Most political observers assume polling errors are random noise or simple processing glitches, but G. Elliott Morris argues that the catastrophic failure to predict Wisconsin's Democratic primary was a structural inevitability baked into the methodology itself. By accessing raw microdata from State Navigate, Morris reveals that the polls didn't just miss the mark; they were measuring a phantom electorate that looked nothing like the voters who actually showed up. This is not a story about bad luck, but a forensic dissection of why our current tools are blind to the realities of low-turnout primary elections.
The Demographic Mirage
The core of Morris's investigation challenges the fundamental assumption that weighting samples to match registered voter files guarantees accuracy. State Navigate, a non-profit that Morris contracted to build their weighting program, generated benchmarks based on predicted participation that inadvertently skewed the sample toward the young and the liberal. "The July sample was 36% seniors and 17% under 30 — roughly half as old and eight times as young as the electorate it was trying to measure," Morris writes. This discrepancy created a feedback loop where the pollsters were essentially asking the wrong people to predict the behavior of a different group entirely.
Morris demonstrates that the published weights were nearly useless because the sample had already self-selected to match the flawed targets. "The problem is that the targets from State Navigate themselves described a much younger electorate than the one that shows up to a low-turnout August primary," he notes. When he re-weighted the data using actual voter file histories from L2, a commercial voter database, the error margin dropped significantly, yet the fundamental disconnect remained. The pollsters were capturing a universe of voters who identified as "Democratic Socialists" or "Extremely Progressive," while the actual primary electorate was dominated by seniors and moderates.
The poll samples were nowhere near that. The July sample was 36% seniors and 17% under 30 — roughly half as old and eight times as young as the electorate it was trying to measure.
Critics might argue that weighting by ideology is dangerous without a gold-standard benchmark, as Morris himself admits, but the alternative—ignoring the ideological skew—proved far more costly. The failure here wasn't just a lack of data; it was a failure to recognize that the "likely voter" models used in general elections do not translate to the chaotic, low-turnout environment of a primary.
The Clock Was Ticking
Even if the demographic targets had been perfect, the analysis reveals a second, time-sensitive error: the race was moving faster than the polling schedule allowed. Morris leverages a rare feature of the State Navigate study—a recontact panel where the same voters were interviewed in July and again in August. This design allowed him to track individual shifts rather than just aggregate noise. "The recontact design is frankly the coolest thing about State Navigate did in Wisconsin this year," Morris observes, noting that it let him watch voters change their minds in real time.
The data shows that support for David Crowley was surging in the final days, a trend that the final published polls completely missed. "I estimate Crowley was gaining about two-thirds of a point a day in the final weeks of the race," Morris calculates, a momentum that likely tipped the scale after the surveys went into the field. When undecided voters in the August poll were allocated based on how previous undecideds had voted, the race transformed from a landslide into a dead heat. "Apply that same break to the 10% of the weighted August poll that was still undecided, and the race becomes a dead heat," he writes, reproducing the actual 0.5-point margin almost exactly.
This section of the argument is particularly potent because it moves beyond static snapshots to dynamic movement. It suggests that in primary elections, the timing of the poll is just as critical as the methodology. A poll conducted five days too early is effectively a poll of a different election.
The Invisible Non-Response
The final layer of the mystery involves a more subtle form of bias: the tendency for certain types of voters to refuse to answer the phone. Morris investigates whether Hong's supporters were simply more eager to participate in surveys, creating an artificial inflation of her numbers. Using a sophisticated statistical method called the "measure of unadjusted bias for proportions," or MUBP, originally proposed in a 2019 paper by Andridge and colleagues, Morris tests the hypothesis that non-responders were systematically different from responders.
The evidence suggests that even within the same demographic groups, the voters who refused to answer were more likely to support Hong than those who picked up. "The older and more moderate voters who did take the poll were still more Hong-friendly than the older and more moderate voters who actually showed up," Morris concludes. This is a chilling finding for the industry: it implies that even if you get the demographics right, you can still miss the election if the people who are most committed to one candidate are the least likely to talk to pollsters.
The older and more moderate voters who did take the poll were still more Hong-friendly than the older and more moderate voters who actually showed up.
This aligns with broader trends observed in recent years, where polls have struggled to capture the behavior of voters who are disengaged from the traditional political apparatus. While Morris notes that the non-response bias in this specific race was smaller than the demographic errors, it represents a persistent blind spot that no amount of weighting can easily fix.
Bottom Line
G. Elliott Morris's analysis is a masterclass in transparency, proving that the biggest polling errors often stem from flawed assumptions about who shows up to vote, not just how the data is processed. The strongest part of his argument is the use of recontact data to prove that late-breaking movement was the deciding factor, a nuance that standard polling aggregates completely obscure. However, the piece leaves the reader with an unsettling reality: until the industry develops better models for primary electorates and accounts for non-response bias, we must lower our expectations for accuracy in these contests. The next primary season will likely see similar misses unless the methodology evolves to match the volatility of the voters themselves.