Quote:
Originally posted by JR@Feb 25 2003, 01:37 AM
when do polls have enough data to become
statistically relevant?
how many users or votes does it take to make a
conclusion that is meaningful?
|
Depends on what one means by "statistically relevant".
If there were just two choices, results would be correct to within 3% after 1000 people voted. It goes as the square root of the number of people questioned. The square root of 1000 is 31. 31/1000 is about 3%. 100 people sampled is a 10% error.
Interestingly, the percentage of the population in question that is polled is not significant. Polling 1000 completely random people out of a population of 250 million is a better measure of the results one would obtain by polling all 250 million people than a poll of 100 people out of a population of 1000 would be.
----------------
A little more in-depth discussion if interested. The 3% number above is the standard deviation. 67% of a population in a normal distribution falls within one standard deviation of the mean.
This means that if we poll 1000 completely random people as to who they will vote for president, there is a 67% chance that will be accurate to within 3% if we had polled ALL of the population.
95% of a population falls within two standard deviations of the mean. So if we poll 1000 completely random people as to who they will vote for president, there is a 95% chance that will be accurate to within 6% if we had polled ALL of the population.
It's important to note that distinction. When we see polls on television or in the newspaper that say "Error +/- 3%", it doesn't mean that if we had polled everyone it WOULD be correct to within 3%, it only means that it will be that accurate 67% of the time.
3 standard deviations is 99.7%. So if on election night, a poll of 1000 says a candidate has 59% of the votes, there is a 99.7% chance that candidate will win .. and a 3 in 1000 chance the trailing candidate will pull it out.