a result, non-flat countries will emulate their more successful peers and be inclined to
adopt flat taxation.
The next section turns to the tests of these hypotheses on the likelihood of adoption
of the flat tax regime.
4 Model Specification
In this section we isolate the rational learning pattern of diffusion from correlated domestic
or external sources of that trend. We collect data from 20 Eastern European and EU
applicant countries over a period of fifteen years, from 1990 to 2005; as such, the unit
of analysis here is country-year. We confronted the potential bias caused by listwise
deletion in the face of missing observations in our dataset by using Amelia (Honaker,
Joseph, King, Scheve, & Singh, 2003) to impute for missing values based on patterns
in the existing observations.23 Reported coefficients and standard errors are averaged
across ten imputed datasets. To correct for the serial correlation that plagues time-series,
cross-section datasets, we employ year fixed effects, which also aid in ensuring that our
results are not simply due to idiosyncracies of a particular year.24
In measuring ideology, we use expert survey data (Benoit & Laver, 2006) (in 2003
the surveys were applied to all countries in Eastern Europe, inter alia).25 For each
23Observations were not missing at random; earlier years as well as less-developed countries tended
to have more missing values. This pattern makes it feasible to employ imputation based on existing
relationships in the observed data. There was 16 percent missingness in our base model for imputation.
24Because of the high degree of correlation with our diffusion variables, we do not employ country
or region fixed effects; it is superior to substantively model space rather than just to include dummy
variables. Due to this diversity of flat tax countries and a battery of parameters which account for country
characteristics in the model specification, we feel sufficiently confident that we address Galton’s problem
(1889) and are able to distinguish between effects arising out of diffusion and those out of clustering and
domestic characteristics.
25There are several ways of measuring party policy positions, including methods based on content
analysis (Budge et al., 2001), roll call voting patters (Hix et al., 2005; Poole, 2005; Poole & Rosenthal,
1997), computerized word scoring techniques (Laver, Benoit, & Garry, 2003; Benoit & Laver, 2003) or
opinion surveys (Hug & Konig, 2002; Konig & Slapin, 2004), all of which have their advantages. But in
17