Adequate sample size for developing prediction models is not simply related to events per variable.
Ogundimu EO, Altman DG, Collins GS. · 2016
Read Adequate sample size for developing prediction models is not simply related to events per variable. by Ogundimu EO, Altman DG, Collins GS. (2016) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5045274.txt.
Description
Emmanuel O. Ogundimu, Douglas G. Altman, and Gary S. Collins test a sample size rule in the Journal of Clinical Epidemiology in 2016. Builders of medical prediction models lean on a rule of thumb: at least ten outcome events for each predictor in the model. Whether that rule holds when many predictors are rare, as real clinical ones often are, was untested. The team drew repeated samples from a general-practice dataset of over two million patient records.
With rare binary predictors mixed in, ten events per variable left bias in the model's estimates. Around twenty events per variable generally removed it, for both hand-picked and computer-selected models. Whether a model even converged depended more on predictor rarity than on the event count. The authors advise making the rule data-driven: aim for twenty events per variable when rare predictors are present, or use a penalized method instead.
Catalog record
| CREATOR | Ogundimu EO, Altman DG, Collins GS. |
|---|---|
| YEAR | 2016 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |