President Trump loves to brag about his polls. But in recent days, he’s seized on a number that might make even his most jaded observers raise their eyebrows — claiming that his support is up 19 percent with Latinos despite shutting down the government over his desire for a wall along the U.S.-Mexico border.

Trump doesn’t include a citation in that tweet, but he appears to be referencing an NPR/PBS/NewsHour/Marist poll that shows Latino support for him at 50 percent. If that’s true, that’d be a big deal. Trump won less than 30 percent of the Hispanic/Latino vote in the 2016 presidential election, and a 50-50 split among Latinos would have a big impact on politics. Even Trump himself sounded a note of incredulity, finishing his tweet with, “Don’t know how my poll numbers are so good.”

That’s the question both he, and we, should be asking.

One of the first rules of polling analysis is that unexpected results need corroboration. If you see a poll number that looks weird, the best response isn’t to accept it on face value or dismiss it out of hand — you should see what the rest of the data says.

And other polls disagree with Marist. The Post/ABC News poll shows Trump’s approval among Hispanics at 18 percent. Ipsos has never had Trump above 40 percent with Latinos, and the Hill also shows low numbers for the president. The Economist/YouGov poll puts Trump at 30 percent with Latinos, and they haven’t seen a recent surge in his numbers.

If Trump’s approval rating among Hispanics/Latinos was suddenly and radically shifting, I’d expect to see evidence in some of these other numbers.

Another helpful rule in polling analysis is not to get lost in the details.

Trump has a high level of Latino support in the Marist poll, but his overall job approval rating is still a not-so-great 40 percent among registered voters in that survey. Moreover, Trump’s standing in poll averages has been falling. Americans blame him more than anyone else for the government shutdown, and that’s taken a toll on his popularity. Trump might have gotten a good Hispanic/Latino number in one poll, but his overall numbers are still bad.

I don’t want to seem like I’m picking on Marist here; I’m not. Marist is a solid pollster. I’ve used them in predictive models in the past, and I’ll use them in the future. Polling inherently involves randomness and error, and even the best pollsters often have really large margins of error around their demographic numbers. Barbara Carvalho, the poll’s director, has said that Trump didn’t place the numbers in appropriate context. Marist did a good job contextualizing the poll here.

The president (or whoever was running his Twitter account at that time), on the other hand, is doing bad polling analysis. The polls show that the president is unpopular, and the results of the 2018 House elections confirmed that. And the polls show, contrary to Trump’s claims, that he slipped rather than surged during the shutdown.

Trump can demand that we “Look to [the] final results!” when judging him. He just might not be happy with the results.

Read more: