Page 122 - DECO503_INTERNATIONAL_TRADE_AND_FINANCE_ENGLISH

International Trade and Finance

Notes It should be noted at the outset that all of the states analyzed in this study are advanced industrial
countries. This precludes, for example, an assessment of whether our findings vary depending on a
state’s level of economic development. It is also clear that caution must be exercised when offering
generalizations based on an analysis of such a limited time period. But since the tendency for advanced
industrial countries to rely on NTBs became increasingly pervasive during the 1980s and virtually no
quantitative cross-national research has been conducted on the issues addressed here, our results
should provide a useful first cut at the hypotheses presented above.
Regression diagnostics
Before proceeding, a number of issues regarding the regression results presented need to be addressed.
First, in a cross-sectional analysis such as that conducted here, one concern is that the errors in the
regression (e ) will not have a common variance. Under these circumstances, the OLS estimates will
t
be heteroscedastic and therefore inefficient. White tests yielded no evidence of heteroscedasticity in
the present case.
Second, as noted above, it is important to ensure that our decision to pool data across 1983 and 1986
was appropriate. It is obvious that if the effects of the independent variables in equations (1) and (2)
on NTBs vary over time, this procedure would be inappropriate. However, analysis-of-variance tests
yielded no evidence that the model is unstable between 1983 and 1986, and Chow tests also yielded
no evidence that any of the individual regression coefficients is unstable across time.
Third, the use of OLS might be inappropriate because the value of the dependent variable (NTB) is a
proportion (and therefore is bounded by zero and one). Under these conditions, OLS estimates may
be inefficient and predict proportions of NTB that exceed one or are less than zero. Since it is well-
known that a Tobit model can be used to deal with these problems, we estimated the parameters in
equations (1) and (2) using that model. The results were virtually identical to those. Apparently, the
fact that the dependent variable is truncated poses no problem here.
Fourth, we should examine the extent to which our results are sensitive to influential observations.
To this end, we estimated the parameters in equations (1) and (2) after deleting observations one at a
time. Our findings indicated that the signs and statistical significance of the regression coefficients
are quite robust with respect to the deletion of individual observations. Regardless of which measure
of economic size was used, there was no case in which the sign of a regression coefficient changed,
and in fewer than 5 percent of the cases did a regression coefficient fail to remain significant once any
observation was removed from the sample. Indeed, this number of changes would be expected by
chance alone.
Fifth, it is important to ensure that our results are not vitiated by multicollinearity. The results of
bivariate correlations and auxiliary regressions indicated that all of those few cases in which
multicollinearity might be a problem involved PR and TARIFF. To determine whether the effects of
the remaining variables in the model were sensitive to the inclusion of PR and TARIFF, we deleted
these variables (as well as PR · SIZE) from the model individually and in combination. The signs,
sizes, and levels of statistical significance of the remaining variables in equations (1) and (2) were
quite robust with respect to the inclusion or omission of these variables.
Finally, it is useful to consider the possibility that variables omitted from our model may influence
the findings. Particularly important in this regard is whether the extent of government intervention
in the economy influences the rate of unemployment and its propensity to impose NTBs. We therefore
included in equations (1) and (2) the ratio of government expenditures to GDP in year t, which is a
measure of government intervention. The results of this analysis indicate that the regression coefficient
of this variable is negative and statistically significant in every case. This may reflect the tendency for
states characterized by high levels of government spending to buffer and compensate societal groups
disproportionately, thereby reducing demands for protection from groups that are adversely affected
by imports. It is important to note, however, that the inclusion of this variable in our models has no
effect on the sign or level of statistical significance of any remaining variable, including UNEM,
UNEM· log CONST, and UNEM· SIZE· log CONST. Nor are the quantitative effects of the variables
in equations (1) and (2)—and, hence, the conditions that maximize and minimize the incidence of
NTBs—influenced in any substantial way by its inclusion.

116 LOVELY PROFESSIONAL UNIVERSITY