The models are pretty complicated and incorporate several data sources, sampling is just one part of it. I've seen (very dry) documents that go into the methodology in detail. The model does introduce some obvious biases through assumptions of representativeness but how those interact with the headline numbers is not straightforward.