1. When we selected this district we expected it to be a Type 2 district; however, the Democratic race ended up being more competitive than we had anticipated (53.2 percent to 46.8 percent).

2. The interviewers were from Creative Data, Inc. They conducted a total of 2,977 successful interviews: 609 in Assembly District (AD) 9, 573 in AD 49, 644 in AD 53,492 in AD 61, and 659 in AD 75.

3. The survey instrument was extensively pretested; for details of the pretesting or for copies of the survey question forms, please contact the authors at rma#usc.caltech.edu.

4. We assume that voters do not base their candidate evaluations on their expectations of the candidate's winning. We do not believe that voters would have any reason to do so, given the design of the survey instrument: voters willing to engage in strategic behavior are not likely to be bashful about it on an anonymous, selfadministered survey.

5. Twenty-three percent of the registered voters in this district were Republican in 1998.

6. However, because distinguishing between hedgers and impact voters requires us to make distinctions between a candidate having "virtually no chance" and a race being "sufficiently close," we cannot ascertain whether the crossers who are not raiding are hedgers or impact voters. Thus we refer to them simply as "strategic crossers."

7. Steinberg's name was inadvertently omitted from questions 5,8, and 9. To analyze voting in the district, we assumed that voters listing "other" as their vote choice voted for Steinberg based on the reported vote in our sample and what we know of the actual vote on election day.

8. These percentages are based on eleven of the seventeen Republican crossover voters for whom we have complete information and for whom we can make an accurate vote prediction.

9. Our counterfactual analysis is based on earlier academic work of ours in which we have developed techniques for determining how voters might have behaved had the set of candidates or parties which they could choose from in a particular election been different (Alvarez and Nagler 1995, 2000b). Our technique for answering this type of question is straightforward. We begin with a wellspecified model that predicts which party or candidate each voter in our survey sample would select from the full and actual set of parties or candidates in the particular election. (By "well-specified," we mean a predictive model which includes as much information about each voter as possible, both demographic and political information. This ensures that we will have the most accurate predictions possible about the behavior of voters.) We estimate the parameters of this predictive model and use them to produce predictions for how much each voter likes or prefers each party or candidate. We can then remove parties or candidates from the set available to each voter in a hypothetical election and examine how all (or how subsets of the electorate) would behave under these restricted choice conditions. Our procedure in this analysis follows the same steps. We first estimate multinomial logit models which predict how much each voter in our exit poll sample likes or prefers all of the candidates in this race. We then examine only the crossover voters to see which candidate they would have preferred in their own party had they not been able to cross over on election day. Based on this counterfactual example, we compute the percentages of votes each candidate would have received in a hypothetical closed primary. The important assumptions which we make in this analysis are (1) that the same types of voters who turned out in the blanket primary would also have turned out had the previous closed primary system been used and (2) that voter preferences for candidates within their own party would have been the same had the closed primary been held instead of the blanket primary. The predictive model we employ uses the voter's gender, age, and racial or ethnic identification as demographic predictor variables. We also use the voter's opinions about Propositions 226 and 227, their ideological stance, their opinions about the state of the California economy, and their notions of what the most important issues were in this election. The multinomial logit results for candidate choice in Assembly District 61 appear in table 6.4. The Assembly District 61 voter choice model correctly predicts 34.4 percent of the voter choices, with a 50 percent correct prediction rate for Soto,45.3 percent for McLeod,42.0 percent for DeMallie,22.7 percent for Skropos,15.2 percent for Thalman,7.1 percent for Wickman, and 0.1 percent for others.

10. These percentages are based on nineteen of the twenty-three Republican crossover voters for whom we have complete information and for whom we can make an accurate vote prediction.