We test twin-panel predictions against real, independent government data — not internal consistency checks. Below is the first published result: a 9-state calibration study on a genuinely held-out behaviour, checked against the National Family Health Survey.
0.700
Spearman correlation vs. NFHS-5, 9 states
80.6%
Directional accuracy — 29 of 36 state-pairs ranked correctly
4.84pp
Mean absolute error against the real state-wise rate
Study 1
State Calibration: Alcohol Consumption
Alcohol consumption varies enormously by state in India — from single digits in some states to over a third of men in others. That real spread makes it a clean test: does the twin panel’s state-level prediction track the state-level reality?
Panel
225 male twins, 25 per state, 9 states
Ground truth
NFHS-5 (2019–21), % of men who drink, by state
Design
Held-out — no alcohol-related field exists anywhere in the twin’s data or prompt, verified before running
Question
Each twin estimates, unprompted, what share of people like them would answer “yes” to a routine health-survey question
Held out means held out. Before this study ran, we audited the entire twin schema, every free-text field, and the prompt-construction code for any trace of alcohol data. None existed — the panel had to infer the answer from real regional, religious, and community context, the same way an actual survey respondent’s answer is shaped by their own life, not a lookup.
Result
Twin prediction vs. real NFHS-5 rate, by state
States are ordered by their real rate, low to high. The panel’s predictions follow the same climb.
% who drink alcohol
Twin-predicted (panel average) vs. real NFHS-5 rate