follow an autoregressive process, we apply the Yule-Walker method. Finally, when the OLS
residuals are homoskedastic and do not follow an autoregressive process, we simply use
OLS to estimate regression model (3).14
Panel A of Table 4 outlines the estimates for our entire sample, covering the 08:1976-
07:2001 25-year period. Recall that corporate bonds carrying a standard call provision
dominate the data during this sample period. As expected, yield spreads are negatively
related to both the three-month bill yield and the slope of the riskless term structure for all
ratings. In general, both level and slope coefficients monotonically decrease with credit
quality. These results agree with Duffee's (1998) results for his callable U.S. bond portfolios.
***Insert Table 4 here***
Panel B of Table 4 reports our results for the 01:1995-07:2001 period. For the
economically callable AA- and A-rated indices, both level and slope coefficients are negative
and statistically significant. Recall that BBB-rated bonds are economically noncallable
during this period. Our estimated level and slope coefficients for the BBB index are both
statistically insignificant. Applying regression model (3) to his U.S. noncallable bond data,
Duffee still finds both level and slope coefficients to be significantly negative, although
weak. He attributes this result to the coupon level effect. Our results support this assertion.
Since the coupon bias is not prevalent in our SCM data, the negative sign for both slope
coefficients disappears.
Recall that Figure 1 demonstrates that the probability of a call is significantly higher
for AA-rated bonds compared with A-rated bonds. The results reported in Panel B of Table 4
are in line with this hypothesis. For the entire 25-year sample period, the estimated slope
coefficients monotonically decrease with credit quality, whereas for the 01:1995-07:2001
period, this monotonicity is reversed in line with the moneyness of the doomsday call. Thus,
applying Duffee's (1998) regression analysis to our data strengthens the hypothesis that
21