Uninformative non-reproduction: a pre-computable power frontier explains which driver–phenotype associations small cancer cohorts can and cannot test

Abstract Background Small cancer cohorts are routinely used to test driver–phenotype associations, and negative results from them enter the literature as evidence that no association exists. Reproducibility frameworks classify such results only as reproduced or not reproduced, conflating a genuine null finding with a study that could never have detected the effect. We asked whether this third possibility can be identified before data are collected. Methods We derived a closed-form detection frontier, d min (n,p) — the smallest standardised effect detectable at 80% power, two-sided α = 0.05, in a two-group comparison of total size n with alteration prevalence p — and calibrated it against empirical power by simulation across three data-generating mechanisms and four group-size configurations (60 points, 2 000 replicates each). We then applied the frontier to a complete scan of nine driver variables against eight expression signatures in TCGA malignant pleural mesothelioma ( n = 87), and to fifteen rare-tumour TCGA cohorts. Results Under ideal-normal data the closed form was accurate to 0.009; under skew and dropout contamination it over-estimated the detection limit by up to 0.086, erring on the safe side in all eight non-normal configurations. Six of fifteen cohorts could detect a medium effect ( d = 0.5) and none a small one ( d = 0.3); the latter would require 416 samples. Of 72 comparisons in the case study, seven (9.7%) had an observed effect at or above their own detection limit and all seven were significant; the remaining 65 fell below it, and all 59 non-significant comparisons lay below their own limit, so none constituted an informative null. A published large-effect association reproduced closely (CDKN2A deletion 41.4% versus 41.5%), whereas a reported medium-effect association could not be addressed: that comparison's detection limit was 0.705 against an observed effect of 0.011. Conclusions A non-significant driver–phenotype result in a small cohort may carry no evidence either way, and this is determinable from cohort size and driver prevalence before any data are collected. Reporting the detection limit alongside negative results would distinguish an underpowered comparison from a genuine absence of association.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11401252/v1
Primary Topic
Cancer Genomics and Diagnostics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Uninformative non-reproduction: a pre-computable power frontier explains which driver–phenotype associations small cancer cohorts can and cannot test

Sanjun Zhao, Xiayu Wu, Xueyi Lin
Research Square
Cancer Genomics and Diagnostics
preprint

Uninformative non-reproduction: a pre-computable power frontier explains which driver–phenotype associations small cancer cohorts can and cannot test

Sanjun Zhao, Xiayu Wu, Xueyi Lin
preprint en

Abstract

Abstract Background Small cancer cohorts are routinely used to test driver–phenotype associations, and negative results from them enter the literature as evidence that no association exists. Reproducibility frameworks classify such results only as reproduced or not reproduced, conflating a genuine null finding with a study that could never have detected the effect. We asked whether this third possibility can be identified before data are collected. Methods We derived a closed-form detection frontier, d min (n,p) — the smallest standardised effect detectable at 80% power, two-sided α = 0.05, in a two-group comparison of total size n with alteration prevalence p — and calibrated it against empirical power by simulation across three data-generating mechanisms and four group-size configurations (60 points, 2 000 replicates each). We then applied the frontier to a complete scan of nine driver variables against eight expression signatures in TCGA malignant pleural mesothelioma ( n = 87), and to fifteen rare-tumour TCGA cohorts. Results Under ideal-normal data the closed form was accurate to 0.009; under skew and dropout contamination it over-estimated the detection limit by up to 0.086, erring on the safe side in all eight non-normal configurations. Six of fifteen cohorts could detect a medium effect ( d = 0.5) and none a small one ( d = 0.3); the latter would require 416 samples. Of 72 comparisons in the case study, seven (9.7%) had an observed effect at or above their own detection limit and all seven were significant; the remaining 65 fell below it, and all 59 non-significant comparisons lay below their own limit, so none constituted an informative null. A published large-effect association reproduced closely (CDKN2A deletion 41.4% versus 41.5%), whereas a reported medium-effect association could not be addressed: that comparison's detection limit was 0.705 against an observed effect of 0.011. Conclusions A non-significant driver–phenotype result in a small cohort may carry no evidence either way, and this is determinable from cohort size and driver prevalence before any data are collected. Reporting the detection limit alongside negative results would distinguish an underpowered comparison from a genuine absence of association.

Research Square
Yunnan Normal University (CN)
Cancer Genomics and Diagnostics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon