Polling Averages

Polls & Trackers

Updated December 19th, 2024 at 2:17 PM



What's this?

Welcome to the HenryDRiley.com polling averages and index tool, designed to help you conveniently explore various polls and my polling averages for different races across the United States. This tool draws inspiration from FiveThirtyEight's poll browsing tool. Every poll added to the FiveThirtyEight polling index is included in the averages. Here's a detailed guide on how to use it:

  1. Select the "Poll Type" from the first dropdown menu. You'll find several options such as Governor, President, Senate, and any other offices or types of polls that are currently on the site. The poll type corresponds to the specific data or election you're interested in.

  2. Choose the race location from the second dropdown menu. This list contains all U.S. states/congressional districts that are currently indexed in the polling averages, allowing you to select the geographic area you're interested in viewing polls for.

  3. Once you've made your selections, the election cycles (by year) that correspond to the specific combination of location and poll type you've chosen will be displayed. Click on your desired year to access a page with all relevant polls related to your selected location and poll type. This page will also have my in-house polling average displayed in a chart and summary at the top.

  4. On the page, you'll find a comprehensive list of polls conducted by various polling agencies, along with details like sample size, FiveThirtyEight pollster rating, and dates. You can compare the polls, analyze trends, and gain insights into the political landscape of your chosen state and poll type.

By following these steps, you'll be able to quickly and easily navigate through the polls and our polling averages based on your preferences.


What polls are included, and how are third parties factored in?

Starting in mid-May of 2024, we have made some changes regarding which polls are added to the averages, and how third parties are handled and adjusted for.

  1. Which polls we include: Typically, a pollster will release multiple versions of their survey. For example, let's say the New York Times/Siena College releases four separate toplines, two from registered voters, and two using their likely voter model. Both the registered and likely voter versions have a two-way or head-to-head result and one including third parties. Our averages always prioritize polls in this order: Likely, Registered, and All Adults. Likely voters are those who say they plan on voting or are identified as probable voters who will actually show up to the polls, registered voters are not necessarily likely to vote but are eligible to do so, and all adults represents the general population. We only include the best version of the poll, but if there is only a registered/all adults version, our weighting system has a factor called LRAV that adjusts for this.

  2. How we handle third parties/independents: Most notably in presidential polls, some pollsters only survey voters on a two-way race (the Democrat & Republican), while others include other candidates as a response option. The issue with our averages up until now is that third parties had much lower averages than they should, as two-way polls just had a third party value of 0%, even if they had over 10% support in most polls where they were included. We now factor this in using the following simple step: If a poll does not include the option for third parties, we simply round the current third party average and use that, then take half of the needed support from the major parties. For example, if a poll has Biden at 44% and Trump at 40%, with no third parties, but our average has third parties at 4%, we would take 2% (half of the third party average) from each major candidate, making the poll results Biden at 42%, Trump at 38%, and third parties at 4%.

    3. Once the current polling average is created by using the weighted average of all polls that are recent enough, the model then creates what I call a "Nowcast" for that specific race using the average. This process is described below.


How are polling averages calculated?

Polling averages are calculated by taking into account four main factors when determining the weight of each poll. Each poll is assigned a weight between 0 and 1, which determines how much of the polling average is based on that specific poll. The four factors are described below.

  1. Recency: The model considers the "end date" of the poll, or the last date when responses were collected, and assigns a weight based on recency. Polls surveyed a day ago are weighted the highest, with the weight declining as polls become older. This weighting is automatically updated in real-time every calendar day.

  2. Pollster Rating: This factor estimates the overall quality and credibility of the pollster on a 0-3 scale. I rely on FiveThirtyEight's Pollster Ratings to match this rating to a weighting factor.

  3. LRAV: LRAV stands for Likely, Registered, Adult, or Voter, indicating the type of respondents targeted and interviewed by the pollster. Likely Voters are given the highest weighting, followed by Registered, then Adults, and finally Voters. I am considering swapping the weighting of Adults and Voters in the future for a more logical ranking. The basic principle is that the opinions of those who are likely to vote should hold more weight than just a survey of all adults, who may not actually vote.

  4. Sample Size: This attribute refers to the number of respondents surveyed in each poll. A larger sample size should be more representative of the voting population, so polls with higher sample sizes are given more weight.

After creating the current polling average based on the weighted average of recent polls, the model generates a "Nowcast" for that specific race using the average. This process is described below.


How is the nowcast created?

The nowcast is generated by taking the current polling average, adjusting for undecided voters to estimate the voteshare, and then converting that into win probabilities.

  1. Adjusting for undecided voters: The model allocates the undecided percentage proportionally. For example, if a candidate is leading 70-20%, the remaining 10% won't be split evenly, but more like 9% to the 70% and 1% to the 20%.

  2. Converting to win probabilities: The model uses a Monte Carlo simulation with a standard deviation/possible polling error of 6% (typical polling error is 3-4%) to determine the percentage of outcomes that result in wins for each candidate, thereby establishing their win probabilities.

  3. An important note: Due to calculation limitations, the nowcast may occasionally show a result of 100%-0%. A 100% chance should be considered "virtually" guaranteed, while a 0% chance is "almost" impossible. In essence, it's more accurately represented as 99.9%-0.1%, as nothing is ever certain. Moreover, if the polling average is an exact tie, the code may behave strangely, showing one party as leading but the race rating favoring the other, or producing confusing outputs. In the case of a tie, the odds are simply 50-50.

Once the nowcast probabilities and voteshares are finalized, they are imported into our election models, either nowcasts or forecasts.