Trump or Harris: Who’s ahead in the polls?

Methodology and sources

We use head-to-head presidential election polls in each state to calculate both an average voting intention and a range of likely values for Donald Trump and Kamala Harris. States are considered ‘tossups’ if a candidate’s polling average lead is less than 5 percentage points, ‘leaning’ if it is between 5 and 10 percentage points, and ‘solid’ if it is greater than 10 percentage points.

How are the averages calculated?

To compute the average at a particular moment in time, we take every state-level poll released in the previous 28 days, and assign it a weight based on how long ago it was released, its sample size, how frequently the pollster releases polls, and how far from the previous average it is. The moving average is the weighted mean of these polls.

When a pollster releases the same poll with multiple sample types, we prefer the ‘likely voter’ version of the poll first, then the ‘registered voters’ poll, then finally the ‘all adults’ poll.

We also add a ‘dummy poll’ in each state on each day that uses the change in national polls compared with the 2020 election result, to establish a weak estimate for states with little polling.

Polls are first weighted according to an exponential decay function, so that a poll released today is weighted fully, while a poll released 4 weeks ago is not weighted at all.

Polls with larger sample sizes are also weighted more than those with smaller sample sizes.

If a pollster has released more than one poll within the last 28 days, each poll by that pollster is weighted less to ensure that every pollster affects the average equally.

Finally, to ensure outliers don’t skew the average, a quadratic ‘kernel weighting function’ is applied to give a lower weight to polls far away from the current average.

How are the error ranges calculated?

To obtain a range of likely values that each candidate could win if an election took place tomorrow, we combine two separate sources of error: sampling error and polling industry error.

Sampling error represents the risk that the views of a randomly chosen subset of the population do not match the views of the entire population. For each candidate, we estimate the range of values our moving average could have taken given the sampling error of each poll.

Polling industry error, or non-sampling error, represents the risk that all the polls are systematically biassed in one direction or the other, as they have been in previous elections. Sources of polling industry error include using skewed samples, voters being undecided until election day, or voters not telling pollsters their true intentions.

We estimate non-sampling error by considering how much the results of previous elections differed from their pre-election polling averages. The values we have chosen are based on a combination of academic and original research.

Once we have error bars for each candidate in each state, we estimate the upper confidence bound of Trump’s lead by computing the difference between the upper confidence bound of Trump’s polling and the lower confidence bound of Harris’s polling in each state. Likewise the minimum of Trump’s lead is the lower confidence bound of Trump’s polling minus the upper confidence bound of Harris’s polling.

Sources: FiveThirtyEight, RealClearPolitics, FT