Trade Adjustment Costs in Developing Countries: - World Bank ...

272Tibor Besedeš and Thomas J. PrusaThe brevity of duration times for SITC data is surprising. The other surprising resultis that as we aggregate from the product level to the SITC industry level the estimatedprobability of survival decreases. This paradoxical result is related to the uniquecensoring problems (numerous product code changes) only present at the productlevel. If we compare the SITC industry results to the modified censoring results usingthe product level data we see that that aggregation works as expected (Figure 17.1).Two important lessons emerge from the SITC analysis. First, our benchmarkcensoring approach is overly cautious; we classify too many relationships as censoredwhen they actually are failures. Second, SITC data confirm that short durationis not a result of overly fine parsing of the trade data. The aggregationexercise confirms that the main finding is not an anomaly. Most trade relationshipsare short-lived.We considered several alternative approaches toward the issue of multiplespells. First, we simply limit the analysis to relationships with a single spell only.We find very little difference between distributions for single spell and benchmarkdata, especially for TS data. The estimated survival function for single spell datahas a similar pattern as benchmark data: high hazard in the first few years followedby a leveling off of the survival function. We do find that the single spelldata have significantly higher survival than the benchmark results, but we findthat most of the difference is explained by the greater fraction of relationshipsthat are censored in the single-spell data. When we re-estimate single spell datausing the modified censoring approach we find the median survival time is nowthree years as compared with two years in the benchmark data. We also exploredlimiting the analysis to first spells—relationships with just one spell and the firstspell of relationships with multiple spells. The results are generally similar to thesingle spell results and are available upon request.Second, we considered the possibility that some of the reported multiple spellsare due to a measurement error. Specifically, if the time between spells is short,it may be that the gap is mis-measured and interpreting the initial spell as ‘failing’is inappropriate. It may be more appropriate to interpret the two spells as onelonger spell. To allow for such misreporting, we assume a one-year gap betweenspells is an error, merge individual spells, and adjust spell length accordingly.Gaps of two or more years are assumed to be accurate and no change is made.In comparison with benchmark data, the average spell length is less than a yearlonger. The 1-, 4-, and 12-year survival rates in gap-adjusted data are about 7 to9 percentage points higher than in benchmark data.4. DOES SHORT DURATION IMPLY POOR MATCHES?While it might appear on average that bilateral trade patterns are stable, a closerlook at individual product trade patterns reveals that trade is fraught with failure—abouthalf of all relationships fail shortly after they get started. The resultssuggest that relationship-specific investments might be important. To explorethis possibility, we apply results from the Rauch and Watson (2003) matchingmodel to trade duration data.