Data Coherence
ACSI Processed Food: Within-Respondent Coherence
Goes inside each respondent rather than comparing totals: do a synthetic respondent's twelve ratings of a food brand relate to each other the way real customers' do? Benchmarked against 1,968 buyers in the publicly released American Customer Satisfaction Index 2015 sample. Synthetic answers reproduce the live pattern of question-to-question links (0.89, against 0.97 between two halves of the live sample), 54 of 66 question pairs fall within 0.10 of the live correlation, and six of nine ACSI model links fall within or at the edge of the live range.
The respondent-level live and synthetic data files are available on request for your own analysis.
Trust in Named News Brands: Within-Respondent Coherence
With the brand named, do synthetic respondents tell brands apart and link their answers the way real people do? Benchmarked against 12,043 U.S. adults in Pew Research Center's American Trends Panel (Wave 57, 2019), each saying whether they have heard of, trust or distrust 30 named news outlets. Synthetic answers reproduce the live pattern of links between answers (0.89, against 0.99 between two halves of the live sample), trust across the 30 brands correlates with live at 0.91, and 27 of 30 brands split by party in the same direction as live.
The respondent-level live and synthetic data files are available on request for your own analysis.
Patient Twins vs. Real Respondents: HINTS 7
Do patient twins answer, and stay consistent, the way the real people behind them do? 200 Simsurveys patient twins, each matched to a real respondent in HINTS 7 (the National Cancer Institute's Health Information National Trends Survey), answered 25 HINTS questions in sequence using the production twin setup: each twin draws on its respondent's real answers to other questions, never the one being asked. Answer totals average a KL divergence of 0.049 (24 of 25 questions at 0.15 or less), the pattern of links between answers matches the real respondents at 0.82, and twins chose exactly the same answer as their real respondent 43% of the time, against about 24% by chance.