40 How fragile is the evidence base? a meta-epidemiologic study of the fragilityindex derived from 374 randomised trials
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
<h3>Background</h3> Recently, there has been increasing interest in addressing the problem of over-relying on threshold p values. Using p<0.05 represents a blunt arbiter of conclusions that are fraught with false positives and false negatives. Furthermore, questionable research practices are sometimes used to ‘game’ the p-value threshold in order to support the researchers’ preferred conclusions. Tools to highlight p-value shortcomings are required to improve interpretation of p-values. The Fragility Index has been proposed as a tool to highlight the ‘fragility’ of evidence derived from a threshold p-value. <h3>Objectives</h3> The primary objective of this study was to measure the fragility of conclusions from randomised trials (RCTs) published in the New England Journal of Medicine using the Fragility Index. Secondary objectives were to estimate the added impact of losses to follow-up on fragility, and to measure correlation between Fragility Index and standardised effect size, sample size, total number of events, and publication year. <h3>Method</h3> All RCTs of established practices that were published in the <i>New England Journal of Medicine</i> between 2000 to 2016 were included if they met the following criteria: (1) reported a dichotomous primary outcome; (2) had only two comparison groups; and (3) used a 1:1 randomization scheme. Data was extracted from each RCT in duplicate. The Fragility index was calculated by converting one patient in the group (control or experimental group) from a ‘non-event’ to an ‘event’ outcome and recalculating a two-sided Fisher’s exact test until the p-value meets or exceeds 0.05. This Fragility Index was calculated for trials with a significant primary outcome using a Fragility Index calculator, and the reverse Fragility Index for all trials with non-significant (p>0.05) outcomes using an R package. Loss to follow up was measured. Univariable linear regression was performed to assess the association between prespecified trial characteristics and the Fragility Index. <h3>Results</h3> Of 611 RCTs published in the New England Journal of Medicine between 2000 and 2016, a total of 374 met the inclusion criteria. The median Fragility Index was 7.5 (range 0 to 141). One-quarter of the trials had a Fragility Index of 3 or less. The number of patients lost to follow-up exceeded the Fragility Index in 66% (247/375) of the RCTs, indicating that the true Fragility Index would be even lower than reported if corrected for losses to follow-up. The Fragility Index was moderately correlated with the standardised effect size, and weakly correlated with sample size and year of publication. Sensitivity analyses did not reveal material differences when accounting for missing data. <h3>Conclusions</h3> Conclusions from RCTs that are based on p-values are very fragile, with a median of fewer than 8 additional events required to change the conclusion from significant to non-significant (or vice-versa). More than one-quarter of all trials would require only 3 additional events to change the conclusion. Furthermore, the majority of trials had a loss to follow-up that exceeded the Fragility Index, indicating that the results would be even more unstable if the Fragility Index was corrected for losses to follow-up. Efforts to increase awareness of the fragility of conclusions based on p-values is urgently required.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Direct model labels (unvalidated)
Per-model category and study-design labels from the labeling rounds. They are machine output, unvalidated, and the disagreement between models ships as data. No study design here is MEDLINE-validated yet.
| Model arm | Categories | Study design | Confidence |
|---|---|---|---|
| gemma | MetaresearchMeta-epidemiology (broad) Domain: Methods · Genre: Empirical About the Canadian research system: no · About a Canadian topic: no | Meta-analysis | low |
| gpt | MetaresearchMeta-epidemiology (narrow)Meta-epidemiology (broad) Domain: Methods · Genre: Empirical About the Canadian research system: no · About a Canadian topic: no | Meta-analysis | high |
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.124 | 0.104 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.004 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.008 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it