B. Resumen en español
B.4. Pérdidas por leakage en guías SWG
B.4.2. Demostración experimental de la relación entre el índice efectivo
This section discusses additional robustness tests. Results are unreported but available upon request.
In DealScan, loans are grouped into “packages,” wherein larger deals often include multiple loans. The multiple loans within a package share a common borrower, but the composition of the lending syndicate and loan terms often differ. The loans within a single package might be contemporaneous (for example, a lending syndicate offering a short-term loan and a revolving credit line at the same time) or might occur at different points in time, as in the case of a renegotiated credit line. In the dataset here
described, there are on average 1.4 loans per package. While loan and syndicate characteristics are not necessarily fixed within a package, some degree of correlation might exist. Further, borrower and country characteristics are clearly clustered at the package level. Accordingly, the assumption of independence, crucial in many of the statistical methodologies applied in this analysis, does not hold across loans originating from the same package. To alleviate this problem, most of the analysis so far presented clusters standard errors at the package level. As a further robustness test, I replicate all of the results presented above with a reduced sample, including only one loan per package. In particular, for each package I select the earliest loan (based on the loan initiation date) or, if multiple loans share the same “earliest” initiation date, the largest loan (measured as total loan value, in USD) amongst the contemporaneous ones. In unreported results, I find that using this reduced sample does not affect any of the findings presented in Tables 4, 5, 7 and 8. Results presented in Table 6 do not employ clustering, due to a lesser dependence problem, as the decision to retain a certain share of the loan is taken at the loan level, not at the package level.
In the sample used here, loans are not only clustered at the package level, but they are also clustered at the borrower level, as borrowers are at times recipients of multiple loans. Hence, I re-estimate all parameters from the various models presented in Tables 4, 5, 7 and 8 by clustering standard errors at the borrower, rather than package, level. The main findings are unaffected.
As an alternative to the Fraser Institute’s measure of protection of property rights, I use the “Investment Profile” score from the International Country Risk Guide and re-estimate all the results presented in Tables 4 to 8. Aside from differences in
parameter estimates magnitudes (largely due to differences in scaling of the two indices), all results and levels of statistical significance are unaffected.
State ownership of the lender is defined in this paper as majority ownership, meaning direct or indirect control of over fifty percent of equity of the institution by governments or state-owned entities. Yet DealScan identifies minority state ownership as well, where “minority” is defined as exceeding five percent of equity. Making use of this data, I replicate this analysis by identifying government lenders in which state ownership, direct or indirect, exceeds five percent of equity. While signs and significance levels of the coefficient estimates presented in Tables 4 to 8 are unaffected, the magnitudes of the estimated coefficients are somehow smaller. The weaker impact could be either due to weaker government interference or to more noise in the data – as discussed, the dataset used in the main analysis has been extensively validated and multiple errors have been corrected, while no such data validation was performed on the variable identifying minority state ownership.
One of the metrics employed as a measure of access to external financing is the number of previous private loans obtained by a borrower over the past five years. This metric is biased (downwards) for the early years of the sample. To check for robustness against this bias, I replicate the analysis by excluding loans initiated during the first five years of the study period, 1980-1984. The findings presented in Tables 4 to 8 are unaffected.
Due to both the size of the economy and reporting biases, DealScan is heavily biased towards loans originating from the United States. In order to check whether the main results are driven by this bias, I first add a binary variable identifying borrowers
with headquarters in the USA to the predictors in Tables 4 to 8 and find core results to be unaffected. As a second robustness check, I exclude loans to borrowers headquartered in the United States from the analysis. While the statistical significance of some of my results is somewhat reduced, likely due to the smaller sample size, coefficient estimates presented in Tables 4 to 8 are largely unaffected. That is not surprising, given that government loans are rare in the United States and mostly associated with foreign government lenders.
Descriptive analysis also indicates that the subset of loans involving state-owned lenders includes a substantial portion (approximately one fifth by count) of loans to Chinese companies. Accordingly, I replicate tables 4 to 8 excluding loans to borrowers based in China. As in the above-described robustness check excluding USA-based borrowers, while the statistical significance of some results is somewhat reduced, likely due to the smaller sample size, coefficient estimates are largely unaffected.