Searcharxiv⌕ Search

arXiv subjects

Howard Kim

Publications and source records attributed to Howard Kim.

2 recordsLinked to original sources

Distributional Validity of a Korean Synthetic Persona Panel: Evidence From the Korea Media Panel Survey

Large language model (LLM) personas are proposed as survey respondents, yet validation outside English-speaking contexts is scarce. We evaluate how well a Korean synthetic persona panel used to condition Gemini 3.5 Flash and EXAONE reproduces digital and artificial intelligence (AI) service-use distributions of the Korea Media Panel Survey. About 8,000 personas per model answered eight service-use items and eight attitudinal constructs; responses were compared with weighted survey estimates. The overall mean absolute error (MAE) was 14-19 percentage points (pp), with binary item-mean correlations of 0.70-0.91 across waves. Segment error across five axes was 14-18 pp, with between-group signed-error ranges of 49.6/34.7 pp (Gemini/EXAONE; 39.5/31.2 without the non-comparable teen cells). Errors were model-specific: an age stereotype (Gemini) versus an acquiescence-consistent level bias (EXAONE). Generative-AI overestimation was consistent with temporal misalignment; short-form underestimation was framing-sensitive and persisted under randomized order (both shown for Gemini). Post-hoc holdout calibration on 30% of the real data, with the correction form selected inside the calibration set, cut cell MAE from 18.3/15.1 to 4.9/4.4 pp, yet direct estimation from that subsample was more accurate than the calibrated panel (3.6 pp), a synthetic-informed shrinkage estimator beat its real-only counterpart by at most 0.7 pp, and the correction did not transfer competitively across waves. The calibrated panel kept an advantage only below roughly 250-860 real responses (at most 2.3 pp over a real-only shrinkage estimator) or, for one model, on unobserved segments. In this setting, synthetic panels are not survey substitutes; their value is diagnostic.

cs.CY↗

Partial pooling beats synthetic respondents as estimators, not as covariates: a preregistered small-sample tournament

Survey estimates for population segments, such as media use by sex and age, rest on few respondents per cell, and a growing literature offers synthetic respondents generated by large language models to fill the gap. That literature has not tested them against the small-area estimators a methodologist would otherwise use. We ran ten estimators, preregistered, on identical household-clustered splits of a Korean media panel across five calibration budgets and three holdout regimes. A multilevel partial-pooling model of the MRP family led every calibration-only estimator at every budget, matching at one hundred respondents what direct estimation reaches with roughly four times as many. On unobserved segments, information fixed as a per-segment prior was inert for synthetic panel and previous wave alike; attached as a bias correction it transferred, at the price of estimation noise. No registered synthetic advantage over MRP survived the paired decision rule. Entered instead as a covariate inside the multilevel model, an exploratory hybrid, the same panel improved on MRP on unobserved segments: the information survives where the registered constructions did not. Interval coverage stayed below nominal throughout and worsened with budget for partial pooling, contrary to prediction.

stat.AP↗