Reply #37: Ok, Stevepol - you asked

[qwghlmian]

Here is a "fraud-detecting poll":

You put a pollster in every precinct, sample every 10th voter coming out, ask them whom they voted for, and do minimal weighting for tiny adjustments for demographics where appropriate.

Here is US "non-fraud-detecting poll":

You put a pollster in select (few) precincts, sample every 10th voter coming out, ask them whom they voted for plus all kinds of other questions, total the results, then, after all the results are collected, shift the results so that they match the actual election results. This would compensate for any clustering skewing effect caused by picking just a few precincts and make the data that you're after (such as how many X voted for Y, and what do X hold more important than Y etc) accurate.

Why would the second type of poll's raw results not reflect correctly the official election results? Because the variation between different precincts in the same state, even with a couple of miles distance between them, can be enormous, and if you look at the precinct election results in the US you will see that if you look at a precinct or a few, you would not be able to predict the results of the election in the state with any kind of consistency. The way MysteryPollster explains it is the "marbles" analogy. If you put 50% red and 50% green marbles in a jar, then shake them vigorously to distribute them evenly, then take 100 out, you would get pretty close to 50 red ones and 50 green ones (maybe 48 and 52, or 54 and 46). But if you put in 50 red ones, then 50 green ones, give it only couple of shakes, then take out 100, you can easily get 30-70 or 80-20 results.

Why would you run the second poll and not the first one? Because of the cost. The cost of the first type of poll is hugely more than of the second type. But no one is stopping anyone in the United States from ponying up that money, organizing a couple of hundred thousand pollsters on the election day (I think there are at least a hundred thousand precincts in the US) and doing the "fraud-detecting" poll.