Mixed1,154 women, 582 held out
Improved A later registered study met the bar: Births, from a voice that knows more than age.
The test at a glance
- What it tests
- Accuracy against reality: who has a baby in the next two years.
- Data
- General Social Survey panels 2006-2010, 2008-2012 and 2010-2014: 1,154 women aged 18-43.
- Held out
- A seeded half, 582 women; every number fitted on the other half.
- Compared against
- Age alone, a regression on the same facts, facts plus voice, and the voice anchored to the population's rate.
- Bar
- Three registered checks: the voice beats age alone, adds to the facts, and anchoring fixes the level.
- Result
- AUC 0.646, level with the regression; best calibrated when anchored; beating age not shown.
Asked the chance that each woman in the public GSS panels (2006-2014) would have a baby in the next two years, a small voiced model ranked them as well as a regression on the same facts (AUC 0.646), and anchored to the population's rate it was the best calibrated of five methods. Scored on whether they had one, 582 of 1,154 women held out. Beating age alone was not shown.
Protocol, result and scorer