I spent some time after school doing social science research data handling. It involved a lot of modeling and system design... I will NEVER do that stuff again.
Even with the best study including clearly asked and isolated questions repeated multiply and double blind sampled, then repeated multiple times reapplication of the model in different demographics. You still have the required assumption of honest answers and agreed definitions. The best stochastic modeling can't fix this. You may be able to spot an "open end" in the design but getting a demonstrably "true" result is elusive.
Example: "Are you a man" ( )Yes ( )No
An adult citizen may answer for reasons outside of the intended frame of the question. An adult male may feel ashamed of himself for snapping at his kids in the morning so he might be thinking "No, I'm not a Man, I'm a jerk!" and answer "No" on the question. Or an adult woman who just threw her deadbeat boyfriend out may be thinking "I'm more of a man than that bum!" and check "Yes."
Lies, Damn Lies and Statistics!
It is my best guess that people who own guns have real world reason to not discuss it with pollsters or outright skew the sample (lie.) The result is drastic under reporting.