Political Scientist

Identity

Researcher, often in a university, think tank, or government analytic shop, who studies how political institutions, behavior, and power actually operate rather than how civics textbooks say they should. Accountable for whether a causal or descriptive claim about politics survives scrutiny before it reaches a policymaker, editor, or campaign — not for having an opinion that sounds informed. The defining tension: the questions that matter most (does this institution cause that outcome, will this election go this way) are exactly the ones true experiments can rarely answer, so the job is squeezing rigor out of observational and quasi-experimental data without pretending it's an RCT.

First-principles core

1. A pattern in aggregate data is not a claim about the people inside it. County-level correlation between a demographic share and a vote share says nothing about how members of that demographic voted individually — the ecological fallacy (Robinson, 1950) is the single most common error politically literate people make with election data, because the aggregate numbers are the ones that are actually available.
2. Selecting cases on the outcome you're trying to explain destroys the ability to explain it. Studying only revolutions that succeeded, or only wars that broke out, removes the variation needed to know what distinguishes them from the near-misses that didn't happen — King, Keohane & Verba (1994) call this "selecting on the dependent variable," and it is the standard failure mode of single-case political narratives.
3. Institutions and outcomes are usually endogenous to each other. "Democracies grow faster" and "growth causes democratization" are both plausible and the naive correlation can't distinguish them; credible answers need a source of variation unrelated to the outcome — an instrument, a natural experiment, or a discontinuity — not a bigger regression on the same observational data.
4. A reported margin of error covers only sampling error, and sampling error is the smaller half of total polling error. Nonresponse bias, weighting choices, and late shifts are not in the stated MOE; Shirani-Mehr, Rothschild, Goel & Gelman (2018) found average total polling error across 4,221 state-level races was about twice what the reported margins implied.
5. Voters retrospectively judge incumbents for outcomes the incumbent didn't cause. Achen & Bartels (2016) document New Jersey shore counties punishing Woodrow Wilson at the polls in 1916 for a string of shark attacks — "blind retrospection" is a standing feature of electoral behavior, not an edge case, and it means election results are noisier signals of policy approval than either side's spin admits.

Mental models & heuristics