In Silico Development of Psychometric Scales: Feasibility of Representative Population Data Simulation with LLMs

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Enrico Cipriani, Pavel Okopnyi, Danilo Menicucci, Simone Grassini

Keywords: synthetic respondents, psychometric scale development, factor analysis, measurement invariance, individual-level validity, preregistered studies

Source: Open primary source (opens in a new tab)

4
Authors
4
Findings
4
Limitations
3
Evidence

Editorial summary

English

Four preregistered studies test whether GPT-4o-mini-2024-07-18 can generate synthetic participant responses useful for scale development before human data collection. The model receives demographic profiles based on representative UK quotas and answers three reworded prompts per item at temperature 1; item responses are averaged. Each study compares about 300 real and 300 simulated participants: 316/322 for a climate scale, 331/331 for ICT-SC25, 301/300 for SAGAT, and 301/300 for AI Anxiety. Study 1 fails to reproduce the factor structure or invariance. Study 2 reproduces the structure with configural and partial metric invariance, but matching-dimension correlations are only 0.21–0.35 and individual-level ICC is 0.19. In Studies 3 and 4, EFA on synthetic data proposes structures that obtain good CFA fit in newly collected human samples and reach approximate configural, metric, and scalar invariance, but not residual invariance. Across all studies, distributional and variance discrepancies remain, while human–synthetic correlations are weak or near zero. The evidence therefore supports, at most, exploratory group-level early prototyping of factor structures; it does not support replacing human validation or simulating individual responses. Internal gender invariance within synthetic data does not establish human fairness and may simply reflect generator regularity. The work uses one model, one country, English, and Likert scales; invalid outputs are imputed by averaging other prompts, and Study 1 includes an acknowledged poorly specified hypothesis. The OSF links were checked, but their public API did not expose an immutable, auditable code-and-data package.

Español

Cuatro estudios prerregistrados examinan si GPT-4o-mini-2024-07-18 puede generar respuestas de participantes sintéticos útiles para desarrollar escalas antes de recoger datos humanos. El modelo recibe perfiles demográficos de cuotas representativas del Reino Unido y responde tres reformulaciones de cada prompt a temperatura 1; las respuestas se promedian por ítem. Los estudios comparan aproximadamente 300 personas reales y 300 simuladas cada uno: 316/322 en clima, 331/331 en ICT-SC25, 301/300 en SAGAT y 301/300 en AI Anxiety. En el primer estudio no se replica la estructura factorial ni la invariancia. En el segundo se replica la estructura, con invariancia configural y métrica parcial, pero las correlaciones por dimensión son solo 0,21–0,35 y el ICC individual 0,19. En los estudios tercero y cuarto, una EFA sintética propone estructuras que obtienen buen ajuste CFA en muestras humanas nuevas y alcanzan aproximadamente invariancia configural, métrica y escalar, pero no residual. En los cuatro estudios persisten diferencias de distribución y varianza, y las correlaciones entre datos humanos y sintéticos son débiles o cercanas a cero. Por ello, la evidencia respalda como máximo un uso exploratorio, grupal y temprano para proponer estructuras factoriales; no permite sustituir validación humana ni simular respuestas individuales. La invariancia de género dentro de los datos sintéticos tampoco prueba equidad humana: puede reflejar regularidad del generador. El estudio usa un solo modelo, un país, inglés y escalas Likert; además, imputa respuestas inválidas promediando otros prompts y reconoce una hipótesis mal especificada en el estudio 1. Los enlaces OSF se comprobaron, pero no ofrecieron mediante su API pública un paquete inmutable y auditable de código y datos.

Research question

Can a synthetic population generated by an LLM reproduce factorial structures and human measurement properties with utility for the initial development of scales?

Method

Four preregistered studies compare real British samples and synthetic profiles. Two replicate existing scales and two develop scales through synthetic EFA followed by human CFA; fit, invariance, correlations, ICC, distributions, and variances are evaluated.

Sample: Study 1: 316 humans/322 synthetic; study 2: 331/331; study 3: 301/300; study 4: 301/300.

Findings

  • Three of four factorial structures are replicated or generalized.
  • Studies 3 and 4 achieve approximate invariance up to the scalar level, but none demonstrate residual invariance.
  • Individual properties, distributions, and variances are not reliably reproduced.
  • The observed utility is limited to early group factorial prototyping.

Limitations

  • A single model, English, United Kingdom, and Likert responses.
  • Imputation of invalid responses through other prompts.
  • Hypothesis H3 of study 1 poorly specified and without valid individual correspondence.
  • The OSF links did not allow auditing an immutable public package through the consulted API.

What the study does not establish

  • Does not validate substitution of human participants.
  • Does not reproduce individual persons or their real covariation.
  • Does not demonstrate gender equity in human populations.
  • Does not guarantee external, predictive, clinical, or cross-cultural validity.

Traceability

Scope: Full text

Version: arXiv:2512.02910v2; complete 56-page PDF; four preregistered studies; OSF share links checked but no immutable public artifact manifest retrieved

Consulted source: https://arxiv.org/abs/2512.02910v2

Review: Codex full-text, 56-page visual, psychometric and artifact audit, 2026-07-18

Approval: Codex fidelity pass, 2026-07-18

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o-mini-2024-07-18

Instruments and metrics

  • Climate-related scale
  • ICT-SC25
  • SAGAT
  • AI Anxiety Scale

Data used

  • Four UK Prolific quota samples
  • Four LLM-simulated demographic quota samples

Evidence and location

  • Samples and protocol of four studies: arXiv v2, methods for Studies 1–4
  • Invariance and validity by study: arXiv v2, results and Tables 2–7
  • Synthesis of hypotheses: arXiv v2, Table 7 and general discussion