Someone once told me, “Numbers will be the death of you.” He may yet be right, but at the risk of a premature demise, I’m going to attempt to enter the mathematical belly of the beast and tackle the argument underway in political circles on the average or mean number of House seats Republicans are likely to lose this fall based on presidential job approval.

Let’s begin with whether using a historical mean to determine projected election losses makes sense. From 1950 on, the average loss of seats for the party holding the presidency in a nonpresidential congressional election is 24.

Republicans currently enjoy a 24-seat majority, and when viewed through a historical lens, their chances of holding that majority seem iffy, statistically speaking. Political pundits and pollsters (myself included) use averages to understand potential outcomes, but when doing so, it’s important to remember that the mean, an average, is a crude model — easy to calculate but not always a good predictor.

In fact, the win/loss record of a president’s party from 1950 onward has ranged from President George W. Bush’s pickup of eight seats in 2002 to President Barack Obama’s loss of 63 seats in 2010. In 1998, Republicans expected to pick up 20 to 30 seats and said as much. We nearly lost the House. But the dubious distinction of the all-time record loss is held by the Democrats — 125 House seats in 1894. So much for averages.

Here’s another example of the risk of relying on averages.

In the last six off-year elections (1994 through 2014), seven, or 58 percent, of the 12 legislative bodies up (House and Senate) flipped control from one party to the other. In the eight off-year elections from 1962 to 1990, there were 16 legislative bodies up and only one flipped, which represented only 6 percent of the time. Looking across all of the last 14 off-year elections gives you an overall average of 29 percent, but I wouldn’t put much stock in that kind of average when trying to predict future flips.

Think of averages this way. What if you used an average to predict student performance? You’ve got a class of fourth graders and the class mean is 80. Individually, actual student scores would cover a range that collectively averages to 80. The class average isn’t a good predictor of each individual student performance.

A little context

When looking at the fall elections, most political commentators, at least for the moment, are falling back on the usual 24-seat-loss statistic as they suggest Republicans are headed for the cliff. That may turn out to be correct, but I’m suggesting some political context might be in order before taking that numerical leap.

Most of the conclusions that predict doom for the GOP are based, not unreasonably, on the impact of presidential job approval on the political environment. But what we should not forget is that President Donald Trump has had strange numbers since Day One.

Watch: Trump Plays to Supporters, Bald Spot in CPAC Address

On Election Day, he had 60 percent unfavorables, according to the exit polls, a number I’ve cited before. But it’s worth repeating because it’s so unusual that we have to consider what it might mean in terms of this year’s elections.

If someone had said in 2015 that the top of the Republican ticket would have this level of unfavorables but that the GOP would win extremely vulnerable Senate races in Wisconsin, Ohio and Pennsylvania to hold the Senate and also win 241 House seats, they probably would have been committed, at least for observation.

And that’s where the numbers get interesting.

The 2016 exit polls showed that Hillary Clinton had a 55 percent unfavorable rating on election day and only 14 percent of those with a negative view of the former secretary of State opted to vote for a Democratic congressional candidate.

On the other hand, Trump had a 60 percent unfavorable rating, but 24 percent of those who didn’t like him much chose to vote for a Republican congressional candidate. This struck me as unusual, the operative word for the last election.

So I took a look at the outcomes of the nonpresidential elections during the past three presidencies, comparing each president’s job approval with how people with a negative view of him voted on the congressional ballot.

Of those who disapproved of the job Bill Clinton, George W. Bush and Barack Obama were doing at the time of the congressional elections, 11 to 17 percent voted for congressional candidates of that president’s party despite their negative views.

Bill Clinton’s 14 percent falls smack in the middle of this range. Trump’s 24-point performance with negative voters, however, falls outside the norm. The question for this November, then, is whether people will associate Trump’s negatives with congressional candidates. The data suggest possibly not at the same scale as with previous presidents.

Factoring in exit poll data from the 2006 Virginia, Montana and Missouri Senate races, a possible 10 percent increase in the vote margins in these states would likely have kept Republicans in control of the Senate. It’s important to point out, however, that favorable/unfavorable and job approval ratings as measurements are similar but not identical. This is still only a theory but a theory based on historical data.

Do we understand completely the idea that 24 percent of those who had an unfavorable view of Trump voted for a Republican congressional candidate? Do we understand completely why 20 percent of those who voted for Trump had an unfavorable view of him — yet still voted for him? The honest answer is not entirely.

The tax cut

Presidential job approval is important, but given Trump’s history, it is less clear how to apply it in his case, particularly when the key issue that got him and Republicans elected (the economy) has been addressed with a significant legislative accomplishment (the tax cut).

Marcus du Sautoy, a British mathematician, has said, “Although the world looks messy and chaotic, if you translate it into the world of numbers and shapes, patterns emerge and you start to understand why things are the way they are.”

When it comes to political data and Donald Trump, I think we’re just getting started.