[Kaiserschmarrn]

My point is that if you poll lets say 10.000 people then you might get unlucky, imagine if by sheer chance 30% of the people randomly selected are libertarians, then if you asked who they would vote for for president you would most likely get the 30% of the libertarians and maybe 10% of the rest of the group and that would leave an abnormally high polling number for Ron Paul (37%). Clearly getting irregular polls is a possibility and yet I almost never see any poll deviating from the norm, it seems they never get unlucky with their polling and get a group of people that are odd in some way or the other.

It's extremely unlikely to select such a sample.

If we assume that we have 5 mio voters, among them 100,000 libertarians. Then the probability of selecting a libertarian out of the 5,000,000 is 100,000 / 5,000,000 = 0.02.

Now we can use the binomial distribution to find the probability of selecting 3000 libertarians (30%) for a sample of size 10,000. Actually, my software returns a probability of 0 for anything greater than 1000, so I've lowered your number a bit:
1) Probability of selecting 500 or more libertarians (i.e. more than 5% of the sample are libertarians): 5.999582e-73
2) Probability of selecting 800 or more libertarians (i.e. more than 8% of the sample are libertarians): 1.167511e-231

To compare, the probability to win the lottery (6 out of 49) is 7.151124e-08.

So you are far less likely to get a sample like you described than to win the lottery.

[Rilzik]

I shouldn't be posting with all these math people around but...

If you have a random sample, after 1000 the margin of error doesn't change significantly enough to warrant the extra cash and for the most part assuming a actual random sample they should be accurate. That's why you only see samples more then that in internet polls or science generally speaking. That's also why I am always skeptical of polls that have samples under 1000.

Of course in the media and especially regarding politics it has more to do with the questions asked and how they are asked.