Hypothesis

Longer commutes would be associated with lower subjective well-being, as earlier studies had reported.

What failed

Using a range of fixed-effects ordered estimators, the authors found no general evidence that longer commutes go with lower well-being, unlike earlier papers.

Broken assumption

Commuting longer reliably lowers people's life satisfaction.

Why

The abstract attributes nothing specific; it is mainly a comparison of estimators, and reports that linear and ordered fixed-effects models agree.

From the published paper: Dickerson et al. Regional Science and Urban Economics, 2014. ↗CC BY

Hypothesis

Calmangafodipir, which reduced patient-reported chemotherapy-induced peripheral neuropathy in a randomised phase 2 trial, would prevent moderate-to-severe neuropathy in patients receiving oxaliplatin in the adjuvant and metastatic settings.

What failed

Both POLAR trials (n = 301 and 291) stopped early after unexpected hypersensitivity reactions. In the pooled primary analysis, 54.3% of calmangafodipir patients had moderate-to-severe neuropathy at month 9 versus 40.3% on placebo (relative risk 1.37, 95% CI 1.01–1.86, P = .045), and serious hypersensitivity reactions were more frequent (3.6% vs 0.8%).

Broken assumption

Targeting oxidative stress prevents oxaliplatin-induced peripheral neuropathy, as a phase 2 trial suggested.

Why

Not only was there no benefit, the drug arm did worse on the primary endpoint. The authors frame the result as showing the difficulty of preventing this neuropathy by targeting oxidative stress.

From the published paper: Pfeiffer et al. JNCI Cancer Spectrum, 2022. ↗CC BY

Hypothesis

If pheromone detection in the hawkmoth runs through an ionotropic OR-Orco channel, then opening the Orco pore with an agonist should add to the pheromone-evoked response in olfactory receptor neurons.

What failed

In tip recordings from intact pheromone-sensitive sensilla, perfusing the Orco agonist VUAA1 did not boost pheromone responses within the first 1000 ms, although it raised the neurons' spontaneous firing in a dose- and time-of-day-dependent way.

Broken assumption

Insect odorant receptor-Orco complexes transduce pheromone signals as fast ionotropic channels in all species.

Why

The authors propose that in M. sexta Orco acts as a slower, second-messenger-dependent pacemaker channel that sets calcium baseline, rather than as the fast pheromone-gated pore.

From the published paper: Nolte et al. PLoS ONE, 2013. ↗CC BY

Hypothesis

Real-world objects, unlike simple coloured squares, benefit from extra encoding time and escape the usual three-to-four-item working-memory limit, as shown behaviourally and in the contralateral delay activity (CDA) by Brady et al. (2016).

What failed

Three replications with more participants and more trials per condition did not reproduce the key result that objects gain more than colours from longer viewing, and found no object-specific CDA increase. Extra encoding time helped both colours and objects.

Broken assumption

Familiar, complex objects have a special working-memory capacity that longer encoding unlocks.

Why

The abstract offers no account of the original result; the data simply give no support for a different capacity for real-world objects.

From the published paper: Quirk et al. eNeuro, 2020. ↗CC BY

Hypothesis

Better e-service quality and greater e-trust would raise online-shop customers' satisfaction and loyalty, and satisfaction would in turn raise loyalty.

What failed

In a PLS structural equation model fitted to 432 online-shop customers in Banten, Indonesia, every hypothesised path was non-significant: e-service quality and e-trust on satisfaction, both on loyalty, and satisfaction on loyalty.

Broken assumption

E-service quality and trust reliably drive online customer satisfaction and loyalty.

Why

The paper reports the non-significant coefficients and their signs without identifying a cause.

From the published paper: Juwaini et al. International Journal of Data and Network Science, 2022. ↗CC BY

Hypothesis

The standard differential-expression tools control the false discovery rate on population-level human RNA-seq data as they do on small controlled designs.

What failed

Permuting condition labels in an immunotherapy RNA-seq dataset made DESeq2 and edgeR report more differentially expressed genes than the real labels did. Across population-level datasets, DESeq2, edgeR, limma-voom and the non-parametric dearseq all failed to control the false discovery rate, which for DESeq2 and edgeR sometimes exceeded 20% at a 5% target. NOISeq controlled it better but had much lower power than the Wilcoxon rank-sum test.

Broken assumption

The parametric models behind DESeq2 and edgeR hold for large human population cohorts.

Why

The authors trace the failure to the parametric assumptions these tools make, which population-scale samples violate, and recommend the Wilcoxon rank-sum test instead.

From the published paper: Li et al. Genome biology, 2022. ↗CC BY

Hypothesis

Heavier instructional scaffolding, a goal-context-constraints prompting template with iterative refinement and reflective guidance, would improve learning outcomes and metacognitive processes in AI-tutored learning over lighter or absent scaffolding.

What failed

Across 175 first-semester students split between full, light and no scaffolding in the same four-hour session, knowledge, critical thinking and reflective use all rose and cognitive load fell, but no significant differences appeared between scaffolding conditions and the conditions produced no significant interaction effects; only descriptive trends favoured the scaffolded groups on higher-order knowledge.

Broken assumption

More prompting scaffolding yields more learning in AI-supported sessions.

Why

The authors suggest short-term gains depend on how learners engage with the AI rather than on scaffolding intensity alone.

From the published paper: Melanou et al. Education Sciences, 2026. ↗CC BY

Hypothesis

Self-supervised learning, which cuts reliance on labels for large balanced datasets like ImageNet, would also beat ordinary supervised learning on the small, class-imbalanced datasets typical of medical imaging.

What failed

Across four binary tasks (Alzheimer's diagnosis and age prediction from brain MRI, pneumonia from chest X-rays, choroidal neovascularization from OCT; mean training sets of 771 to 33,484 images), supervised learning beat the chosen self-supervised paradigms in most small-training-set experiments, even when only part of the data was labelled.

Broken assumption

Self-supervised pretraining gains seen on large, balanced natural-image datasets carry over to small, imbalanced medical datasets.

Why

The abstract does not give a mechanism; the authors conclude that the choice of paradigm depends on training set size, label availability and class balance.

From the published paper: Espis et al. Scientific Reports, 2025. ↗CC BY

Hypothesis

T cells reacting to SARS-CoV-2 spike protein in people never exposed to the virus would be explained by near-identical epitopes shared with common human pathogens, especially endemic coronaviruses.

What failed

Epitope prediction combined with similarity searches against common human pathogens found that the SARS-CoV-2 epitopes most similar to beta-coronavirus ones sit in replication-associated proteins, not in the host-interacting spike protein.

Broken assumption

Cross-reactive T cell memory to a new virus comes from near-identical epitopes in related pathogens.

Why

On this computational evidence the authors conclude near-identical epitopes do not drive the observed pre-formed immunity to structural proteins; what does is left open.

From the published paper: Stervbo et al. Scientific Reports, 2020. ↗CC BY