7
Although analyses always include all available data, some of the dependent variables
often suffer from considerable missing data, reducing sample size in many analyses. At times,
for instance, there is insufficient data to test the PBC hypothesis within the African context.
Six different dummy variables are employed in the analyses, representing the different
pre-election and post-election effects discussed above. The dummy variable utilized in any
particular table is always defined at the base of that table. Election (1), used for annual
datasets, is set equal to 1 for election years and 0 in all other years. Given that most elections
occur in the last quarter of the year, this coding best captures electoral effects just prior to
elections. Election (2), for quarterly inflation data, is set equal to 1 for the election quarter and
the quarters just prior to and just after an election. This identifies inflationary surges around
elections. Election (3), for quarterly data, is equal to 1 for election quarters and the three
preceding quarters. This captures policy manipulation during the year preceding an election.
Election (4), for quarterly data, is equal to 1 in election quarters and the three subsequent
quarters, and this is used to identify post-election devaluations. Election (5), for annual
unemployment data, is set equal to 1 during the eight quarters after elections if the chief
executive is right-leaning, and conversely set to -1 during the eight quarters after elections if the
chief executive is left-leaning. Partisan is set to 1 during each quarter where the chief executive
is right-leaning and -1 during each quarter where the chief executive is left-leaning.
5
Specification of Empirical Tests
The specification follows that of Alesina, Roubini, and Cohen (1999), namely pooled
time-series cross-sectional analysis using fixed-effects estimation to address cross-sectional
omitted variable bias, and including two lagged values of the dependent variable to address
autocorrelation problems.
6
5
I cannot in good faith offer the reassurance that the results are robust to alternative leads and lags. Some
results do remain robust, but I have not yet extensively subjected all analyses to a variety of alternative
leads and lags. The lags and leads utilized in the reported analyses are mostly modeled after those by
Alesina, Roubini, and Cohen (1999).
6
The optimal specification number of lags was determined by Akaike and Swartz tests (Mills, 1990: 138-9),
and in some instances this suggested inclu ding up to seven lags of the dependent variable. As a practical
Convention