# How many samples account for all the variance?

I have a data set of 100,000. If I want to pick a subset (sample), how can I ensure that it's a "good" sample? I know that I can make it unbiased by picking at random, but how do I know how many to pick to account for all the variance in the dataset?

-
Could you expand a little on what you mean by "account" for "all" the variance? As it stands, the vague terminology means this question does not have a clear interpretation; I'm concerned different readers will have different understandings of what is meant. –  whuber Jul 31 '11 at 19:56
Define good, define unbiased and define account for variance. None of these are sufficiently technical, as being unbiased is a property of a statistic, not the sampling mechanism itself. I can offer you my definitions, but you may not like them. Consider this: there's a concept of measurable sampling mechanism, in which every unit from a finite population has a non-zero probability of selection. If that property is satisfied, one can construct unbiased estimates of the population total using Horwitz-Thompson estimator. Nothing else in sampling world can be guaranteed to be unbiased. –  StasK Aug 9 '11 at 18:52
You can do some preliminary calculations to obtain the standard error for the parameter(s) you want to estimate, given certain sample size. Typically the standard error varies with $\frac{1}{\sqrt{N}}$, so that a sample 4 times as large as another will have a standard error half as large.