ElectorateIQ doesn't store a pre-built population of twins. It keeps the distribution of voter types for every census tract and assembles a panel for whatever area you select, at run time. Here's exactly what goes into a synthetic voter.
The foundation is public microdata. From the Census Bureau's ACS and PUMS we build a synthetic population for every census tract in the country — realistic distributions of age, sex, race, ethnicity, education, income, and household structure, with the real covariances between them preserved.
A state voter file (owned where available) supplies real party registration and turnout by area. States without an owned file get modeled party from a national model, calibrated to each county's actual 2020 lean.
Grounded in real de-identified records — never a reconstruction of a specific person.
Demographics alone can't simulate political behavior. So each twin inherits the full answer vector of a real survey respondent — drawn from ~61,000 respondents in the Cooperative Election Study, each of whom answered ~150 questions. Those become the twin's seed attitudes, donated intact so they stay internally consistent with one another.
A voter, then, is an observed profile (Census + voter file) plus a donated attitude spine (a real person's coherent answers). Not a label we assigned — a real pattern of belief.
~116 battery items per twin, kept as one coherent vector.
Matching donors on demographics alone would let a ZIP in rural North Dakota and one in Los Angeles draw from the same pool. But place genuinely shifts attitudes within a demographic cell — among White Republicans, "abortion should be legal" runs 24% in rural areas versus 35% in urban ones. So ElectorateIQ matches each twin to a donor who is not just demographically similar but geographically close.
The demographic cell — party, age, education, race, sex — is always the primary match, never thinned.
Among matches, it prefers donors closest on urbanicity, then region, then county lean — picking among the nearest handful for variety.
A rural deep-red-county Republican and an urban blue-county Republican of identical demographics now carry genuinely different real attitudes.
The rule of thumb: if the answer already lives in the twin's data, serve it. If it requires reacting to something new, use the Oracle.
Vote, party, turnout, and standard issue positions come straight from each twin's donated answer, aggregated over the panel and weighted by who actually lives in the area. No model call.
For anything not already a survey answer — a new message, ad, candidate, event, or a values topic the survey never asked — an ensemble of Simsurveys' proprietary Consumer and Social models reads the twin's most relevant seeds plus the stimulus and estimates the likely reaction.
The twins decide who is answering and what they already believe. Two custom-built Simsurveys models decide how they react to something new. A general-purpose chatbot has neither — which is why it returns a confident average instead of a real distribution.
A proprietary population model built and validated on real survey data. It carries the political reactions and is the workhorse of the ensemble.
A second proprietary model trained on decades of social-attitude research (GSS, WVS). It handles the values questions — same-sex marriage, marijuana, the death penalty — that the political record never asked.
Reproducing this would take both halves. The models are years of proprietary training on validated population data; the twins are a custom donor-matching, locality, and persistence system. Standings come straight from real donated answers, reactions from the combined models — and it's the two together, not either alone, that make the read hold up against real results.
ElectorateIQ is designed to help you understand voters and optimize messaging — not to forecast an outcome. We hold ourselves to that line deliberately.
| We do | We avoid |
|---|---|
| Model statistically realistic electorates | Claiming exact voter prediction |
| Maintain coherent, consistent voter identities | Recreating identifiable individuals |
| Use probabilistic behavioral modeling | Deterministic political assignments |
| Emphasize message testing & strategy | Positioning as election-prediction AI |
| Validate at the aggregate level | Overclaiming accuracy |
A donor represents "someone demographically and geographically like this person" — never a specific local individual. That's why every result is validated at the aggregate level, and why reactions to new stimuli should be read relatively: which message moves more, which audience responds best.
Request access and we'll stand up the states and districts you're working.