Misleading poll averages

And thoughts about early voting

Two months remain until the big showdown in November. My predictive model, which performed so well from 1952 to 2016, failed utterly in 2020, because I assumed that people were honest. Does polling give us any better insights?

One could argue that the polls did do a better job in forecasting the “official” 2020 election outcome. According to Real Clear Politics, in the average of major national polls, Trump performed 4.3% worse on this day in 2020 compared to 2016, thus foreshadowing his eventual election “loss”.

However, if you accept that argument, then you should also note that Trump is doing 5.0% better today than he did in 2020: and thus 0.7% better than he did in 2016, which he won. Needless to say, he wants to maintain that advantage. Presumably Tuesday night’s debate will reveal Harris to be an empty pantsuit, less knowledgeable than her Democrat predecessors.

But how accurate is the methodology of averaging polls? Look at the results listed on RCP's table of multi-candidate national polls. Notice that some of them contain the abbreviation LV (likely voter) and others the abbreviation RV (registered voter). Broader polls may include the abbreviation A (adult); these are not included in RCP’s average.

However, RCP is misleading readers by including the RV polls. Those who do not vote do not count. (I know; these days that is debatable.) RCP can easily incorporate voter sentiment by using only the LV polls in the average. This is particularly important in the swing states.

There are other ways to improve the average, such as trimming the set of poll numbers to remove outliers, or giving extra weight to observations near the center of the data set. We can accomplish both goals by employing the trimean as a measure of central tendency.

The last thing we could do is to weight the polls by the sample size, but that would be more time-consuming and potentially misleading, for those who remember the debacle of “Dewey defeats Truman”.

Let us compare the RCP average of all LV and RV polls to the Surak trimean of LV polls, computing the margin in favor of each candidate.

National: RCP = Harris 2.1%; Surak = Harris 0.2%

Wisconsin: RCP = Harris 1.5%; Surak = tie

Michigan: RCP = Harris 1.2%; Surak = Harris 0.9%

Nevada: RCP = Harris 0.6%; Surak = Trump 0.5%

Georgia: RCP = Harris 0.1%; Surak = Trump 0.2%

Pennsylvania: RCP = tie; Surak = Trump 0.5%

North Carolina: RCP = Trump 0.7%; Surak = Trump 1.4%

Arizona: RCP = Trump 1.6%; Surak = Trump 2.5%

Nationally, and in each of the seven swing states surveyed, the Surak methodology implies more slightly more support for Trump than the RCP methodology, enough to flip two of the states.

Of course, all of these states are close enough so that turnout will matter a great deal. Contrary to what some pro-life leaders are insinuating, along with alleged Russian influence peddler Lauren Chen, everything is at stake this year.

Some Republican leaders, including Trump himself, have called on conservatives to fight the left by taking part ourselves in early voting. I discourage this, however. Democrats should be given as little information as possible prior to Election Day about the magnitude of cheating that will be required for them to steal another election successfully.

One commenter on X offered a brilliant suggestion one or two months ago, saying that all counties within a state should report their results at the same time. It should never be the case that all the rural conservative counties provide their vote totals, leaving Milwaukee County, Wayne County, Clark County, Fulton County, Philadelphia County, Durham County, and Maricopa County to manufacture votes out of thin air to counteract their neighbors. Only when all counties offer complete vote totals should the results be released for all the counties in the state.