Polls, Polls & Polls

Thought I read somewhere the polls, besides Rasmussen, showed trump ahead of Obama at this time.

I think the dumbest of the polls are those who match up the president against a potential contender. These polls are popularity contests and bear no semblance of what happens in our presidential election. It doesn't matter if 60% of the people in the nation prefer candidate A over candidate B, what matters in the election is what states prefer which candidate and the population (electoral votes) each of those states has.

Yeah, that sure makes it seem like some pollsters are pulling our legs.

Take all polls with a grain of salt, but this doesn't look good for the Democrats less than 11 months from November 2020.

The only polls that matter more so are regional or state specific, national polling is useless in a Republic that elects a President based on Electoral College votes. And as we saw in 2016, the only poll that matters is election day.

Sampling and polling is a much more complicated and intricate process than most people know. A few points about political polling, without getting too technical:

(1) One of the primary factors determining the validity of the results is the randomness of the sample. As has been mentioned, political polling is not random for several reasons. If you could take all 200 million American voters, put them in a fishbowl and blindly pick random samples, you could determine which ones are men and which ones are women, which ones had hair and which ones were bald, etc. You could not determine whom they would vote for without asking them, and then the results get muddled with dishonesty, embarrassment and political correctness. If the sample is nonrandom, such as polling only people willing to answer the phone and respond, the results become even more unreliable.

(2) The "sample error", meaning the degree to which the poll results differ from the results if the entire population were tested, is determined primarily by sample size. All other things being equal, the larger the sample, the smaller the sample error. If you could poll all American voters, and they answered honestly, your sample error would be +/-0%. As the sample size decreases, the sample error gets larger. When I see every political poll claim a +/-3% sample error, regardless of whether 500, 1,500 or 10,000 people were polled, I know the +/-3% is a boilerplate footnote that can't be statistically accurate. It's a bullshit disclaimer.

(3) Sample error is only relevant if the confidence interval is known. Typically, but not always, a 95% confidence interval should be used. The sample error can be claimed to be smaller if the confidence level is dropped to 90%, or 80%, or 50%. Without knowing that number, the sample error is meaningless. In the poll above, where only 1,000 of 200 million voters were polled, and it was all done by phone, and the selection could not be guaranteed as being random, and respondees may have had reasons to be disingenuous, and where a bullshit sample error is claimed, and where the confidence level is unknown, the results aren't worth the bandwidth they take up to release.

Luca, I got my MBA in 1989, and you're really wanting to test my memory, aren't you?  OK, without getting into the math:

Sample Error is the difference between one's calculations from a sample (or a subset of the entire population) and the entire population itself.  In this case, the population is 200 million voters, the sample size is 1,000 voters, and Trump got 440, or 44.0% of those.  The +/-3% (if it were actually calculated, rather than just a disclaimer pollsters tack on at the end of every poll result) means if you took another 1,000 samples, your results could be expected to differ from the first sample by up to 3%, or Trump could end up with anywhere from 38% to 44% of the vote.  If your sample size were 100 voters rather than 1,000, and Trump got 44 of those 100, your poll would still show him with 44..0%, but the sample error would be much larger because of the uncertainty associated with the smaller sample.

Confidence Interval (I shouldn't have used Confidence Level) is the probability that the actual population percentages lie within the sample percentages and sample error.  Or in other words, the probability that the calculated percentages and sample error accurately reflect the entire population.  The default confidence interval in most experiments is 95%, but I have no idea what it is for political polls, especially the shady overnight ones that can't possibly sample very many people or eliminate the various forms of polling bias.

It takes a long time and a lot of work to conduct a legitimate poll, and I'm pretty convinced most of these political ones aren't worth any more of a shit than the news outlets that report them.