Table 4: Multilevel meta regression results
(1) T-stat |
(2) T-stat |
(3) T-stat | |
T-stat |
0.00503 |
-1.182*** |
-0.108 |
(0.009) |
(0.106) |
(0.412) | |
no. of explanatory variables |
0.0267*** |
0.0422*** | |
product effects = 1 |
(0.002) 0.811*** |
(0.000) 0.400*** | |
logged dep. var. = 1 |
(0.060) 0.769*** |
(0.092) 0.339*** | |
country effects = 1 |
(0.104) 0.0419 |
(0.082) -0.174 | |
published = 1 |
(0.057) 0.103+ |
(0.186) 0.0124 | |
N (in logs) no. of years (logs) no. of countries (logs) annual = 1 time effects = 1 cross section = 1 other preferences included = 1 gravity regression = 1 robust s.e. = 1 single country analysis Constant |
2.520 |
(0.057) 0.859 |
(0.013) (0.008) -0.501+ (0.017) 1.249*** (0.066) -3.395*** -0.432*** (0.110) 0.0385 1.223*** (0.240) 0.0797 |
(2.889) |
(1.148) |
(0.953) | |
ln σu Random |
1.876*** |
1.178*** |
1.130*** |
(0.422) |
(0.289) |
(0.237) | |
ln σe Residual Constant |
3.249*** |
1.909*** |
0.352*** |
(0.055) |
(0.056) |
(0.059) | |
Observations |
173 |
173 |
173 |
LR test vs. linear regression |
3.648 |
17.82 |
141.0 |
Variance partition component (%) |
6 |
18.8 |
82.6 |
Standard errors in parentheses. All variables divided by the standard error.
+ p < 0.1, * p < 0.05, ** p < 0.01, ∙" p < 0.001
5 Conclusion
This paper has attempted a meta analysis of the agoa trade preference literature. The findings of the paper
include evidence of publication bias and also the presence of a genuine empirical effect in the literature.
The evidence of publication bias is corroborated by funnel and Galbraith plots presented in section (3) —
which show clear signs of asymmetry in the plots. Secondly, the evidence of a genuine effect in our MRAs
is also consistent with results obtained in the MST presented in the results section. There are some concerns
though, of the changing signs of our coefficients of interest and non-significance of others (that is, (β0∕α0
and β1 /α1). An explanation of this, might be due to the conservative number of studies included in the
meta analysis. In addition, the presence of several coefficients in each study which requires appropriate
modelling of the MRA might also be an issue. In attempting to resolve this we used a multi-level model as
16