ANALYSIS
Public opinion polling is useful for gauging the attitudes of people living in a democracy. But a poll can only accurately describe an electorate if the people included in the survey reasonably approximate the population understudy. In election years with polls of “voters”, this poses an important methodological challenge: researchers must decide which voter characteristics to reproduce in their survey samples and how much “weight” to assign to them.
While polling cannot control voters' decisions about who they ultimately vote for, pollsters can take steps to make their sample as close as possible to the electorate they’re trying to approximate. Polling can do this by setting “quotas” for key demographics during data collection and by weighting the data to reflect those quotas when they are not met. Weighting adjusts the influence individual respondents have on survey estimates so that the weighted sample matches selected characteristics of the population being studied. Importantly, weighing does not have a single "correct" setting.
Researchers must decide which characteristics to weigh on and what benchmarks to use for those characteristics when doing data collection and analyzing the results. Both quotas during data collection and weighting are done with an eye toward benchmarks from the census, voter registration data (if states collect and provide such information), exit polling from prior elections, and voter turnout data. Also, part of this is whether we are talking about people who are registered to vote and those who are likely to vote, with each group being appropriate depending on the question being asked or when the survey is being fielded. For example, in Illinois, only about 50% of registered voters cast their ballot in midterm elections. It’s closer to 70% for presidential elections. Which benchmarks are included and what they’re set to can have a big impact on results.
Thinking about this leads us to ask: how much can election estimates change depending on the assumptions researchers make about who will vote and what the electorate will look like? And does the impact of these choices vary on how close to the election we are?
To help people understand the impact of these types of decisions, the following is an analysis of two recent surveys on public opinion among registered voters in Illinois regarding the 2026 Illinois Governor’s race that we collected. Fieldings occurred using an online panel of registered voters from 8/12/16 to 8/26/26 and 10/1/26 to 10/22/26. These surveys were not for any political campaign and was funded by the non-partisan Center for State Policy and Leadership at the University of Illinois Springfield and the University of Illinois System’s Institute of Government and Public Affairs.
Full transparency: we're not big fans of asking people the standard "who would you vote for?" question to assess the state of a race. We prefer approaches that ask voters to make predictions based on their collective knowledge, which we have found to be more accurate for predicting election outcomes and vote totals, though less accurate for competitive primaries and Electoral College forecasting. But this survey provides a useful opportunity to demonstrate how demographic decisions about sample composition can affect the results a poll produces.
To do this, we explored a couple of ways of looking at our survey data. The first uses pretty standard demographic benchmarks including a voter’s partisanship, age, race, sex, and education. Second, we further refine our weighting from the “standard” above to include the state region (City of Chicago, Chicago Suburbs, North/Central Illinois, and Southern Illinois), self-reported political ideology, and income. These benchmarks come from prior exit polling in Illinois and our own previous polling work. Unlike characteristics such as age or race, there is no definitive administrative source telling us the ideological composition of Illinois voters. Ideological self-identification can also change over time and can depend on the wording and response options used in a survey. We therefore chose a relatively cautious benchmark of 34% liberal, 34% moderate, and 32% conservative, based on prior Illinois polling and exit poll data. This benchmark may understate the liberal share of the 2026 electorate. Still, we believe it is preferable to make a stronger assumption about the electorate's composition for now, in the absence of clear empirical evidence.
The third adds in respondent recalled 2024 vote as. Third, we incorporate prior presidential vote recall for those who voted in the 2024 U.S. presidential election. Vote recall is not a perfect measure because some respondents may misremember or misreport their previous vote, and since we have private ballots there is no way for us to confirm who folks voted for. Nevertheless, prior presidential vote is increasingly used in political polling because it can provide information about a sample's composition that is difficult to capture through demographics alone. In 2024, 54% of Illinois voters cast their ballots for Kamala Harris, 43% voted for Donald Trump, and 2% voted for other candidates.
However, the above raises an important point. Not everyone who is registered to vote actually does so. So, if we want to estimate what the election might look like among people who actually vote, we need to build a likely-voter model. There are several ways to do this, and people have their preferred methods. We base ours on three things: whether the respondent plans to vote in the 2026 election and whether they are highly motivated to do so.
Looking at the table above, you can see how the sample's demographics, our choices, and time shape the results. In particular, August while we didn’t see huge swings for Pritzker and Bailey, Corbett did “best” in August when we incorporate 2024 vote choice and look at all registered voters but drops once we start building out a likely voter model. We do see a big jump for Governor Pritzker in the original likely voter October model while we see a big drop for Corbett. However, respondents often overestimate the likelihood they will vote. The further out from the election you are, the more voters say they will vote. In August, 68% of the respondents were likely voters. Come October, it’s down to 61%. But normally governor elections usually have around a 50% to 55% voter turnout. So our last model uses the October data but sets the weight for likely voters down to the high end of that range. When we do that, we see Governor Pritzker’s vote share drop down from the original voter model down to 56%.
Importantly, each estimate is based on the same underlying survey respondents for the two data collections. What changes from model to model are the assumptions we make about how those respondents should be weighted and which respondents should be considered likely voters. Keep in mind also that our models have an estimated uncertainty of ±3.50, while our likely voter models are about ±4.50 for both data collections.
Are we saying this will be the outcome of the election? No. The point we’re hoping to leave you with is this: When evaluating a poll, it is tempting to focus primarily on the reported horse-race numbers. But those numbers are the product of a series of decisions about who is included in the sample, how respondents are weighted, and who is assumed to participate in the election. Reasonable researchers can make different decisions about those questions, which are defensible and valid, and consequently produce meaningfully different estimates from the same underlying data. Understanding those choices is therefore essential to understanding what a poll actually tells us. Something to keep in mind as we all try to read the tea leaves for what the November 3rd results will be.
*The October 2026 data will be available in the future on the IGPA datahub, which can be accessed here.
