Background

A growing body of research gives LLMs personas and uses them in place of survey respondents. But when only demographics such as age, sex, and party are provided, responses skew toward stereotypes and lose diversity, because psychological and social context such as values, life satisfaction, and social perceptions is missing. This project made maximal use of already-collected government-approved statistics to create virtual Seoul citizens with context, and used them to forecast the Seoul mayoral race in the 9th nationwide local elections in 2026.

What I did

The work was divided into five stages.

Project workflow Data integration → propensity score summary → generation of unobserved variables → fidelity enhancement → output

Data integration. Because the needed variables were not all in one survey, six government-approved statistics, including the Social Survey, the Economically Active Population Survey, and the Survey on Human Rights Awareness, were combined one after another by statistical matching based on PFI (Parametric Fractional Imputation). Using common variables such as sex and age as links, unique variables from dataset A were transplanted into B, with calibration to preserve each dataset’s marginal distributions. The result was a Seoul citizen dataset of 30 variables × 10,000 people.

PFI-based data integration PFI integration that generates unique variables in another dataset using common variables as links

Propensity score summary. The 30 items were grouped into three axes—socioeconomic status, subjective well-being, and social perceptions—and reduced to 6 latent scores each with IRT (R mirt) and ordinal factor analysis (R lavaan).

Generation of unobserved variables. Demographics and propensity scores were turned into persona text and fed to gpt-4o-mini, gemini-2.5-flash-lite, and gemma-3-27b (temperature 0.7, 10 repetitions), which returned voting participation and candidate preference as JSON.

Fidelity enhancement. Each model showed clear biases: GPT predicted very low turnout, Gemma high support for a particular candidate, and Gemini high support for third parties. We therefore used opinion polls as external information and corrected responses in two ways: reweighting, and regenerating with the information provided (RAG).

Fidelity enhancement methods Model-specific biases and two poll-based correction ideas

Results

Before correction, support for candidate Jeong Won-o varied widely across models: 59.3% for GPT, 16.9% for Gemini, and 95.2% for Gemma. After reweighting, all three models converged to around 51–52% for Jeong Won-o and 45–47% for Oh Se-hoon (51.78% vs. 45.67% for GPT after an additional correction for non-response). The RAG approach barely reduced bias, with GPT at 85.02% and Gemma at 96.32%. The conclusion was that reweighting over-reflects external information, while RAG does not convey information well and is sensitive to how it is provided.

Results by correction method Candidate support by model before and after correction: reweighting vs. RAG

We also built a dashboard for exploring candidate support and voting tendencies by group. It divides the 30 variables into 9 representative persona types so that choices can be viewed by type.

Chuniverse