If you don't have time or patience to navigate the numbers and findings below, all you need to know is this: two-way vote outcomes obscure a lot of variation in voting behavior — standing up, sitting down, and switching. When we look at all kinds of people across all kinds of behaviors, we find some who have perfect political homes right now, and more who are feeling pushed away or looking for their best fit.
Two groups of loyal voters are gradually losing their political homes: Declining Democrats and Declining Republicans. Two groups of new voters are looking for one: the Newly Motivated and New Republicans.
Winning means making room for four additional tent poles: faith, order, respect, and individual control over outcomes.
It's hard not to feel the chaos of American politics right now; it's front and center in the news headlines and underscored by what appear to be wild swings in the electorate in each election. Every pundit seems to have the answer: Michiganders refused to vote; Latinos swung toward Republicans; Democrats became discouraged. Individually, these and the many other explanations are probably true, but they did little to diminish our sense of chaos. To make sense of it, we needed to understand how all kinds of groups were shifting across all kinds of voting behaviors.
The result: seven clusters of Americans, identified by tracking 2,839 people across three elections in the American National Election Survey — grouped not by demographics but by what they actually did. Who switched parties, who showed up for the first time, who dropped out, who stayed put.
Multinomial mixture models find the groups of demographics which predict patterns of voting behavior. This lets us go further than "swing voter" or "young voter" to understand who is changing, how, and why.
Demographics tell you who's in each group. They don't tell you why someone who voted Democrat in 2016 and stayed home in 2024. Each cluster shares a worldview — what a good person does, who's responsible when things go wrong, what should happen when you play by the rules.
To find the stories underneath the behavior, we used a demographic bridge to map our seven clusters onto the General Social Survey — 16,100 Americans answering questions about God, trust, race, fairness, punishment, and whether life is working. Within the same party, the stories diverge sharply.
The two firm clusters split cleanly on who controls your fate. Firm Republicans put "work hard" first for what children should learn; 38% say people are poor because they lack willpower. You succeed or fail on your own. Firm Democrats say systemic discrimination causes race disparities in economic success (56%) — but 53% also say most people can be trusted, the highest of any cluster. It's not just the system they expect to influence their outcomes but the community.
Firm Republicans: In 2024, 80% voted R, and 7% were new, standing up for the party for the first time.
Firm Democrats: 94% voted D in 2024. 1% went from D to non-voter. 0% defected to R.
These two worldviews — sole architect vs. interdependent community — are the anchors the entire system presently circulates around. Every other cluster is being pulled toward one or the other.
The two declining clusters are the only ones that rank obedience as a top value for children to learn. Both are people of faith — 58% of Declining Democrats and 65% of Declining Republicans say they know God exists, no doubts.
Declining Democrats: 83% voted D in 2024, but 13% went from D to non-voter, and 3% went from D to R.
Declining Republicans: 74% voted R in 2024, but 4% went from R to non-voter, and 3% went from R to D.
Both groups feel increasingly politically homeless. The current has moved — their stories about order, service, and faith haven't.
Three clusters are moving from non-voter to voter. One is flourishing. The other two aren't. The difference explains where they land.
New Democrats: 40% report excellent health — the best of any cluster. They rate their financial position highest of all seven groups.
The Newly Motivated: Just as young as New D, they are matriculating to the electorate and breaking about 3:1 for R. 33% say people treat them like they're not smart — no other group comes close.
New Republicans: Much older, they are not just becoming eligible but coming off the bench. The unhappiest group in the system: 29% say they're not too happy, only 14% report excellent health, 28% say their finances are getting worse.
For some, showing up at the polls is a natural part of their story — it's their turn. They reliably choose Democrats. Other voters coming off the bench are motivated by something different; their story isn't working like they think it should and they're seeking allies to help them achieve the flourishing they want.
Line these seven groups up by where they sit politically and they don't make a left-to-right spectrum. They form two circulation loops anchored by Firm D and Firm R. These two anchor stories are pulling new voters with matching stories toward them.
The two-party system is inherently convective. Shifts in the prevailing culture — beliefs about how the world works, how it should work, and our roles in it — move the anchors of the currents. Instead of rebalancing, the system receives and competes for new voters and pushes voters with nonmatching stories out into open water.
Bottom line: there are voters who will vote if they find an ally that fits their story — someone who can help them get what they want. What would it take to make space for more of those voters in a political party?
The short answer is to create space for people motivated by their faith in God, for people motivated by maintaining social order, for people motivated to seek respect from others, and for people who need to feel agency and control.
Explore the seven stories below.
Tap a mark to navigate. Open the evidence drawer on each card for verbatim GSS question text and the numbers behind each beat.
We used Claude Code to read codebooks, implement the statistical models we designed, and present these results on this website. We didn't use Claude to design research questions, interpret results, or write copy. Because AI is very bad at knowing what's important and is even worse at writing compelling prose.
ANES Panel: American National Election Study 2016–2020–2024 panel. N=4,054 stacked observations from 2,839 unique panelists. Recalled vote transitions across three elections.
General Social Survey: GSS 2016, 2018, 2021, 2022, 2024 waves pooled. N=16,100 (unweighted). Weights: wtssps. Split-ballot design means sample sizes vary by item.
K=7 multinomial mixture model (flexmix) on recalled vote transitions — who switched parties, who showed up for the first time, who dropped out, who stayed put. Concomitant model uses demographics (age, sex, race, education, region, marriage, church attendance) to predict cluster membership. Party ID and ideology were deliberately excluded to avoid circularity.
Multinomial elastic net (glmnet, alpha=0.5, 10-fold CV) trained on the ANES panel to predict cluster from 14 demographic features. Applied to GSS respondents to assign predicted cluster membership — giving us ~80 worldview variables per cluster without ever asking ANES respondents about their beliefs directly.
Standardized Mean Difference — how far each cluster's average is from the population mean on a given variable, in standard deviation units. An SMD of +0.50 means that cluster is half a standard deviation above average. Values above |0.30| are notable; above |0.50| are strong.
Panel attrition: The Newly Motivated cluster's demographic profile (young, non-college, Hispanic, unmarried) makes them the highest-attrition group in the panel — N dropped from 254 to 63 valid observations across waves. The weighted electorate share (15%) is the more reliable population estimate.
Bridge model limits: The demographic bridge can only capture worldview variation that correlates with demographics. Beliefs that cut across demographic lines will be muted.
Cluster 4 (New D): Small sample (N=192 in ANES, N=207 weighted in GSS). Interpret with appropriate caution.
All percentages are weighted. GSS exact question wordings are preserved in the evidence drawers on each story card.
Analysis pipeline: 40_recalled_vote_transitions.R → 43_flexmix_stacked.R → 45_map_to_gss.R → 48_expanded_factor_analysis.R → 50_narrative_profiles.R. All scripts available on request.