Since today is Super Tuesday I took quick read of Brian Knight and Nathan Schiff's Momentum and Social Learning in Presidential Primaries.
1. I liked the paper - at about 20 pages it covered a lot:
This paper provides an investigation of the role of momentum and social learning in sequential voting systems. In the econometric model, voters are uncertain over candidate quality, and voters in late states attempt to infer the information held by those in early states from voting returns. Candidates experience momentum effects when their performance in early states exceeds expectations. The empirical application focuses on the responses of daily polling data to the release of voting returns in the 2004 presidential primary. We find that Kerry benefited from surprising wins in early states and took votes away from Dean, who held a strong lead prior to the beginning of the primary season. The voting weights implied by the estimated model demonstrate that early voters have up to 20 times the influence of late voters in the selection of candidates, demonstrating a significant departure from the ideal of "one person, one vote." We then address several alternative, non-learning explanations for our results. Finally, we run simulations under different electoral structures and find that a simultaneous election would have been more competitive due to the absence of herding and that alternative sequential structures would have yielded different outcomes.
2. As expected I was lost by Equation (6) which I think needs to be explained a little. I don't believe it was derived from the earlier equations but I may be wrong. It may work better as part of an appendix.
3. More details on estimation of the multinomial logit would also be good for doofuses like me -- again as an appendix. It looks like they take a state-by-state approach to estimation but this wasn't clear and for a data-centric approach, I think some discussion on how the independent variables were created would also be good.
4. It's hard to tell how general these results are since it is only applied to the 2004 Democratic primaries and a cookbook approach (again in an appendix as part of item 3) would help future analysts work through other future data.
5. Figures 3,4,5 might work better as one figure with 3 panels. Maybe some of the other weight figures as well.