“If it only gets better from here, then what’s there to change about it?“ – Joji

Background

During the course of the West Virginia primary season, State Navigate polled 9 State Senate races in West Virginia’s Republican primaries. Polling these types of campaigns is a difficult task; unlike high-profile races for statewide offices or congressional seats, state legislative elections, especially primary elections, are often sleepy affairs, with relatively few voters paying attention, and turnout wildly unpredictable.

Very few firms publish state legislative primary polling publicly, so we are unaware of any industry-wide analysis of the historical accuracy of state legislative primary polls. In an analysis of their 2020 work, the polling firm Change Research found that their state legislative general election surveys had an average absolute error of around 6 percentage points; we would expect the average absolute error of state legislative primary polls to be much higher; at the presidential level, for example, primary polls are about 5 points less accurate than general election polls. 

Did we pick the winners?

With those benchmarks in mind, let’s dive into State Navigate’s 2026 West Virginia primary polls and see how we did. First, the least scientific question: did we predict the correct winners in each race?

District	Poll Dates	Poll Leader	Actual Winner	Winner Called?
SD-17	5/3-6	Takubo	Takubo	Yes
SD-17 special	5/3-6	Charnock	Charnock	Yes
SD-9	5/3-6	Roberts	Roberts	Yes
SD-2	4/27-29	Heaney	Heaney	Yes
SD-3	4/27-29	Azinger	Azinger	Yes
SD-3 special	4/27-29	Barnhart	Barnhart	Yes
SD-8	4/27-29	Wheeler	Wheeler	Yes
SD-14	3/21-22	Harmon	Taylor	No
SD-1	2/27-28	Chapman	Chapman	Yes

In 8 of the 9 races we surveyed, we did correctly identify the leader. The one miss was in District 14, a 3-way race that we polled back in March. Several factors likely contributed to the error in that case. 

First, having 3 candidates on the ballot likely contributed to indecision among the electorate. Of all the polls we conducted in West Virginia, this one had the highest number of undecided voters (59%), even after we pushed undecideds on if they were leaning toward a candidate in the race. 

Second, this survey was conducted almost 2 months before the election. In a low-profile race like a state legislative primary, many voters have not fully tuned in to the election at that time, and campaigns and outside groups often hold back on spending until the month prior to the election. 

Third, this was one of the closest races that we surveyed; Taylor won the election by 4 percentage points, the second-closest race in our dataset. Our poll showed Harmon ahead by only 3 percentage points, so despite the topline miss, our poll was still within the margin of error.

Fourth, this had the fewest respondents out of any of our surveys. The reason for this was a hiccup in texting voters; we only reached half of our desired voter universe due to being incorrectly marked as spam by AT&T; this is the only time we’ve ever run into this issue, and we made sure that in the later surveys we resolved it with the help of our text vendor, Alliance Forge.

Did we get the margins right?

Next, let’s consider the margins in each race. And here, we’re deviating from typical polling error analysis techniques. 

Normally, when we calculate polling error, it’s not done by considering how closely the margin between the top two candidates matches the actual margin, but by how closely each candidate’s vote share matches the actual vote share. This is done to account for factors such as multiple candidates on the ballot. 

For example, imagine you had a poll with Candidate A at 45%, Candidate B at 40%, and Candidate C at 15%, and the election results showed Candidate A at 52%, Candidate B at 47%, and Candidate C at 1%. If you just analyze the margin between the top two candidates, the poll appears perfect: it predicted a margin of 5%, and that’s what happened in the election. But clearly something is very wrong with the survey, since it estimates the third candidate doing so well, only to see them barely get any votes at all.

A more traditional method would be to look at each candidate’s performance in the poll and the election, and take the average error for each. If you use that approach, the average error for this poll is 9.3 percentage points per candidate, which isn’t a particularly strong showing.

However, in our case, we’ll deviate from this best practice. The reason for this choice is that, due to the nature of the races we’re surveying, we have an unusually high number of undecided voters in our polls. As mentioned above, a majority of voters in the 14th district were undecided; on average, 31% of respondents in our polls said they had not yet decided whom to vote for. As a result, every single candidate in our surveys was underestimated. On average, candidates earned 13% more in their elections than our surveys predicted.

So let’s focus on the margins rather than individual candidate vote predictions, despite the individual-candidate approach being a superior method in a perfect world. We get the following results.

District	Poll Dates	Winner	Runner-Up	Poll Margin	Actual Margin	Error	Error within MOE?
SD-17	5/3-6	Takubo	Pritt	1	12	11	no
SD-17 special	5/3-6	Charnock	Jarrouj	26	13	13	no
SD-9	5/3-6	Roberts	Antolini	6	11	5	yes
SD-2	4/27-29	Heaney	Dobkin	3	9	6	yes
SD-3	4/27-29	Azinger	Fehrenbacher	23	22	1	yes
SD-3 special	4/27-29	Barnhart	Harshbarger	10	21	11	no
SD-8	4/27-29	Wheeler	Bartlett	2	3	1	yes
SD-14	3/21-22	Taylor	Harmon	-3	4	7	yes
SD-1	2/27-28	Chapman	Eddy	25	18	7	yes
					Average	6.89

On average, the absolute error (the difference between the actual margin and the margin in our actual polls) in our 2026 West Virginia primary surveys was less than 7%. Considering the benchmark discussed above from Change Research, of around a 6% error in state legislative general election polling, plus the added difficulty of surveying state legislative primaries, this is a fairly strong performance.

There were three races in which the difference between our poll result and the actual election result fell outside the published margin of error for our survey. The first was the special election in District 3; our poll showed Barnhart ahead of Harshbarger by just 10 points, but Barnhart ultimately won that race by 21 points. Interestingly, we simultaneously polled the regular election in the 3rd District, and the margin in that poll was within 1 point of the actual election results. The difference between the two appears to be undecided voters: in the regular election, 19% of respondents were undecided, but in the special election, 36% were undecided.

The other two races where the error in our poll fell outside the margin of error were the final polls in the two District 17 elections. (Note: we polled this district twice, but only the final survey is included in this analysis. If we included both surveys, our average error would change from 6.89% to 6.91%.) We noted in our analysis when we published this survey that we had concerns about its accuracy; we observed a significant difference in respondents’ demographics between our two District 17 surveys, particularly by region and education. As a result, this poll had a higher design effect than our other surveys; if we recalculated the margin of error to account for this effect the results would not have been outside the margin of error.

What can we learn for the future?

One potential lesson for our future polls, particularly in small-population races like state legislative races, might be to push our undecided voters a bit more. Even the polls conducted the week prior to the election in Districts 17 and 9 had an average of 18% undecided voters. It’s possible that a future survey design that pushed more respondents to choose a candidate might improve our results.

In addition, as we found in District 17, toward the end of the campaign, voters in these small districts may experience polling fatigue. This district was one of the most competitive in the state, and we were not the only pollsters hitting up voters for their opinions; we’re aware of at least 2, and possibly more, surveys conducted around the same time as ours. With so few voters available, it’s quite likely that the same respondents were being tapped again and again to answer polls. This could be a contributing factor to the nonresponse bias we experienced in our second poll in District 17.

There are a few ways to possibly ameliorate this kind of problem, all of which come with trade-offs. First, we could poll the most competitive districts a little earlier (be the first annoying text message, rather than the fifth), to try to catch voters before they tap out of the polling game. The trade-off here is that the further out we are from Election Day, the more voters will be undecided in these lower-profile contests.

We could also extend the field period for these surveys to perhaps give voters we’re trying to capture a little more time to respond to our requests. A longer field period might help us target the voters we see as missing from the sample (in the case of District 17, rural and non-college-educated voters) to reduce the design effect. The trade-off here is that extending the field dates increases costs.

Final Thoughts

All in all, our West Virginia polls appear to have performed at least as well as, if not better than, the industry standard in state legislative primaries. While we cannot expect any polling to be perfect, our work contributed to the public dialogue in these races and, in most cases, added useful information to help the electorate make informed choices. But it may be helpful for us in the future to consider methodological choices that reduce the number of undecided voters, because in several cases, high undecided percentages appear to have contributed to errors.