each district, which is measured by the proportion of votes in each district in the former
election.
25
However, there are some districts in which only two of the major party
candidates were contested either in the former election or in the 2000 election, which
caused a missing data problem. Generally, these partially contested electoral data pose a
serious problem in the analysis of multiparty electoral data. In my knowledge, the most
appropriate way to deal with this problem available up until now is to employ the
multivariate t imputation model developed by Honaker, Katz and King (2002). But as a
second best method, I will treat any observation with one missing value either in the 2000
election or in the 1996 election as “missing variables at random” and exclude them from the
computation of Vote.
26
4.3. Designing the Regression Test
The basic OLS model is specified as
(1) Vote
ij
= +
1
*List
ij
+
2
*Incumbency
ij
+
3
*Party Strength
ij
+
4
*List
ij
*Incumbency
ij
+
where the dependent variable, Vote
ij
, is each candidate’s vote share in district i and party j.
This model will examine the general impact of the campaign on voters’ voting behaviors. If
my Sincere Persuasion Hypothesis about CAGE’s campaign were correct, List would be
statistically significant and negative for swing voters’ regions. In a plain word, listed
candidates would receive fewer votes in regions mainly consisting of nonpartisan voters in
25
There were small changes in the electoral district in Korea after the 1996 election. In the case of the
combination of districts I averaged votes from two districts, and in the case of the separation I used the same
value.
26
This temporary prescription to partially contested electoral data surely has its own problems. As I said in
footnote 24, I am still working on a project to update the analysis using the method Honacker, Katz, and King
(2002) suggested.
Convention