Documentation

Task battery

Each experiment presents a fixed battery of calibrated either/or choice tasks. For Prospect Theory this is 13 tasks: five mixed gambles against a certain £0 (identifying loss aversion λ), four gain gambles at small and large probabilities (identifying probability weighting γ), and four 50/50 gain gambles against certain amounts (identifying value curvature α). Each task is constructed so its indifference point sits at a known parameter crossover, meaning every answer rules parameter regions in or out.

Estimation

Estimation is two-stage. A coarse grid search (512 candidate (α, λ, γ) triples, ties resolving towards the Kahneman–Tversky empirical medians) finds a robust starting point, then a maximum-likelihood refinement under a logistic choice rule produces continuous estimates. A weak prior centred on the grid winner regularises the fit when a respondent's choices are perfectly consistent, and standard errors are derived from the curvature of the penalised likelihood where it is well conditioned.

Data quality

One response is accepted per device via an idempotent respondent token, and responses sharing a network are flagged (not excluded) in the researcher's results, since shared networks are common in classroom settings. Raw choice data and fitted parameters are exportable as CSV for external analysis.

Limitations

With a short series of binary choices, the data can be consistent with several parameter combinations; standard errors are reported where estimable, and results should be read as behavioural profiles rather than precise psychometric measurements. Estimates are regularised towards population-typical values when responses carry little information. Experiments are currently available for Prospect Theory and Hyperbolic Discounting, with further models planned.

Method

This is the silicon sampling method of Argyle, Busby, Fulda, Gubler, Rytting & Wingate (2023), “Out of One, Many: Using Language Models to Simulate Human Samples”, Political Analysis 31(3). Conditioning a language model on socio-demographic backstories makes its answers statistically resemble those of the matching human subgroup. This platform applies the method to behavioural-economics parameters rather than attitudes.

Personas

UK personas are sampled from the ONS 2011 Census Microdata Teaching File (a public 1% sample of England and Wales); US personas from ACS 2023 PUMS microdata covering all states. Because whole records are sampled, age, sex, income, education, occupation, and region keep their real joint distribution. Sampling is seeded, so any audience is exactly reproducible. Each generated audience records the model, prompt version, and seed used.

Constraints

The model's only permitted output is a choice of A or B for each task; anything else is rejected and retried once, then discarded. Option order is randomised per respondent to cancel position bias. Raw model text is never stored or displayed; every number and sentence a researcher sees is computed by the platform from validated choices.

Validation

Synthetic data approximates human responses and should be validated against them. When an experiment holds both cohorts, the results page reports each parameter's human and synthetic means, the variance ratio, and a two-sample Kolmogorov–Smirnov test, alongside the published literature benchmarks. Synthetic respondents are always flagged, in the interface and in every export. Validation audits pin every run to a single model, so measured noise is response noise rather than disagreement between providers.

Monthly allowance

Every paid plan includes 500 synthetic respondents and 10 audience generations per calendar month; a single generation can hold up to 100 respondents. Validation audits draw from the respondent allowance but do not use a generation. Pausing a generation returns its unused respondents to your allowance until you resume. Current usage is shown on your account page and in the generate dialog.