LLMRpanel: calibrated silicon samples for survey and experiment design

Quantitative silicon sampling with the methodological stance baked into the objects:

Details

• panel_from_margins(): persona panels drawn from population margins you supply (ACS/ANES-style), with the persona text rendered from a template.

• panel_instrument() with item_likert(), item_choice(), item_open(); vignette_design() and conjoint_design() for factorial stimuli.

• panel_from_data(): the joint-distribution counterpart, drawing personas from microdata rows.

• panel_administer(): every persona answers every item, with item- and option-order randomization recorded per response.

• panel_calibrate(): compare silicon marginals to human benchmarks; until it runs, every print method shows an UNCALIBRATED banner.

• panel_bias_audit(): option-order effects and refusal/parse rates – the response-style artifacts silicon respondents are known for.

• conjoint_instrument() and amce(): forced-choice conjoint items and their average marginal component effects, with respondent-clustered standard errors.

• panel_power(): analytic two-arm power for the planned human study, priced from the silicon pilot.

• panel_report(): the design-stage report, banner included.

The stance, printed rather than preached: silicon panels are instruments for the design stage – pretesting questionnaires, piloting vignette and conjoint designs, stress-testing instruments – and for measuring model behavior. They estimate human population quantities only to the extent that calibration against human data earns that reading, case by case.

See also

Useful links:

• https://github.com/asanaei/LLMRpanel

• https://asanaei.github.io/LLMRpanel/

• Report bugs at https://github.com/asanaei/LLMRpanel/issues

Author

Maintainer: Ali Sanaei sanaei@uchicago.edu