∆EPS, then we may have difficulty discriminating earnings manipulation
from overall behavior. Consider the regression of ∆EPSt+ 1 =α+β∆EPSt.
We estimated this relationship using fiscal-year observations but condition-
ing on ∆EPSt>5 to exclude contamination of threshold effects. The esti-
mated βproved to be close to zero (−0.05) in our sample; thus, a reasonable
null hypothesis is that the ranking of firms by ∆EPS over the next year will
be unrelated to this year’s performance. Moreover, in both the cases below,
we find that the “strongly surpass threshold” never underperforms the “sur-
pass threshold” group, reflecting some persistence in earnings growth in the
neighborhood of interest.
If there is significant EM, we expect that the “meet threshold” group—
many of whose members having presumably borrowed earnings—will un-
derperform the groups immediately above (“surpass threshold”) andbelow
(“miss threshold”). This conjecture assumes that the EM effect exceeds
any normal persistence in performance. This is a sharp and quite unusual
prediction.
Table 18.1 column heads show the definitions of our four groups. Then it
reports the mean and median of relative performance by group for the year
following the formation of the groups. First note that, in our benchmark
658 DEGEORGE, PATEL, ZECKHAUSER
Table 18.1
Next Year’s Relative Performance by Groups Formed around the No-Change
Threshold of the Formation Quarter’s ∆EPS; Subsample of Firms with ∆EPS ≥0 for
Formation Quarter
Groups by Performance in Formation Year
D.
A. B. C. Strongly
Annual DEPS Miss Meet Surpass Surpass
(Cents/Share) in Threshold Threshold Threshold Threshold
Formation Quarter −5 to − 10 to+ 4 +5 to+ 9 +10 to+ 14
No. of Observations 1,143 2,220 3,688 4,049
Performance in
following year:
- Mean ∆EPS 9.48 7.10 6.53 9.44
- Median ∆EPS 13 8 9 12
- Wilcoxon test 4.02 −.77 −7.54 N.A.
p-values (reporting
column relative to
next column) .0001 .4389 .0000
Note. N.A.=not applicable.
The Wilcoxon test compares a group’s performance in the postformation year with that of
the next group. Under the null hypothesis that the distributions of performance of the two
groups being compared are the same, the Wilcoxon test is distributed standard normal
(N(0,1)).