Do Culture Wars Drive Our Elections?

Forecasting the 2022 US Senate Races

Thirty-five US Senators will be elected in November 2022. Typically, the incumbent party loses seats in the midterms. If history repeats, Republicans will regain control of the Senate. And, with President Biden’s low approval ratings, this outcome can seem even more likely.

But conventional wisdom and polling are so unreliable! Pollsters forecasted that Hillary Clinton would win by a landslide in 2016 as would Joe Biden in 2020. Clearly, polls are a blunt tool if not dangerously misleading.

Better Election Forecasts

I published four forecasts for the 2020 presidential election using Google Trends rather than election polling data. Three out of four of my forecasts were more accurate than Nate Silver’s FiveThirtyEight.

My last forecast was just before the November 3 election. It was especially accurate and beat FiveThirtyEight by a wide margin.

The key to success was finding loaded, “culture war” variables that predicted the 2016 presidential election. Then I used these variables to predict 2020.

With this article, I launch a new forecasting series. This time I use culture war variables to forecast the 2022 Senate elections.

Culture War Variables

The underlying premise behind this and my prior election forecasts is that we are in the midst of a culture war. It has simmered throughout US history, peaked during the Civil War, and now threatens our democracy once again.

But how can roiling tides of sentiment be measured and used to predict election outcomes? My method is summarized in two steps:

1. Find search terms or keywords people type into their browsers that are loaded with cultural values, judgments, and underpinnings.
2. Select loaded terms that mirror previous election patterns or are emerging manifestations of the culture war.

Loaded Terms that Mirror

Loaded cultural terms are often subtle. For example, in April 2020 the search term “tanning” mirrored Biden’s percent of the popular vote per state in November 2020.

Some state populations are more interested in having a deep tan and are inclined to dismiss the associated risk of skin cancer. As tanning searches become more popular in a state, the likelihood of that state voting for a Republican versus a Democrat increases.

More broadly, the tanning variable gauges the state’s respect for science and medicine, the willingness to sacrifice a short-term superficial gain for long-term health, the importance of physical appearance over inner beauty and intelligence, and so on.

Some culturally loaded terms have been around for decades but their predictive power ebbs and flows. Take the term guns. When the term is not highly politicized, its use in search can distinguish gun-favoring states from the others. But with the recent Uvalde tragedy, every state is interested in guns, so its predictive power disappears. It is still useful to track but may not regain its predictive power before the 2022 midterms.

Covid was an issue country-wide in 2020, so it was not a strong predictive variable then. Now the term is a culture war emblem. Republican-leaning states tend to dismiss covid, often with disdain, which is reflected in their low search activity for the term. On the other hand, Democratic-leaning states are still concerned and search for covid-related information more often. These dynamics make Covid a powerful prediction variable at this time.

Top 10 Predictor Variables

The base set of predictor variables includes old and new variables, for a total of 30. They were found by looking for search terms that mirrored 2016 and 2020 presidential election results. Then Google search data was collected for April and May 2022 for all 30 variables. The top 10 variables for each month are listed in the table below. They appear in order of monthly search volume.

Five variables were strongly related in both April and May to the 2020 presidential election: Walmart, Whole Foods, Covid, Wikipedia, and vaccine.

The Preliminary Forecast

When using all 30 variables for April and May, Democrats are forecasted to have a net gain of one seat in the Senate.

Note: We are five months away from the election. Much can happen between now and then. Consequently, this forecast is preliminary.

Q&A

What forecasting method did you use?

My method follows three steps.

1. Start with a known event in the past. For example, Joe Biden won X% of the popular vote in State Y in 2020. This data is gathered for all 50 states.
2. Select variables that accurately predict #1 using data that precedes the event. For example, state-by-state search patterns for “Wikipedia” in October 2020 mirror the actual November 2020 presidential election outcome, so it is selected.
3. Using current data, apply a broad set of selected variables to the 2022 Senate race. For example, state-by-state search patterns for Wikipedia in April and May 2022 are used to forecast the 2022 Senate races in November for each state.

Why tie variable selection to the 2020 presidential outcome?

This approach avoids bias. Only variables that are proven to be effective in predicting the past are used to forecast the future. This forecasting method is widely accepted and is especially popular among technical financial traders.

Sure, things are different now. Biden is less popular, inflation is high, and so on. But my method accurately predicted Biden’s victory by selecting variables that accurately predicted Donald Trump’s victory over Hillary Clinton. So, I am using the method again.

Pundits make forecasts based on theory and opinion. My forecasts are rooted in math and science, and my accuracy is measured and reported.

The risk of my approach is that some variables that capture important 2022 election dynamics were irrelevant during the 2020 election. The full set of 30 variables addresses this risk and is used to create a secondary forecast.

For example, new variables such as “Truth Social” and “NFT” are among the 30 variables. Some, like NFT, may show a strong connection to the 2020 election and will be used to create the top-10 variable forecast. Otherwise, they will remain in the set of 30 variables, and the 30-variable forecast will be compared and assessed relative to the top variable forecast.

Do you weight variables by search volume? Some keywords are much more popular than others.

I tested variable weighting during my 2020 election forecasting effort. The results indicated that weighting reduced forecast accuracy, so I do not use it.

How do you account for gerrymandering?

I don’t. Gerrymandering will impact 2022 election results for state House and Senate seats, but not for US Senator elections.

How do you account for recent voter restriction laws?

State-level efforts to limit Democratic voting and discount Democratic votes are very real and worrisome but are not directly addressed in this forecasting method. Such attacks on fair election procedures will distort election outcomes, especially in states like Georgia, Michigan, and Wisconsin. The model may reflect a state’s inclination to support voter restriction, but I currently have no way of testing or proving this.

Why are there no economic variables in your model?

Economic dynamics are nested within many of the 30 variables, but their presence is subtle. Walmart and Whole Foods reflect economic and other dynamics. The same is true for opera and organic.

Terms such as inflation, income, and wages are not effective predictors. If they were, they would be included in the model.

Conventional wisdom says our current high inflation spells serious trouble for Democrats in November. My analysis says inflation as a stand-alone issue is not a deciding factor. Other dynamics that are more closely tied to our culture wars will drive election outcomes in the US.