New subscriber. I came as much/more for the adult statistical analysis and discussion as the juicy topics covered. I figure if I keep reading on a regular basis, (going full Bayesian here) that I might actually move my mean understanding of statistics ever-higher.
About ten years ago I answered my phone to an unknown number from another country.
It was a teenager in a call centre who gave me a barrage of questions about politics with a few market research questions thrown in at the end. She absolutely raced through the questions, couldn’t pronounce the names of political parties, and clearly had no knowledge of our interest in the subject matter she was surveying. I doubt that her data entry was very careful.
In other words, huge potential for the introduction of non-sampling error
Anyway the political opinion poll appeared in my national media the following weekend with n≈1,000. Ever since I’ve been very sceptical about political opinion polls. I think they are exclusively done as marketing by polling companies.
Interestingly when I was surveyed enough there was a topical question clearly paid for by one of the political parties that was never destined for public release
But wouldn't this snowball? In your example, for instance, the 5 prior polls may have already been themselves adjusted according to the Bayesian method. If so, is the implication here that this election seems much closer than it actually will be? And would the best strategy of looking at the polls in any state be to follow the diachronic direction and pay less attention to the actual spreads?
For readers, note that this is a general problem with Bayesian inference, one that crops up a lot in academic output where people are reluctant to ignore prior work due to the citational economy they use. It can very rapidly turn into circular reasoning: someone publishes something plausible but wrong (doesn't replicate or whatever), other authors then include this result in their priors, adjusting their own results as a consequence, which then gets published reinforcing the wrong result, and that new result is then incorporated into yet more priors polluting them still further. That this is happening isn't always obvious. COVID modelling has lots of examples of it in action.
“…there is another way to look at this. If you are a pollster, and you want to present readers with the current best estimate…”
Who said that pollsters wish to present their reader with accurate, or at least, best estimates? Cynic here. I maintain there are two types of polls—internal and external. The political camps pay for internal polling. They need as much accuracy as they can get, bad or good news. Then there are the external polls, done for the “rubes”. These are done to influence the electorate. They may be bought by the political parties, newspapers, or whomever—but they are not designed with accuracy in mind. They are designed to persuade, demoralize, obtain audience as in clickbait. Anything but accuracy. “He who pays the piper calls the tune.”
What we are seeing now is what always happens toward the end of the election process, a regression of the pollsters to the “true” numbers that will shortly be revealed. This is done to save some semblance of credibility for the polling organization.
New subscriber. I came as much/more for the adult statistical analysis and discussion as the juicy topics covered. I figure if I keep reading on a regular basis, (going full Bayesian here) that I might actually move my mean understanding of statistics ever-higher.
Thanks.
There's an old saying in US politics, polls are designed to influence public opinion, not reflect it. So what kind of poll can you trust?
The kind of poll that does NOT get publicized. Only the privately commissioned polls that are not released to the public can be trusted.
About ten years ago I answered my phone to an unknown number from another country.
It was a teenager in a call centre who gave me a barrage of questions about politics with a few market research questions thrown in at the end. She absolutely raced through the questions, couldn’t pronounce the names of political parties, and clearly had no knowledge of our interest in the subject matter she was surveying. I doubt that her data entry was very careful.
In other words, huge potential for the introduction of non-sampling error
Anyway the political opinion poll appeared in my national media the following weekend with n≈1,000. Ever since I’ve been very sceptical about political opinion polls. I think they are exclusively done as marketing by polling companies.
Interestingly when I was surveyed enough there was a topical question clearly paid for by one of the political parties that was never destined for public release
But wouldn't this snowball? In your example, for instance, the 5 prior polls may have already been themselves adjusted according to the Bayesian method. If so, is the implication here that this election seems much closer than it actually will be? And would the best strategy of looking at the polls in any state be to follow the diachronic direction and pay less attention to the actual spreads?
Yes if everybody does this and lies about it, aggregating the data will be very incorrect in the end due to the lack of independence.
For readers, note that this is a general problem with Bayesian inference, one that crops up a lot in academic output where people are reluctant to ignore prior work due to the citational economy they use. It can very rapidly turn into circular reasoning: someone publishes something plausible but wrong (doesn't replicate or whatever), other authors then include this result in their priors, adjusting their own results as a consequence, which then gets published reinforcing the wrong result, and that new result is then incorporated into yet more priors polluting them still further. That this is happening isn't always obvious. COVID modelling has lots of examples of it in action.
“…there is another way to look at this. If you are a pollster, and you want to present readers with the current best estimate…”
Who said that pollsters wish to present their reader with accurate, or at least, best estimates? Cynic here. I maintain there are two types of polls—internal and external. The political camps pay for internal polling. They need as much accuracy as they can get, bad or good news. Then there are the external polls, done for the “rubes”. These are done to influence the electorate. They may be bought by the political parties, newspapers, or whomever—but they are not designed with accuracy in mind. They are designed to persuade, demoralize, obtain audience as in clickbait. Anything but accuracy. “He who pays the piper calls the tune.”
What we are seeing now is what always happens toward the end of the election process, a regression of the pollsters to the “true” numbers that will shortly be revealed. This is done to save some semblance of credibility for the polling organization.
good point. everyone hates Bayes. but being dumb is legit.....
The b statistics that dare not speak its name - nice :D We live in a world of frequentists after all.