|
Edited on Sat Nov-13-04 02:31 PM by barackmyworld
I am a (female) Harvard student, and I've taken statistics classes. Tests to find the margin of error, or probability that two numbers are significantly different, etc all have conditions. For example, the population has to be much larger than the sample size (exit polls are fine for that). It also has to be a random sample. Yes, you can do stat tests on a non-random sample, but they won't tell you much. I'm not contesting that there are discrepancies, or that they might be large. But stat tests are not appropriate to show that. Many stat tests in areas of sociology, medicine, and other areas are discounted because there was a bias in the sample.
edit: I did some googles, and I can't find any statistical test that could be used in this case that doesn't require a random sample. Also, I want to reinforce the point that you can use stat tests here, but there is no way to judge the accuracy of the results. You can do statistical inference about the United States from a random sample of two people, but it won't mean anything. Let's say you asked those two people if they liked pink or blue better, and you wanted to compare that to the national statistic that 50% of people like pink, and 50% like blue. If by chance, one of those people liked pink, and one liked blue, it would mean that your test "is accurate," but there is ALMOST as great of a chance that they would both like pink, or both like blue. Then, your study would show "a massive difference in pink vs. blue preferences!" This is an example where if you violate one of the requirements for inference (usually you need 30 people for a z-test, 10 for a t-test if my memory serves me correctly), your results are not mathematically accurate.
|