f138a9b10568461af6ab1d83.csv (0.6 KB): Break estimates (min per km per year) with HAC and block-bootstrap 95% intervals, AIC and leave-one-out error.
Parse passed: 53 men and 47 women rows, matching the page tables; two paces checked by hand. Dead end: a first script named inspect.py shadowed the standard library under python -I; renamed. Result: men's 2016 slope break is -0.00335 min/km/yr (HAC 95% CI -0.00419 to -0.00252) but a quadratic no-break trend fits as well or better (AIC -453.1 vs -447.5), so the break is model-dependent. Women's interval from the bootstrap includes zero.
Session 1 ended: succeeded Summary: Wikipedia's marathon record tables parsed cleanly (53 men's rows, 47 women's rows, hand-checked). A linear model with a slope break at 2016 shows the men's record falling faster afterwards, by 0.0034 min/km per year (HAC 95% CI -0.0042 to -0.0025). That result depends on the model: a quadratic trend with no break fits as well or better. The women's result is weaker and its sign depends on the start year. Findings: ## Design and data Cohort-style time series of dated public records. It is not a trial. Source: the Wikipedia "Marathon world record progression" page, downloaded 2026-10-11. I read it as records (parsed table rows), not as a primary source. No IAAF or World Athletics lists were read. **Parse check.** - Men: 53 parsed rows. Women: 47 parsed rows. These match the page tables (54 and 48 rows, less the header). Both are above the 15-row threshold. - Two paces were checked by hand. Hayes 2:55:18.4 gives 4.1547 min/km. Da Costa 2:06:05 gives 2.9881 min/km. - Pace is time divided by 42.195 km. - I dropped rows marked "Disputed" and rows that only cite a newspaper claim. Men: 2 disputed rows. Women: 8 disputed rows, mostly short or unmeasured courses before 1984. - The women's main series is the mixed-race record ("Mx" or untagged rows). The women-only ("Wo") rows are a separate series. They have only 5 records since 2005, so I did not model them. - Progressive records come from a running minimum over date. Series are the record pace at each year end, 1980 to 2026 (47 points). The list holds 18 record events since 1980 for each sex. The women's count includes the early-1980s rows. - The year-end series is a step function and strongly autocorrelated. Intervals use a Newey-West (HAC) estimate with 4 lags, and a moving-block bootstrap (block 5, 2,000 draws). ## Model Pace = a + b(t-2016) + c·D + d·D·(t-T), where D = 1 from the break year T. The key number is d, the slope change in min/km per year. A negative d means the record falls faster after T. ## Results (break at 2016) | Group | Slope before | Slope after | Slope change d | HAC 95% CI | Bootstrap 95% CI | |---|---|---|---|---|---| | Men | -0.00366 | -0.00701 | **-0.00335** | -0.00419 to -0.00252 | -0.00553 to -0.00128 | | Women (mixed-race) | -0.00838 | -0.01512 | **-0.00674** | -0.01030 to -0.00317 | -0.01523 to +0.00311 | All slopes are in min/km per year. Scale for men: d is about 0.2 s/km per year. Over the 10 years to 2026 that is about 0.034 min/km (2.0 s/km), or about 85 s over the full distance, relative to the pre-2016 line.Show 40 more lines
**Break at 2019 (ratification date of the Kosgei record).** - Men: d = -0.00305 (HAC -0.00435 to -0.00174; bootstrap -0.00591 to -0.00014). - Women: d = -0.01041 (HAC -0.01476 to -0.00605; bootstrap -0.02368 to +0.00377). **Model comparison, 2016 break against a linear trend with no break.** - Men: AIC -447.5 vs -419.9. Leave-one-out RMSE 0.0086 vs 0.0116 min/km. - Women: AIC -302.4 vs -297.8. Leave-one-out RMSE 0.0400 vs 0.0427. This is a small gain. - For the 2019 break, women gain nothing (AIC -297.2). **Placebo test.** I fitted the same slope break at each year from 1990 to 2010 (21 years), on the same 1980 to 2026 window. - Men: the 2016 break is more negative than all 21 placebos. The placebo median is -0.00234, so most placebos also point downward. The 2016 rank is 1 of 22. - Women: the 2016 break is more negative than all 21 placebos. The placebo median is +0.0039, so the women's placebo breaks are mostly positive. The rank is 1 of 22. - With 22 candidate years, a rank of 1 corresponds to a permutation p of about 0.045 at best. These are not independent tests. ## Sensitivity checks that weaken the result - **Curvature.** A quadratic trend with no break has AIC -453.1 and leave-one-out RMSE 0.0083 for men. This beats the linear-break model (AIC -447.5). Adding a 2016 break to the quadratic gives d = -0.00172 (HAC -0.00345 to +0.0000036). The interval touches zero, so the men's break is model-dependent. - **Men's end date.** Without the 2026 record (data to 2025), d = -0.00276 (HAC -0.00376 to -0.00176). With data to 2023, d = -0.00355 (HAC -0.00484 to -0.00226). The 2026 record is on Wikipedia but I did not verify it against a primary source. - **Women's start year.** The women's sign flips if the window starts in 1970. Then d = +0.0020 (HAC -0.0048 to +0.0088). The women's series is short and uneven. It is flat from 2003 to 2019, then falls in steps. - **Women's quadratic.** The quadratic plus break model gives d = -0.0149 (HAC -0.0245 to -0.0054) and AIC -308.6. It is the best women's model by AIC, but it rests on 14 to 18 events. - **Level change.** The level-change term is not stable. Its bootstrap interval includes zero for men, and it is positive for women at 2016. ## What the record list cannot show - It cannot separate shoes from pacing teams, course choice (flat, fast courses and the Berlin and Chicago circuits), more or better-paid athletes, or changes in how the lists are kept. - Few points and autocorrelation make the intervals approximate. HAC is a poor fit with 47 correlated points. The bootstrap intervals are wider for that reason. - A record is the single best result among many. The list does not show how the top 10 or the median elite runner moved. - Wikipedia is a secondary record. The source column mixes IAAF, ARRS and press reports. ## Present view A faster fall in the men's record after 2016 is likely under a linear-trend model, and it stands out against the placebo years. A curved trend explains most of it, and a record list cannot name a cause. For women, the data support no firm break estimate. What would change my view: top-10 or median-elite times by year from a public results database, and a primary-source list. A trial of shoes in the same runners would test shoes directly. ## Files - /work/parsed_records.csv, /work/events_*.csv, /work/results.json, /work/break_table.csv - Scripts: /work/scripts/parse.py, analyse.py, sens_fig.py - No app published.   [Break table CSV](/media/2026/10/a2becfbdf3b483705656df16c1c48cc1cad7c883f138a9b10568461af6ab1d83.csv)