Hypothesis

Most complex-trait GWAS variants act by changing the baseline expression of nearby genes, so eQTL colocalization and related methods should recover the known causal genes.

What failed

Across 220 gene-trait pairs where the causal gene is known from protein-coding variants, baseline-expression methods implicated few of those genes: colocalization 8%, transcriptome-wide association 2%, and regulatory annotations plus distance 4%, despite eQTLs lying near most GWAS hits.

Broken assumption

Trait-associated non-coding variants mostly coincide with homeostatic expression QTLs measured in standard datasets.

Why

The authors argue the results point to a deficit in the assumed model itself rather than just limited data, and call the gap 'missing regulation'.

From the published paper: Connally et al. eLife, 2022. ↗CC BY

Hypothesis

Scale-free (power-law degree) structure is a near-universal property of real-world networks, with broad consequences for the structure and dynamics of complex systems.

What failed

Applying statistical tests across nearly 1,000 social, biological, technological, transportation and information networks, strongly scale-free structure was empirically rare: for most networks log-normal distributions fit as well or better than power laws, and social networks were at best weakly scale-free.

Broken assumption

Power-law degree distributions are a universal signature of real-world networks.

Why

The authors organise the competing definitions of 'scale-free' and show the prevalence claim does not survive severe testing; a handful of technological and biological networks do appear strongly scale-free, so they call for new theory for the non-scale-free majority.

From the published paper: Broido et al. Nature Communications, 2019. ↗CC BY

Hypothesis

Adding partial endothelial trepanation to anterior lamellar keratoplasty would be at least as safe as big-bubble deep anterior lamellar keratoplasty, with fewer conversions to a full-thickness corneal graft.

What failed

Only 14 eyes of 14 patients were enrolled. Conversion to full thickness happened in 5 of 13 surgeries (3 DALK, 2 PET; p = .592), and no secondary outcome differed between groups. The trial was terminated for low inclusion and lack of equipoise, leaving the primary outcome underpowered.

Broken assumption

Partial endothelial trepanation makes anterior lamellar keratoplasty safer against full-thickness conversion.

Why

The authors state no conclusions can be drawn on the primary outcome from this underpowered trial, but that the PET technique was not as safe as expected; they report the results as an ethical obligation.

From the published paper: Wisse et al. Journal of Trial and Error, 2025. ↗CC BY

Hypothesis

Eight weeks of MBSR would change grey matter volume, density or cortical thickness, as widely cited earlier studies reported.

What failed

Pooling two three-arm randomised trials (218 meditation-naive participants: 75 MBSR, 73 matched active control, 70 waitlist), MRI showed no MBSR-specific structural change against either control, at the whole-brain level or in regions taken from prior MBSR studies.

Broken assumption

A short mindfulness course produces measurable structural brain change beyond what an equivalent active program does.

Why

The authors point to major methodological limitations in the earlier studies; with a larger sample and an active control, the reported neuroplastic effects did not appear.

From the published paper: Kral et al. Science Advances, 2022. ↗CC BY

Hypothesis

The honeybee’s roughly 20% higher mutation rate follows from its exceptionally high recombination rate, so the closely related bumblebee, with far less recombination, should mutate measurably less.

What failed

The bumblebee’s crossover rate was confirmed to be much lower (8.7 against 37 cM/Mb), yet its directly measured mutation rate was 3.6 x 10^-9 per haploid genome per generation (95% CI 2.38-5.37 x 10^-9), about 5% above the honeybee’s rather than lower, with no significant difference between species.

Broken assumption

A species’ recombination rate sets its genome-wide mutation rate.

Why

The authors do find a direct coupling between recombination and mutation within the genome, but so weak that it leaves almost no footprint on between-species differences.

From the published paper: Liu et al. Molecular Biology and Evolution, 2016. ↗CC BY

Hypothesis

Replaying newly learned foreign words during sleep would boost recall in 11–13-year-olds even more than in adults, because slow-oscillation activity peaks in adolescence.

What failed

Words cued by loudspeaker during post-learning NREM sleep were not recalled better the next morning than uncued words. Frontal theta power rose with successful cueing as in adults, but the expected spindle-power increase did not appear.

Broken assumption

More slow-oscillation activity during sleep means cued memory reactivation works better.

Why

The paper does not settle the cause; it discusses adolescents' learning abilities and pre-sleep performance levels as possible reasons for the missing effect.

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

Hypothesis

India's 2008 RSBY insurance scheme for poor households would reduce their out-of-pocket health spending.

What failed

Difference-in-differences on three national survey waves (N = 346,615) found no effect on the likelihood, level, or catastrophic share of inpatient spending, nor on outpatient spending. The chance of any out-of-pocket spending actually rose by 30%, and non-medical spending rose 5%.

Broken assumption

Enrolling poor households in hospital insurance lowers what they pay out of pocket for care.

Why

The abstract does not give a mechanism; it concludes the scheme was ineffective at reducing the spending burden on poor households.

From the published paper: Karan et al. Social Science & Medicine, 2017. ↗CC BY

Hypothesis

Combining a motivational component (mental simulation) with a volitional one (implementation intentions) would increase regular physical activity.

What failed

Neither trial found any significant main or interactive effect on physical activity: a cluster-randomised study of 92 students with self-reported activity at four weeks, and a 2 x 2 trial of 184 fitness-centre members with attendance records at four and nineteen weeks.

Broken assumption

Self-regulation techniques that worked in earlier studies transfer to sustained real-world physical activity.

Why

The authors emphasise that they followed protocols consistent with previous research, used an objective behavioural measure and had sufficient statistical power, and so suggest the components may not change physical activity and that format and moderators need testing.

From the published paper: Meslot et al. Frontiers in Psychology, 2016. ↗CC BY

Hypothesis

Adding avelumab to three cycles of neoadjuvant platinum chemotherapy would produce enough radiologic responses in resectable stage IB–IIIA NSCLC to justify a full phase 2 study.

What failed

Of 15 patients in stage 1, four had a radiologic response (one complete), fewer than the six required to continue, so the study was terminated. The regimen was tolerable (26.7% grade III/IV toxicity) without extra surgical complications.

Broken assumption

The response gains from adding checkpoint inhibitors to chemotherapy in advanced NSCLC carry over to the neoadjuvant setting.

Why

The paper reports only that adding immunotherapy did not appear to raise the response rate in this setting, and so the study failed its primary endpoint.

From the published paper: Tfayli et al. Cancer Medicine, 2020. ↗CC BY