How Many Metas?
Let's build a simple model. Assume I'm a New Hampshire Democratic primary voter. I want to vote for the candidate I like best—let's say Dean—but have a stronger preference for picking someone who's electable, and I'd also like for whomever's going to win to come out of the primarys strong. That is, if (say) Kerry's going to win anyway, I'd rather he be the clear favorite so he can start focusing on Bush rather than sniping at other candidates. The Iowa results, then, have the clear potential to set off a social cascade effect. And I think it's pretty straightforward to show that the level of calculation people go through in deciding how to vote will determine the strength of that cascade.
Call my intrinsic preference—the candidate I like best on policy and personality—my first order calculation. The second order calculation weighs in the probabilities of victory gleaned from Iowa, which will factor into my reasoning, but may not be dispositive. But then there's my third order calculation: I assume other people in NH are also doing a second order calculation, which may further increase my sense that Kerry's likely to win. The effect is further magnified as I do more metacalculations. Obviously, that doesn't mean we all ultimately end up converging on Kerry—some people want to make a statement, and they'll vote for Kucinich. But there's definitely a feedback effect of some magnitude.
Here's the more controversial part: Because most people don't actually do many metas in their own calculations, the speed of the news cycle, which feeds back the previous iteration of the calculus for us, should correlate with the magnitude of the cascade effect. That is, each poll that's conducted reflects people's prior electability estimates and then feeds back into them. I mention this because Adam Clymer's got a Times op-ed today suggesting that a bigger gap between Iowa and New Hampshire might have muted this follow-the-leader effect. But if the feedback effect I'm imagining is sufficiently strong, just the opposite might be the case.