We invoke three adjustments to the data samples obtained under sample selection criteria 1 and 2 prior to their adoption in formal empirical tests of bond efficiency, quality, and liquidity.7
First, we balance the pre- and post-CDS samples under- lying the efficiency tests. Panel B of Table 1 reveals substantially greater number of observations for the per- iod after the introduction of CDSs than the preceding period. The pre- and post-CDS sample sizes are unba- lanced because the inherent structure of the two main data sets (Bloomberg CDS trades and TRACE bond transac- tions) prevent us from obtaining longer pre-CDS transac- tion history. Most CDS introductions take place early in the chosen 2002–2008 sample period (294 of 332 or 89% of CDS introductions occur in the first three years). The TRACE database commences in mid-2002. Hence, rela- tively limited pre-CDS bond trade history exists for 89% of CDS introductions. To alleviate this disparity in sample sizes, we create balanced pre- and post-CDS samples for regressions underlying tests of market efficiency by trun- cating the post-CDS period to just two years. We eliminate all transactions that occur more than two years after CDS introduction.
Next, we impose event windows in the analysis of bond liquidity and quality. To minimize possible confounding effects of unrelated events that could arise over long time spans (say, if entire pre- and post-CDS time periods are used), we consider a four-year window ([ 2, þ 2] years) surrounding the CDS introduction event to assess the impact of the event on bond quality and liquidity. Given the sparse nature of bond trades, it appears appropriate to use two years of pre- and post-CDS trades for the tests implemented. This chosen window also reconciles with the preceding balanced sample approach we adopt in tests of bond efficiency.
Finally, we use control samples to benchmark the results obtained for the bonds of firms with CDS introduc- tion. We consider two types of control samples: pooled unmatched control sample and pair-wise matched control sample. As Goldstein, Hotchkiss, and Sirri (2007) and Davies and Kim (2009) point out, both approaches of forming control samples have their own merits and applicability, yet each has likely limitations. An unmatched pooled control sample could include bonds that are different from the event sample bonds, which could affect the results. If a pair-wise matched control sample is used, results could be sensitive to the attributes used for matching and the particular choice of bonds selected. Therefore, we follow Goldstein, Hotchkiss, and Sirri (2007) and use both approaches. We form two control samples of bonds by firms with no CDS introduction as follows.