studies and gravity regressions reduce the reported t-statistics holding all else constant. On the contrary,
annual data and single country studies contribute to higher t-statistic values. All the variables in the preced-
ing are significant at 1% with the exception of the number of years coefficient that is significant at the 10%
level of significance. The random component of the model (ln σu) and the residual standard error (lnσe)
are significant in all three columns. These indicate that the studies are largely different from each other. The
variance partition component (VPC) corroborates this evidence together with the I2 tests discussed earlier
which all point to heterogeneous studies. The VPC calculated are 6%, 18.8% and 82.6% for columns (1)
- (3) respectively. They indicate the level of variance in study level t-statistics that can be attributed to
differences between the various studies in the sample. In column (3), the VPC is large and implies that,
82.6% of the variance is due to differences between the studies. On the contrary, the basic MRA in column
(1) attributes only 6% of the variance to differences between the studies. Thus one observation from the
table, is that increasing the moderator variables and controlling for some of the variation between studies
tends to account for the larger values of the VPC observed in columns (2) - (3). In the appendix, we include
Mueller (2008) to check if the study’s exclusion affects the results presented in this section. Table (4) in the
appendix shows that this is not the case. We observe similar estimates consistent with that presented here
for the multi-level model. In addition, excluding the six large coefficients does not change the results and
the results are similar to those presented here (the results are not presented in the paper).
In concluding, the number of explanatory variables, logged dependent variables, single country studies,
cross sectional data and published studies are consistent in all three tables and maintain the same signs.
The FAT/PET and MST in most cases have provided evidence of a genuine empirical effect and publication
bias. Thus corroborating earlier evidence presented in the funnel and Galbraith plots of section (3). We
do not find strong evidence, that textiles & apparel significantly increase the reported t-statistics. On the
contrary, there is some evidence on agriculture increasing the t-statistic values relative to other products, all
things equal. Last but not the least, we do find some evidence for the number of countries and composition
of countries. However, the direction of the effect is not certain—we do obtain both positive and negative
relationships of the coefficients in the results above.
15