Preprints
Papers posted by their authors, each published together with its CheckPaper review.
- Daily Maximum Temperature and Pedestrian and Cyclist Deaths in the United States, 2000–2024: A Pre-registered County-Day Panel AnalysisHot weather is associated with more road traffic deaths, but national evidence for people walking and cycling in the United States, and on which age group is most affected, is limited. All deaths recorded in the Fatality Analysis Reporting System (FARS) for 2000–2024 were linked to county daily maximum temperature (Tmax) from NOAA nClimGrid-Daily, giving 28,372,440 county-days in 3,107 contiguous US counties. Following an analysis plan time-stamped before any crash record was obtained, we fitted Poisson pseudo-maximum-likelihood models with county-by-year-by-month, day-of-week and holiday fixed effects, precipitation terms and Tmax bins at lags 0–3. Standard errors were clustered by state and date. The cumulative rate ratio (RR) for Tmax of at least 90 °F versus 60–70 °F was 1.010 (95% CI 0.967–1.055) for 135,546 pedestrian deaths and 1.143 (1.026–1.273; Holm-adjusted p = 0.093) for 19,965 cyclist deaths. For 679,655 vehicle-occupant deaths it was 1.032 (1.008–1.057). RRs did not differ between children (0–14 years), adults (15–64) and older people (65+) among pedestrians (p = 0.64) or cyclists (p = 0.92), and neither ratio to the occupant RR was significant after Holm adjustment. A pre-specified falsification test with next-day temperature was positive for all three groups, consistent with a one-day offset in how the temperature grids are dated; exploratory models aligned with that dating gave larger associations. Hot days were associated with a modest rise in cyclist deaths, no detectable change in pedestrian deaths and no age group at distinctly higher relative risk.4/10Major revision
- Reproducibility of the Association Between Daily Step Count and All-Cause MortalityKey Points. Question: Whether the published association between daily step count and all-cause mortality in US adults is reproduced when the analysis is repeated in public survey data under a pre-registered plan. Findings: In this cohort study of 2378 adults aged 40 years or older, 8000 compared with 4000 steps per day was associated with a hazard ratio for death of 0.43, compared with the published 0.49; both primary estimates met the pre-specified replication criteria. Meaning: The association between more daily steps and lower mortality was reproduced in direction, shape and approximate magnitude. Importance: An influential estimate of the association between daily steps and mortality has not been independently re-analyzed. Objective: To repeat the primary analysis of Saint-Maurice et al (2020) in public data and to examine sources of any difference. Design, Setting, and Participants: Pre-registered replication of a prospective cohort analysis in the National Health and Nutrition Examination Survey 2005-2006, including adults aged 40 years or older with valid accelerometer data, linked to the public-use National Death Index file. The analysis plan and replication criteria were locked before mortality data were opened. Exposure: Median daily step count across valid days of hip-worn accelerometry. Main Outcomes and Measures: All-cause mortality through 2015, with cardiovascular and cancer mortality as secondary outcomes. Replication required the same direction, a dose-response pattern and overlapping 95% CIs for 8000 and 12 000 vs 4000 steps per day. Results: Of 2378 participants (weighted mean age, 56.9 years; 54.1% women), 484 died during a weighted mean follow-up of 9.38 years. In the fully adjusted model, the hazard ratio for 8000 vs 4000 steps per day was 0.43 (95% CI, 0.35-0.52), compared with the published 0.49 (95% CI, 0.44-0.55), and for 12 000 vs 4000 steps per day 0.30 (95% CI, 0.19-0.46), compared with 0.35 (95% CI, 0.28-0.45). All 17 published estimates with a replicated counterpart agreed in direction with overlapping CIs; 11 replicated estimates lay inside the published CIs. Alternative analytic choices gave estimates of 0.42 to 0.44, whereas follow-up through 2019 gave 0.53 (95% CI, 0.44-0.64). Conclusions and Relevance: The association between daily step count and mortality was reproduced in direction and dose-response shape, with a somewhat stronger association that is compatible with sampling variation and an association concentrated early in follow-up. Because the design is observational, these estimates describe an association rather than the effect of increasing daily steps.5/10Major revision
- Upward Mobility across U.S. Commuting Zones: A Pre-Registered Replication of Chetty, Hendren, Kline and Saez (2014)Chetty, Hendren, Kline and Saez (2014) reported that the chances of children from low-income families moving up the income distribution differ widely across U.S. commuting zones. They also reported that places with more upward mobility have less segregation, less income inequality, better schools, more social capital and more stable families. This replication used the authors' public commuting-zone files and an analysis plan, with numerical tolerances, that was fixed and time-stamped before any outcome statistic was computed. Two outcomes were examined. The first is absolute upward mobility, the expected income rank of children whose parents are at the 25th national percentile; the second is the probability that a child from the bottom income quintile reaches the top quintile. Of 479 published quantities that could be graded, 471 were reproduced within the pre-set tolerances, 8 approximately, and none failed. The spread of upward mobility across areas (standard deviation 5.68 rank points), the ranking of the 50 largest areas, all 20 correlations of the five main factor groups and the five-factor regression (R2 = 0.757) matched the published values. Decile maps show the same geography, with high mobility in the Great Plains, low mobility in the Southeast and the industrial Midwest. New pre-registered estimates show that the probability of reaching the top quintile has the same correlate pattern (19 of 19 signs agree; rank correlation 0.985), with coefficients about one-tenth smaller. These relationships are associations across places, not causal effects.5/10Major revision
- Egg intake and all-cause and cardiovascular mortality in US men and women: prospective cohort studyObjective: To estimate the dose-response association of egg intake with all-cause and cardiovascular mortality in US men and women, and to test whether it differs by sex. Design: Prospective cohort study with an analysis plan time-stamped before mortality data were obtained. Setting: National Health and Nutrition Examination Survey (NHANES) 1999–2018, linked to the National Death Index through 2019. Participants: 45,304 adults aged 20 years or older (22,094 men, 23,210 women) with a reliable 24-hour dietary recall. Main outcome measures: All-cause mortality, and cardiovascular mortality in the 1999–2014 cycles. Survey-weighted Cox models with restricted cubic splines of egg intake were adjusted for sociodemographic factors, energy intake, diet quality without eggs, smoking, alcohol, physical activity, body-mass index and cardiometabolic conditions. Results: Over a median of 9.2 years, 7,315 participants died, 2,222 of cardiovascular causes in the 1999–2014 cycles. No primary test remained below 0.05 after Holm adjustment. In men, the hazard ratio for all-cause mortality was 1.08 (95% confidence interval 0.96 to 1.22) at 1 egg/day and 1.17 (1.02 to 1.35) at 2 eggs/day versus none; this gradient was confined to the first 10 years and weakened after adjustment for dietary cholesterol. In women, the hazard ratio was 0.86 (0.76 to 0.97) at 1 egg/day, a non-linear association sensitive to exposure definition and follow-up time. Cardiovascular mortality was not associated with egg intake (1 egg/day: men 0.99, 0.82 to 1.20; women 0.96, 0.78 to 1.18), but a 10% higher hazard could not be excluded from about 0.3 eggs/day. Evidence for a sex difference was weak (2-df interaction P=0.15; linear P=0.046). Conclusions: In this observational study, egg intake was not robustly associated with mortality, and associations do not establish causation. No intake was shown to be safe. In men, intakes of about two eggs a day or more were associated with modestly higher all-cause mortality, mainly in the first decade of follow-up; in women, no higher mortality was observed within the range of intake studied.5/10Major revision
- Daily step count, a plateau in all-cause mortality, and differences between women and men: prospective cohort study using NHANES accelerometryObjectives To estimate whether, and at what daily step count, higher step counts stop being associated with lower all-cause mortality, and whether this differs between women and men. Design Pre-registered prospective cohort study. Setting US National Health and Nutrition Examination Survey (NHANES), with accelerometry in 2005–2006 and mortality linkage to 31 December 2019. Participants 2373 adults aged 40 years or older with hip-worn accelerometry (1187 women, 1186 men); replication in 4917 adults from 2003–2006 and 5651 adults with wrist-worn accelerometry in 2011–2014. Main outcome measures All-cause mortality, from design-based Cox models with restricted cubic splines. Plateau location was estimated as a hinge breakpoint and as the step count reaching 90% of the maximal risk reduction, with bootstrap confidence intervals, and classified with a pre-specified decision rule. Results Over a median of 13.6 years, 702 participants died. Relative to 4000 steps/day, hazard ratios were 0.52 (95% confidence interval 0.43 to 0.63) at 8000, 0.43 (0.35 to 0.52) at 10000 and 0.39 (0.29 to 0.52) at 12000 steps/day. Non-linearity was nominally significant (P=0.035; multiplicity-adjusted P=0.070), and the plateau was classified as present but imprecisely located: breakpoint 6980 steps/day (4900 to 13400), 90% point 9920 (6910 to 18380). The curve did not differ by sex (P=0.72), but the women minus men breakpoint difference (−5130, −8800 to 5950) did not establish equivalence. Ten-year absolute risk at 4000 versus 10000 steps/day was 17.8% versus 8.5% in women and 21.8% versus 11.0% in men. Both replication cohorts gave the same classification without sex interaction; excluding early deaths or participants in poor health removed the plateau. Conclusions Most of the mortality difference associated with daily steps lay below about 8000 to 10000 hip-measured steps/day, but the plateau location was imprecise and sensitive to reverse causation. No difference between women and men was detected. Keywords: step count; all-cause mortality; dose-response; accelerometry; sex differences; NHANES5/10Major revision
- Short Sleep, Depression and Their Association Among US Adults Before and After the COVID-19 Pandemic: A Pre-registered Repeated Cross-sectional Analysis of NHANES 2005–2023Depression among US adults rose after the COVID-19 pandemic, but no design-based national estimate shows whether short sleep changed over the same period or whether the association between short sleep and depression differed. We conducted a pre-registered, survey-weighted repeated cross-sectional analysis of 42,222 adults aged 20 and over in eight National Health and Nutrition Examination Survey cycles from 2005–2006 to August 2021–August 2023. Short sleep was defined as less than 7 hours of usual workday sleep, analysed as two series because the sleep item changed in 2015, and depression as a Patient Health Questionnaire-9 score of 10 or more. Age-adjusted prevalences, adjusted prevalence ratios from survey-weighted Poisson models and their differences between 2017–March 2020 and August 2021–August 2023 were estimated overall and by sex and age group, with Holm control across three primary hypotheses. The age-adjusted prevalence of depression was 12.6% in 2021–2023, 4.1 percentage points (95% CI 2.2 to 6.0) higher than in 2017–March 2020 and the highest since 2005–2006, with the largest difference at ages 20–39 (7.5 percentage points). Short sleep was 3.2 percentage points lower (95% CI −5.2 to −1.3), at 20.9%, mainly among men. Short sleep was associated with depression in every cycle (adjusted prevalence ratio 1.71 in 2021–2023 and 1.41 in 2017–March 2020); the ratio of these ratios, 1.21 (95% CI 0.89 to 1.63), was too imprecise to establish or exclude a change. Because the depression screener changed mode, response fell and the data are cross-sectional, these differences describe populations at two periods rather than effects of the pandemic.5/10Major revision
- Site-specific Cancer Burden in China and the United States, 1990–2023: A Pre-registered Decomposition of Between-country Differences into Modifiable Risk FactorsBackground. China and the United States have similar overall cancer mortality but different leading cancers. Earlier comparisons placed each country’s risk-attributable fractions side by side without splitting the absolute between-country gap into risk components. Methods. In a pre-registered descriptive secondary analysis, age-standardized incidence, mortality and DALY rates for stomach, liver, oesophageal, lung, colorectal, breast and prostate cancer were compared from 1990 to 2023 using GBD 2023. GLOBOCAN 2022 and 2024 and IARC Cancer Causes estimates provided breakdowns by sex, infection-attributable incidence and a cross-source check. The China minus United States difference in age-standardized mortality was split into components attributable to GBD risk clusters by Shapley allocation of the published joint attributable fraction, and a component not attributed to GBD-assessed risks. Results. In 2023 all-cancer mortality was 112.4 per 100 000 in China and 114.2 in the United States. Mortality was higher in China for stomach (16.5 vs 2.8), oesophageal (10.9 vs 3.5), liver (6.4 vs 4.3) and lung cancer (32.9 vs 25.7), and lower for colorectal, breast and prostate cancer. Since 1990 the stomach and oesophageal differences narrowed by about 70%, and the lung difference changed from −13.1 to +7.2. Tobacco (5.2) and air pollution (4.2) components exceeded the lung gap; tobacco accounted for 48% of the oesophageal gap; 75% of the stomach gap and most of the breast and prostate gaps were not attributed to GBD-assessed risks. IARC infection-attributable incidence was 29.5 per 100 000 in China and 15.8 in the United States, mainly from H. pylori and hepatitis B virus. Stomach, liver, oesophageal and lung cancer differences were larger among males. Liver and colorectal ratios depended on the data source. Conclusions. The lung and oesophageal contrasts align with tobacco and air pollution, whereas most of the stomach, breast and prostate contrasts lie outside the GBD risk set. The components are accounting quantities without uncertainty intervals, for both sexes combined.5/10Major revision
- Parental Income, Occupation and Education and Children’s Outcomes in China and the United States: A Pre-Registered Cross-Database ComparisonThis pre-registered study measures how strongly parents’ income, occupation and education predict their children’s outcomes in China and the United States. It compares published estimates based on the China Family Panel Studies and other Chinese surveys with public US data from Opportunity Insights, the General Social Survey and the World Bank’s intergenerational mobility database. No Chinese microdata were analysed. Each China–US pair receives a comparability grade before any difference is computed. On the closest income measures, the rank-rank slope is 0.494 in China (children born 1981–1988) and 0.287 in the United States (1978–1983), a difference of 0.207 (95% confidence interval 0.190 to 0.225). Intergenerational income elasticities reverse order between the headline estimates and a common father–son framework. On identical schooling measures, persistence is higher in China for the 1980s birth cohort (difference 0.154) and lower for cohorts born before 1970. Movement out of farming depended more on parental farm origin in rural China in 2002 than among rural-origin US respondents. Within China, persistence is higher for daughters than sons and for rural than urban hukou holders; US metro and nonmetro slopes differ by 0.019. Cross-country differences in intergenerational persistence depend on how each measure is defined.4/10Major revision
- Off-Hours and Weekend Pushes on GitHub Before and After AI Coding Assistants: A Pre-Registered Interrupted Time Series, 2018–2026AI coding assistants became mainstream in 2022, and it is unclear whether they changed when developers work. This pre-registered study followed a random sample of 4,873 developers from January 2018 to July 2026. Their public GitHub pushes were converted to local time from profile locations and commit time-zone offsets. Two shares were measured, pushes made outside 09:00–18:00 local time and pushes made on a local weekend. An interrupted time series around the launch window of GitHub Copilot and ChatGPT, with pandemic, composition and platform controls fixed in advance, compared each month after December 2022 with the extrapolated 2018–2022 trend. The off-hours share lay 0.85 percentage points below its trend (95% confidence interval −2.27 to 0.57) and the weekend share 1.25 points below it (−2.77 to 0.26). Both results are inconclusive, because neither departure differs from zero, and neither is precise enough to show equivalence within 1 point. Increases of the size seen during the COVID-19 pandemic, when both shares rose by about 3 points, are not consistent with the data. Within-developer estimates were close to zero, placebo breakpoints produced departures of similar size, and 26 pre-specified robustness checks did not alter this conclusion. Among regions with enough developers, only Eastern Asia changed, with a fall in the off-hours share that follows an unusually steep pandemic-era rise. Northern America, Northern and Western Europe and the United States showed no departure. Public push timing does not show a population-wide move into evenings or weekends after AI assistants spread.4/10Major revision
- Late-Night Play and Blunders in Online Chess: A Pre-Registered Within-Player Study of Local Time of Day on LichessA common belief among online players holds that chess played late at night is worse chess. Evidence on decision quality at night is sparse, however: the one within-player study of online chess compared mornings with evenings and did not examine the hours after midnight. This study tested, under a pre-registered analysis plan, whether the local hour of play is associated with blunders and with rating-adjusted results in the public Lichess game database. Six months of rated games from 2024 and 2025 were drawn for a hashed one-sixteenth sample of players, and local time was assigned from profile country flags, with hourly activity used to choose a zone within multi-zone countries. The analysis covered 14.1 million games by 17,084 players, of which 1.28 million carried engine evaluations. Hour effects were estimated within player and speed, with controls for session length, position in the session, breaks and preceding results, and were judged against pre-registered smallest effects of interest. Between 00:00 and 04:59 local time, the blunder rate was 0.20% of the day rate lower than between 10:00 and 17:59, and an increase of 5% or more was ruled out by an equivalence test. This held for four alternative blunder definitions, and no rating band or speed showed a significant late-night increase. Score relative to expectation was 2.26 Elo-equivalents lower at night. This association was concentrated in Bullet games and among early-active players, was sensitive to a one-hour shift of the window, and was absent when the session and prior-result controls were omitted. The rating-adjusted peak fell at 18:00–19:59, and no hour was clearly best for blunders. Evening moves were 1.1% faster than morning moves, without more evening blunders. Position within a session and runs of losses were associated with deficits of 3.5 to 9.6 Elo, several times the size of any clock-hour contrast. In online chess, the course of the current session is associated with much larger differences in results than the hour on the clock.5/10Major revision
- ChatGPT and the English-Language Gap in Science: A Pre-Registered Difference-in-Differences Study of 74,627 Authors in OpenAlexLarge language models were expected to narrow the disadvantage of researchers who write in English as a second language, but evidence comes mainly from studies of detected adopters. This study tests whether outcomes of authors in non-English-speaking countries changed relative to authors in six English-speaking countries after ChatGPT’s release in November 2022. A 1% random sample of OpenAlex author identifiers yields 74,627 authors followed from 2017 to 2025. A pre-registered difference-in-differences design with author and field-by-year fixed effects, inference by wild cluster bootstrap over countries and Holm correction compares 2024–2025 with 2017–2022 on four outcomes. Output rose by 22.5% relative to English-country authors, but the interval includes zero (Holm p = 0.396). Top-quartile journal share rose by 2.8 percentage points and log days to acceptance fell by 0.039; neither survives correction and both fail the pre-registered pre-trend check. International collaboration did not change detectably. Non-English-country authors had been converging on English-country authors before 2022, and a post hoc trend adjustment leaves little change. The data do not show that ChatGPT levelled the playing field in its first two years, although individual gains and small group effects cannot be excluded.5/10Major revision
- Temperature and Noise Complaints to 311 in New York City, 2010–2026: A Pre-Registered Time-Series AnalysisNoise complaints are a large share of requests to New York City's 311 service line, but their association with temperature has not been estimated with calendar control over the full public record. This pre-registered ecological time-series study related 8,632,969 daily-window and 4,306,668 night-window 311 noise reports from 2010 to October 2026 to Central Park temperature. Quasi-Poisson distributed lag non-linear models over lags 0–3 controlled for day of week, holidays, season, long-term trend and the pandemic period, with autocorrelation-consistent inference. Days at the 90th percentile of daily mean temperature had more reports than days at the median (rate ratio 1.134, 95% CI 1.066–1.205), and nights at the 90th percentile of night temperature had more reports than median nights (1.158, 1.076–1.246). Both curves peaked near the upper quartile of temperature and fell below the median level at the 99th percentile. Heat-wave nights under the primary definition showed no clear increment beyond the temperature curve (1.012, 0.944–1.085), and alternative definitions gave estimates on both sides of this value. The temperature association differed by complaint type, being largest for street and sidewalk, vehicle and air-conditioning equipment reports, and by borough, and it declined from the lowest-income to the highest-income areas. Twenty-two sensitivity analyses kept the direction of the primary estimates. Because 311 counts reflect both sound and the propensity to report, these associations describe reporting and do not measure sound or the conduct of any area or group.5/10Major revision
- Screen character names and US baby-name popularity, 1986–2025: a preregistered event study with matched control namesFilms and television series are often credited with making baby names popular, but US evidence rests on case curves for single names without comparison names. This preregistered stacked event study links the first names of main characters in 864 major US-released films and series (release years 1986–2020, from Wikidata and Wikipedia) to US Social Security Administration birth counts for 1980–2025, nationally and by state. Each of 1,332 names was compared with five control names matched on pre-release level and trend. Over the release year and the next three years, character names were on average 0.180 log points (19.7%) above their controls (95% CI 0.148 to 0.211), about 354,000 excess births concentrated in a few titles. Part of this gap predates release: the names stood about 0.10 above their controls in held-out lead years, and fake release dates eight years earlier produced a mean of 0.121. A post hoc comparison of each name’s true and fake dates attributes about 0.10 (0.05 to 0.14) to the release window. No name met the preregistered rule for being made popular, because placebo calibration could not resolve small enough p-values; a post hoc count suggests about 81 names above chance. The excess kept rising through ten years, so no decay half-life was identified. Girls’ names gained slightly more than boys’ names (difference 0.052, −0.016 to 0.113), films and series did not differ clearly, only action and adventure titles stood out among genres, and between-state variation was small and fragile. Screen releases coincide with lasting gains for character names, about half of which reflect trajectories the names were already on.4/10Major revision
- Hugging Face Daily Papers Listing and Later Citations of AI Preprints: A Pre-Registered Matched Cohort StudyCurated daily feeds now shape which machine-learning preprints researchers read, but whether a place on such a feed goes with later citations has not been examined against comparable papers. This pre-registered matched cohort study compared arXiv papers listed on Hugging Face Daily Papers with never-listed papers from the same primary category and release week, matched on propensity scores from arXiv features and adjusted for pre-release author prominence and institution. Citations were counted at 6, 12 and 24 months after the first arXiv version. At 24 months, the primary confirmatory test, listed papers had an adjusted geometric mean of citations plus one that was 203.3% (95% CI 184.5 to 223.4) higher than that of matched controls, based on 11,401 papers. The difference was 107.1% (95% CI 100.9 to 113.5) at 6 months and remained large in every sensitivity analysis. Among listed papers, citations rose with upvotes recorded seven days after listing, with an elasticity of 0.623 (95% CI 0.488 to 0.758) at 24 months. A pre-registered regression discontinuity at the top-1 and top-3 daily ranks was not feasible, because no cell reached the required number of close-vote days, and descriptive comparisons at these cutoffs showed no consistent rank advantage. A Daily Papers listing is a strong early marker of later citations, but the design cannot separate an attention effect of the listing from selection of papers that would have been cited more in any case.4/10Major revision