ifPROTECT* lies between μ1 and μ2, 3 ifPROTECT* lies between μ2 and μ3, 4 ifPROTECT* lies
between μ3 and μ4, and 5 ifPROTECT* lies above μ4. The disturbance term is assumed to be
normally distributed.
In this model, the probability that protect will take on the value 1, say, is the probability that
PROTECT* = β,Y + ε <μ1 (4)
(where Y is the vector of all independent variables), which is equivalent to the probability that
ε<μι-β,Y (5)
or
φ(μι-β,Y) (6)
where φ is the standard normal distribution. The probabilities that protect will take on any of the
otherpossible values are similarly constructed: they depend onthe respondent’s characteristics, Y; the
vector of estimated coefficients, β; the cutoff points, which are parameters to be estimated alongside
β; and the standard normal distribution.
It follows that the marginal effect of changing an independent variable on the probability of a
given outcome depends not only on β, but on the standard normal density function evaluated using a
particular choice of Y. A significant positive coefficient implies that changing the relevant
independent variable increases the probability that protect takes on the value 5, and reduces the
probability that protect takes on the value 1. The marginal effect of changing such a variable on the
probability that protect takes on the values 2-4 is, however, apriori unclear. Initially, we will simply
estimate ordered probit models, and will loosely speak of variables being either positively or
negatively related to protectionism; marginal effects will be estimated later.
In nearly all cases, we include a full set of country dummy variables, to take account of
country-level effects operating on all individuals within a country (coefficients not reported). Each
column in Table 4 indicates whether these dummies have been included or not. For all other variables
14