with the summary statistics reported in Table (1) in the appendix.
The summary statistics and the studies used in the meta analysis are presented in the appendix (Tables (1)
-(2)). The results of the random effects meta analysis are presented in the appendix also. Figure (1) presents
a visual representation of the summary meta analysis by study. The length of the diamonds in the graph
represent the confidence intervals of the pooled study effects of each study. Three studies have a negative
pooled estimate (Cooke, 2009, Nouve and Staatz, 2003, Seyoum, 2007). Of the three studies only Nouve
and Staatz (2003) reports a significant pooled estimate. The remaining pooled coefficients are positive and
three of these are significant. Table (3a) in the appendix presents the underlying data for the forest plot
discussed above. Table (3b) presents the tests of heterogeneity. Five of the studies display high levels of
heterogeneity reporting I2 values of 69.7% - 99.9%. Five out of the remaining six studies have an I2 value
of zero indicating that there is no real variance between their coefficients. The pooled estimate of all studies
of 0.124 is significant and also displays high heterogeneity (I2 = 99.9%). Thus, the pooled agoa impact is
13.2% ((exp0.124 -1) × 100). Table (3c) reports the z - values and their respective probability values for
tests of significance of the pooled estimates.
Due to the heterogeneity present in the studies (Table (3b)), pooled estimates from the FEM would be bi-
ased and hence its results are not reported in the appendix. However, for comparative purposes the estimate
of the FEM was 0.007 (0.7%) which is much less than the 13.2% estimate reported by the REM. The results
shown in the appendix excluded one study (Mueller, 2008)—the inclusion of this study reduces the pooled
effect to 0.121 (12.9%)2. On the contrary, excluding the six large coefficients (305.1 - 769.5) reported
in Nouve and Staatz (2003) did nothing to alter the results presented in the appendix—the pooled effects
remained the same.
Meta-Significance Tests
We present the MST results (tests of genuine empirical effect) in Table (1). Results in columns (1), (3)
and (4) point to the presence of a genuine empirical effect. The regressors—the log of degrees of freedom,
sample size and square root of degrees of freedom are all significant at the 1% level of significance in each
of the 3 columns respectively. However, columns (2) and (5) find no effects. The square root of degrees of
freedom and precision are insignificant also in the remaining columns. In column (5) we followed Stanley
(2008) by shrinking the t-statistic value to zero and using it as the dependent variable. All three estimates
of the genuine effect are greater than zero and less than half in columns (1) and (3) of the table. A t-test of
α1 = 2 is rejected in all three cases (with t values of 12.41, 12.40 and 2302 respectively) with the result—
αι < 1.
2In the FEM model the pooled estimate was 0.006 after including Mueller (2008) indicating a marginal decrease of 0.001.
10