For instance, you have to filter out most posts with links (because they tend to have very little associated data, or are spam), anything in languages you don't have parsers or datasets for, etc. Then you can cut on character count (because there are a surprising amount of useless tweets that are under 5-9 characters)...
As others have and will continue to state, this further cuts down the sample size of an already fairly small sample (500+ million in the world[0]). If you want to be specific and only look at a single nation, you're cutting that sample size based on who has location turned on and/or broadcasts it in their profile. This is a significantly smaller number.
Edit: This is obviously a better sample size than a random poll of individuals, however, but that doesn't mean that the cumulative quality will be drastically higher.
[0] - http://www.statisticbrain.com/twitter-statistics/ (sources from three different input points)