It's pretty easy to make more out of a statement like that than the data really implies, and people have been rightly skeptical. But in chipuni's discussion thread (friends-locked), bernmarx links to the original study [PDF], and so we have the opportunity to unpack it a bit.
Is it statistical trickery?
Page 26* breaks it down with more clarity; the 21 percent includes 8 percent who are "absolutely certain" about their belief in "God or a universal spirit"; 7 percent "fairly certain", and 6 percent "not certain". The next page subdivides that 21 into 6 percent who believe in a personal God, 12 percent an impersonal force, and 3 percent other/both. Related questions get similar levels of response; 7 percent of atheists believe the Bible is the word of God, 18 percent believe in life after death, and 14 percent believe "angels and demons are active in the world." This isn't a fluke on a badly worded question.
For sample size, N=515 atheists (P. 111), which means the numbers are +/- 4.3 percent at 95% confidence. Methodology is described starting on P. 174 and is apparently fairly standard for surveys of this type.
(Side note: Could people be flat-out lying, giving random or otherwise useless answers? I suppose it's possible. I can't find any data on how often that occurs or how it's controlled for. Anyone with more of a statistical background is encouraged to speak up in comments.)
And the study's put out by the Pew Forum on Religion and Public Life, not some no-name think tank with an axe to grind.
Is it linguistic trickery?
Briefly, maybe, depending on how picky you want to be and where you want to place the blame.
"Atheist" was determined by self-identification and was an affirmative choice, not a catch-all category (P. 210). They had to either specifically pick "atheist" out of a list, or if they said something like "nothing in particular", had to clarify "atheist" when specifically asked about it. There might be confusion about the term "atheism," but if there is, it's in the mind of the survey respondents, and the problem isn't with the questioners but with the general population.
The question asked was "Do you believe in God or a universal spirit?" (P. 227) Then, of the ones who answered yes, they were asked of the certainty of their belief and to clarify between personal being and impersonal force (P. 228-229). The argument that this might cause some atheists to answer "technically, yes" due to the "universal spirit" part has some merit -- look at the number of atheists who say they are absolutely certain of their belief in a personal God, and the number plummets to 3 percent. But, again, if this is the case it's a problem with definitions among the general population, and not the survey itself.
The real question is, what causes this disconnect? The survey itself tells us little. Perhaps a fraction of the American irreligious misread "atheist" to mean "against God," and should more properly be classifying themselves as Satanists. Perhaps a fraction misread "atheist" to mean "disbelieve in a personal God," and could better call themselves Deists. It's easier to believe in linguistic error than in willful stupidity.
Does it show that atheists are stupid, crazy and/or hypocritical?
No, no, no. "Atheist" has a handful of different and almost mutually exclusive meanings, to the point where Wikipedia has to distinguish between "strong" and "weak" atheism to even write a meaningful definition of the word. It may seem obvious that "atheism" should imply an active disbelief in higher powers, but people use the term in all sorts of ways (even sometimes when they just mean "agnostic"), and that doesn't make them prima facie wrong.
And let's be clear here: We're talking about 21% disagreement. If you can get 79% of any group of allegedly like-minded people anywhere to agree on anything, you're doing pretty well. This goes double for small and often insular subgroups like atheists (or open source advocates, furries, LARPers, etc). And it's arguments about what the group is actually about that draw the most heat. Believe me, I'm Otherkin, I know.
Still not convinced? Alright. Then let's assume atheists have a 21% nutbar population. This still puts them approximately 7% ahead of the population at large.
* * * * *
Also, apropos of nothing: props to hafoc for utterly out-Frosting me in the tech support poetry thread.
* The PDF has four pages of cover, credits and index; I'm following the numbering scheme of the document itself. In order to immediately locate any reference, jump to (page number + 4) in the PDF you download.