values. We show the list of islands in Table A1. Note that we conducted a panel data analysis of
data from 1975 to 2001 (26 years).
In this paper, we adopt a partial adjustment model and its modification to analyze the data.
First, we set up the following partial adjustment model (note that the subscript i denotes individual
islands and the subscript t denotes time):
Yit-Yit-1 =θ(Yit*-Yit-1)+uit (1)
where Y is the per capita taxable income, θ is the partial adjustment coefficient, and Y* is the
ideal per capita taxable income given by Yit* = α+ βGit + γTit + δNit where G is the per capita
fiscal expenditure, T is the per capita number of tourists, and N is the population size.
Substituting this formula into (1) and adding an error term, we have the following equation for an
empirical analysis in this study:
Yit =(1-θ)Yit-1 +θα+θβGit +θγTit+θδNit +uit.
If we assume Y* is given by a distributed lag model with M lags8, we have
Yit = α+ β0Git + β1Git -1 + β2Git -2 L + βMGit-M + γ0Tit + γ1Tit -1 +γ2Tit-2 L +γMTit-M
+δ0Nit + δ1Nit-1 +δ2Nit-2 L +δMNit-M
If we introduce a parameter that represents the difference between a variable at time t and that at
time t-1, that is, ΔGit = Git - Git -1 , ΔTit = Tit - Tit -1 and ΔNit = Nit - Nit-1 , we then rewrite the
above equation as follows:
Yit*=α+βGit+φ0ΔGit+φ1ΔGit-1+φ2ΔGit-2L+φM-1ΔGit-M+1
+γTit +ψ0ΔTit +ψ1ΔTit-1 +ψ2ΔTit-2 L +ψM -1ΔTit
-M+1
+δNit +ξ0ΔNit +ξ1ΔNit-1+ξ2ΔNit-2L+ξMΔNit-M+1.
Thus, the fiscal and tourism multipliers and the effect of population size in the long run are
denoted by β, γ, and δ, respectively9. Using this formula, (1) is then modified as:
Yit=(1-θ)Yit-1+θα+θβGit+θφ0ΔGit+θφ1ΔGit-1L+θφM-1ΔGit-M+1
+θγTit+θψ0ΔTit+θψ1ΔTit-1L+θψM-1ΔTit-M+1
+θδNit+θξ0ΔNit+θξ1ΔNit-1L+θξM-1ΔNit-M+1+uit.
Applying this model to panel data allows us to relax the constancy of the intercept α . In
particular, the variable α can be divided into three separate components as follows:
αit = μ+ηi +τt
where μ is a constant over all time and islands, ηi is a variable depending only on the individual
island i, τt is a variable depending only on the time t. We have two popular models for this
specification. One is a fixed effects model that treats both ηi and τt as nonstochastic variables,
and the other is a random effects model that assumes both ηi and τt are random variables with