Trade Finance in the Recovery of Trade Relations after Banking Crises
229
Third, there are products that exit more than once (multiple spells). The general approach of the literature to control for multiple spells in duration models is to include in the regressions a multiple spell dummy equal to 1 if the relation has at least one exit during the sample period. However, to control for the fact that multiple spells are time-varying within a relation, a different definition of multiple spell is considered, with the construction of a variable equal to the number of spells before time t. This approach, the authors believe, is theoretically more correct than the standard approach of the literature because it does not consider a relation to be characterized by multiple spells until its first observed reentry, but only after it.16 Due to the high level of disaggregation of the dataset, throughout all the analysis the assumption is that there is a representative firm for each trade relation. This allows the analysis to refer to “experience” and “size” as two measures of heterogeneity among exporters. Because size and experience are not the same (in the sample, the correlation is 0.19), they capture different characteristics of exporters. Notes 1. The World Bank’s forecast is 15.7 percent, and the Organisation for Economic Co-operation and Development’s forecast is 12.3 percent. 2. Using data on U.S. imports at the Harmonized System (HS) 10-digit level of disaggregation from 157 countries between 1995 and 2009, the authors have extrapolated all relations that were interrupted at the occurrence of a banking crisis in the exporting country. 3. The authors chose the United States as the destination country because the original trade data used (from the Global Trade Atlas and the Center of International Data at the University of California, Davis) contains information at the 10-digit level of disaggregation only for trade flows in and out the United States. 4. After inclusion in the regression of the market share of a product to control for product heterogeneity, results do not change. 5. The variable experience cannot be included in a Cox regression because it is highly correlated with the duration of a spell, which is the conditioning variable in duration models. 6. The effect of experience should be interpreted with caution, since it captures both the negative duration effect (the fact that the probability of exit decreases the longer a product has been in the market) and the presence effect (learning by exporting). The authors are only interested in the latter effect, which has an economic interpretation. 7. The indicator of long-term external financial dependence (EFD) comes from Rajan & Zingales (1998) and is computed at International Standard Industrial Classification (ISIC) three-digit industry level. For short-term financial dependence, we use trade credit dependence (TCD) from Levchenko, Lewis, and Tesar (2009), computed at the North American Industry Classification System (NAICS) four-digit level (the original measure is from Fisman and Love [2003]). In the data, the correlation between EFD and TCD is very high and equal to 0.7. 8. For a similar approach, see Besedes (2007), section 3.3.2. 9. Similar results can be shown when using trade credit dependence (TCD). 10. From figure 12.5, it might seem that the variable size is not constant across time. This is controlled for in the regressions by stratifying the sample (see annex 12.1). 11. A Cox regression was also estimated for different groups of export size. The results, available under request, show that neither the financial dependence variable nor the other control variables have a size-specific effect.