Hypothesis

Because heavy precipitation intensifies at higher temperatures, flood-producing streamflow extremes in non-urban catchments would rise with temperature in the same way.

What failed

Using global catchment datasets, there was little sign that heavier rainfall at higher temperatures led to matching increases in streamflow; most regions showed streamflow extremes decreasing with higher temperature. Only for the rarest events in small catchments did the two scale together.

Broken assumption

More intense extreme rainfall in a warmer climate translates directly into more intense flooding.

Why

Catchment conditions such as initial soil moisture dampen the streamflow response; these matter less only as events get more extreme and catchments smaller.

From the published paper: Wasko et al. Scientific Reports, 2017. ↗CC BY

Hypothesis

Sentiment extracted from news headlines would carry predictive power for stock market direction and returns that could be exploited.

What failed

Across about 1.86 million headlines, explicit sentiment scores, even split into firm-specific and non-firm-specific parts, lacked robust predictive power; neither FinBERT sentiment, past returns nor implied volatility gave a consistently exploitable edge. Only weekend and holiday sentiment carried a modest signal.

Broken assumption

Public news sentiment contains return-relevant information that prices have not already absorbed.

Why

The authors note most news content is neutral or objective, and that markets appear to anticipate rather than react; they read the results as support for weak- and semi-strong-form market efficiency.

From the published paper: Davidović et al. Journal of risk and financial management, 2025. ↗CC BY

Hypothesis

Patient-reported outcome measures reflect objective functional recovery after surgical repair of an acute Achilles tendon rupture, as measured by heel-rise limb symmetry.

What failed

In 76 men followed a median of 41 months, median ATRS was 97 and median AOFAS 98, yet median limb symmetry was 87.5% and only 40.8% reached the 90% threshold; neither score correlated with symmetry (rho = 0.190, p = 0.100; rho = 0.218, p = 0.058), and the scores did not differ between the high- and low-symmetry groups even though heel-rise repetitions did.

Broken assumption

High patient-reported scores mean objective function has returned.

Why

The authors conclude the questionnaires do not capture persistent limb-symmetry deficits and should be read alongside performance-based tests.

From the published paper: Doğruöz et al. Journal of Clinical Medicine, 2026. ↗CC BY

Hypothesis

Women with more individual resources (such as education, work or assets) would be protected against intimate partner violence.

What failed

In DHS data on 580,000 women across 30 sub-Saharan African countries, individual resources showed no protective association with abuse, although they were linked to lower acceptance of it. Resource inequality was associated with more abuse, and employed women faced higher risk in communities that accepted wife-beating more.

Broken assumption

Giving a woman more resources raises her bargaining power enough to protect her from partner violence.

Why

The authors point to community attitudes: where wife-beating is more accepted, women's employment goes with greater risk, so individual resources do not translate into protection.

From the published paper: Cools et al. World Development, 2017. ↗CC BY

Hypothesis

Because metformin dampens inflammation, adding it to 14 weeks of supervised resistance training would augment the muscle response in healthy adults aged 65 and over.

What failed

Among 94 randomized participants the placebo group gained more lean body mass (p = .003) and thigh muscle mass (p < .001), and more thigh muscle area (p = .005) and density (p = .020), than the metformin group, with strength gains trending lower on metformin; fibre hypertrophy and satellite-cell and macrophage abundance were unaffected.

Broken assumption

Suppressing training-related inflammation helps rather than hinders muscle hypertrophy.

Why

The authors report that metformin raised AMPK signalling and tended to blunt the training-induced rise in mTORC1 signalling, and that placebo but not metformin lowered the type I fibre percentage.

From the published paper: Walton et al. Aging Cell, 2019. ↗CC BY

Hypothesis

Children lose substantial reading and math skills over the summer, and achievement gaps grow faster in summer than during the school year, consistently across recent large datasets.

What failed

Comparing three recent U.S. elementary-school datasets from the 2010s (ECLS-K:2011, NWEA and Renaissance), most patterns did not generalize beyond a single test: summer losses, gap growth by income, ethnicity and gender, and growth in score variance each appeared on some tests and not others, and where gaps grew they did not consistently grow faster in summer.

Broken assumption

Summer learning loss and faster summer gap growth are robust patterns that hold across achievement tests.

Why

The paper reports that the patterns depend on which test is used; it does not identify a single cause and recommends showing that a summer-learning pattern replicates before drawing broad conclusions.

From the published paper: Workman et al. Sociological Science, 2023. ↗CC BY

Hypothesis

Slow, affective stroking would strengthen ownership of a virtual body more than fast stroking, as it does for the rubber hand illusion.

What failed

A first study found more subjective ownership with slow (3 cm/s) than fast (30 cm/s) stroking. A second study again found the usual synchronous-versus-asynchronous effect but did not reproduce the slow-versus-fast difference, so the authors could not conclude that affective touch enhances the full-body illusion.

Broken assumption

The affective-touch boost seen for a rubber hand transfers reliably to ownership of a whole virtual body.

Why

The abstract does not explain the discrepancy between the two studies and calls for further research.

From the published paper: Jong et al. Experimental Brain Research, 2017. ↗CC BY

Hypothesis

A smaller GBLUP model restricted to genotyped animals, but carrying information from non-genotyped relatives, could replace full single-step GBLUP and cut the size and computational cost of livestock genetic evaluations.

What failed

Absorbing the equations of non-genotyped animals into those of genotyped animals did give solutions equivalent to full ssGBLUP, but the authors judge the model impractical: it pushes pedigree-based additive covariances into the residual covariance matrix, making it dense and hard to invert, and non-genotyped animals' breeding values cannot easily be back-solved from the genotyped ones. The alternative pseudo-phenotype route was itself hard to compute.

Broken assumption

An algebraically equivalent smaller model is also computationally cheaper to run.

Why

The residual covariance matrix stops being diagonal and its inverse is dense, which raises the cost of every downstream multiplication; the conditional structure needed for back-solving is lost.

From the published paper: Nilforooshan et al. Journal of Trial and Error, 2026. ↗CC BY

Hypothesis

Carryover from trial to trial — larger congruency effects after a congruent trial than after an incongruent one — would explain why threat congruency effects in the dot-probe task are small and unreliable.

What failed

In two large samples (n = 241 and n = 82) there was a small congruency effect for fearful faces (d = 0.15) and a reverse effect for threatening images (d = 0.11), but no carryover effects of threat in either experiment; Bayesian analyses gave moderate to strong support for the null.

Broken assumption

Trial-to-trial carryover explains the weak and unreliable threat congruency effects in the dot-probe task.

Why

The authors conclude carryover does not influence attentional bias to threat, so it does not account for the small effects and poor reliability reported for the task.

From the published paper: Maxwell et al. Journal of Trial and Error, 2022. ↗CC BY