Beware the Polling Subgroups

I was ironing clothes this morning, half listening to Morning Joe on MSNBC (I jump around to the various cable channels, comparing journalistic malpractice, and usually Fox & Friends wins to a long shot).

They mentioned a Florida Decides poll, noting Donald Trump was ahead in that state -- even with Republican Hispanics.

Me being me I thought, "Wow, did they survey enough of those to really compare?" It's hard to say, but probably no. Here's the crosstabs we'll be looking at. Take your time, peruse them, because it'll help, then go down to the crosstabs of party identification and candidate preference. They're after the story.

Trump leads overall in Florida. OK. Got that. He leads among men and women. OK. And even among those ages 18-34. 

About that age thing. This is mostly a landline robo-poll. Some surveys were sent to smart phones not as calls, but actual surveys you have to fill out. Here's the methodological stuff:

This research was conducted using blended sample, mixed mode: respondents reachable on their home telephone (69% of registered voters) were interviewed on their home telephone in the recorded voice of a professional announcer. Respondents not reachable on a home telephone (31% of registered voters) were shown a questionnaire on their smartphone, tablet or other electronic device.

Finally, Trump leads among Cuban Hispanics by two percentage points over Carson and Rubio, and by a much wider margin among non-Cuban Hispanics (41 percent, to 2nd place Carson's 15 percent). Think about that. Total GOP respondents was 922. What they don't tell you is how many were Hispanic. Or black. Or white. We don't get the raw numbers, but it's very likely the number of Hispanics in this sample was relatively small, with balloons the margin of error to the point where it's hard to say if Trump, despite what MSNBC said, is actually winning among Florida Hispanics.

In other words, beware when pollsters, or journalists, provide differences among subgroups. The margin of error for an entire sample does not apply to subgroups. Ever.