The American National Election Studies is a collaborative research program between the University of Michigan and Stanford University. Since 1948, it has surveyed American voters before and after every presidential election — and most midterms — tracking political attitudes, behaviors, and social identities over time.

The dataset is the gold standard for political science research. It covers party identification, voting behavior, candidate evaluations, policy attitudes, racial identity, authoritarianism, social trust, and hundreds of other constructs — across more than 70 years and tens of thousands of respondents.

Variable naming (VCF codes): Every ANES variable has a code like VCF0301 (party identification). Lower-numbered VCFs (0001–0200) are administrative and demographic — they cannot serve as research outcomes. Variables above VCF0200 are substantive attitudes and behaviors, suitable for both predictors and outcomes.

Survey windows: ANES conducts a Pre-election wave (attitudes before voting) and Post-election wave (reported behavior, retrospective evaluations). Pulse models are trained on the Pre-election wave for attitude outcomes and Post-election for behavioral outcomes like vote choice.

Political scientists organize explanations of voter behavior around a funnel of causality — a principle introduced in The American Voter (Campbell et al., 1960) and still the backbone of ANES research design. The idea is simple: causes closer to the individual's biography are more distal (acting early and indirectly), while causes closer to the election are more proximal (acting late and directly). The funnel narrows as we approach the vote.

The funnel shapes how Pulse constructs its models. A distal variable like education is a legitimate predictor of party ID, issue positions, and vote choice alike. A proximal variable like a candidate thermometer is a powerful predictor of vote choice — but using it to "explain" vote choice is nearly circular: a voter who rates the Democratic candidate highly almost certainly votes Democratic. That is not a causal discovery; it is a measurement redundancy.

In practice this means Pulse's LASSO predictors reflect genuinely independent contributions — not artifacts of asking the same question twice in different forms. Users who build custom hypothesis tests should keep the funnel in mind: placing a proximal predictor alongside a distal one is valid, but placing two variables from the same funnel layer that share conceptual content risks multicollinearity and tautological inference.

Work through these steps at your own pace.

Pulse 2.0 trains LASSO regularized regression models on ANES data. Each model predicts one outcome variable (DV) using all valid predictor variables (IVs) simultaneously. Models are evaluated out-of-sample to ensure estimates are not overfit to the training year.