Almalieque wrote:
However, if you were to ask "Do you support abortion" or any other hot topic that is so heavily divided, it becomes increasingly more difficult to be that accurate with the same amount of numbers due to the increased weight of other factors such as the demographics, how the question was worded, how the people were contacted cell vs email vs land line, time of day of the contact, etc. It's not impossible, but it becomes much more about being attentive to details at that point.
Absolutely. You reach a point where there are too many factors to accurately model in a complex environment such as the real world. Nearly every algorithm from mass spectra alignment software, to climate prediction software, to market analysis, to election polling all suffer from this same flaw. In the end you're never going to be able to get a perfect approximation of reality.
So what do you do? Well, you make assumptions, you simplify, condense and ignore certain variables you think will have a minor affect, or that you don't have data on, etc.
In a laboratory setting this would be something like removing (or I suppose simply not adding?) part of an equation, varying a parameter, or changing its weighting through a range of reasonable values and observing how the accuracy of your algorithm changed. For polling I'd reiterate what I said above, the more closely the results match the predicted ones, the less likely it is the poll needs to be altered. Assuming you've been testing the methods on previous data as well of course. Also, throwing out that the more complex algorithm (the one with more values for more more possible factors) is not necessarily the more accurate one. People tend to ***** up more complicated things, and some factors may be best ignored if they prove difficult to model accurately.