Our days surely are the era of polls and surveys. Every day you read the latest numbers about how many consumers use this or how many voters think that or whatever. When the newspaper citing the survey is not a miserable one, it will not only give the percentage, but also the number or people asked for the poll. And those sample sizes often are not nice round figures, but weird numbers (1,003 people seem to be popular). Why don’t statisticians ask exactly 1,000 people? I know that the larger your sample is, the more accurate will be your results. But I don’t think 1,000 vs 1,003 makes all that much of a difference. Is their a sensible reason for odd sample sizes?
In any survey, some people won’t answer, and some won’t answer all the questions. 1003 may reflect the number of people that answered every question.
Sample sizes are calculated based on how much accuracy you want your results to have. Statisticians normally want to be able to say that results are accurate within 5 percentage points, nineteen out of twenty times.
Depending on how many questions you have, the nature of the questions and responses, and, as ultrafilter pointed out, the response rate, you’ll get a different required sample size.
My company oversamples a bit because a few respondents will be ‘bad’ for some reason. So, if you want 1000 respondents, you might do 1020 surveys. 1003 is what you end up with.
Another aspect of the very small oversampling amounts is simply multiple people doing it in parallel. You don’t know if a given survey is going to be completed until it actually is completed. If you’ve got a bank of 5 people each making calls in an attempt to get 1000 surveys done over a couple of nights, when the thousandth actually completes, there may still be three other people on the phone engaged in active surveys. They aren’t going to stop in the middle say ‘Sorry, we don’t need you anymore, goodbye.’ and hang up. They finish the survey, and assuming it was completed appropriately, the results will get included in the pot.