As we head into the real election cycle, already we're hearing about how the polls are skewed because they sample too many Democrats, or other excuses like that. I recognize the desire to manipulate the results into something more appealing, but it displays an ignorance of how statistical polling works, as 2012 proved.
For a poll to have any statistical value, the sample must be random. A selected sample (whether it's self-selecting, like an online poll, or selected by the moderation, like focus groups) has little to no statistical value, or relevance to the population as a whole.
There are various ways to increase the accuracy of a purely random sample though, and the most common and effective way is known as "stratification". Stratification relies on dividing the population (prior to selection of a random sample) by certain fixed and mutually exclusive demographic traits, and then randomly selecting samples of each strata in proportion to that strata's representation in the population as a whole.
For example, a pollster may divide the population as a whole into "Men" and "Women" - and then sample from each strata in proportion to the numbers of men and women in the population as a whole.
Now, from the complaints I've seen, people seem to think that the population should be stratified by political identification. This causes two major problems, statistically. First of all, the poll is meant to measure political opinion, which is not a fixed quantity. Trying to stratify according to the results that you want makes the entire poll meaningless. Secondly, stratification requires constant and immutable demographics in order to maintain it's statistical value. Political Identification and opinion are not constant, they change all the time.
When you see the Party Identification breakdown on a poll, those are results of the poll, not stratification. If a poll has 45% Democrats responding, that means that 45% of the randomly selected people identified as a Democrat while responding to the poll - not that they pre-selected a sample of 45% Democrats.
All of the "unskewing" theories come from that misconception - and that's why they're all wrong, as we saw in 2012.
For a poll to have any statistical value, the sample must be random. A selected sample (whether it's self-selecting, like an online poll, or selected by the moderation, like focus groups) has little to no statistical value, or relevance to the population as a whole.
There are various ways to increase the accuracy of a purely random sample though, and the most common and effective way is known as "stratification". Stratification relies on dividing the population (prior to selection of a random sample) by certain fixed and mutually exclusive demographic traits, and then randomly selecting samples of each strata in proportion to that strata's representation in the population as a whole.
For example, a pollster may divide the population as a whole into "Men" and "Women" - and then sample from each strata in proportion to the numbers of men and women in the population as a whole.
Now, from the complaints I've seen, people seem to think that the population should be stratified by political identification. This causes two major problems, statistically. First of all, the poll is meant to measure political opinion, which is not a fixed quantity. Trying to stratify according to the results that you want makes the entire poll meaningless. Secondly, stratification requires constant and immutable demographics in order to maintain it's statistical value. Political Identification and opinion are not constant, they change all the time.
When you see the Party Identification breakdown on a poll, those are results of the poll, not stratification. If a poll has 45% Democrats responding, that means that 45% of the randomly selected people identified as a Democrat while responding to the poll - not that they pre-selected a sample of 45% Democrats.
All of the "unskewing" theories come from that misconception - and that's why they're all wrong, as we saw in 2012.
