from Elsevier Science-NL, Sara Burgerhartstraat 25, 1055 kV Amsterdam,
The Netherlands.
- In view of this difference, the data used for the 1981 regression were reexam-
ined, and no irregularities were found. The decline in significance was particu-
larly noteworthy for the R&D variable.
- Thus, the procedure involved matching firms present in the original 303-firm
Compustat database with identical firms in the NSF/Census database on a
year-by-year basis: (1) pairwise elimination of cases with missing data; (2)
reestimating the original models using the Compustat data; (3) substituting
the NSF/Census R&D data for the Compustat R&D data in the relevant
firms; and (4) reestimating the original models, once again using the hybrid
NSF/Census–Compustat database.
- This implies that the richer content of the NSF/Census data set regarding the
R&D activities of a firm can be brought to bear on questions impossible to an-
swer with the Compustat data alone.
- Closer examination of the firm-level data might help to explain the reasons for
these differences.
- The fact that the variances were significantly different in 1978 and 1979 and
the means differed in the opposite direction in 1979 raises questions about the
way R&D expenditures were reported in these years.
- It is noteworthy that the differences in variances of the two series become
insignificant in 1980–1982 when the sample size drops from 12 to 11,
thus suggesting that a single firm could have accounted for the 1975–1979
differences.
- This result is noted in Guerard, Bean, and Andrews (1987).
- Thus, the COMP303 data set covers firms with a wider range of R&D expen-
ditures and has a true zero point. Inasmuch as the constant term is not signifi-
cantly different from zero for any of the samples across the five years, the
COMP303 model seems the most plausible. To put it another way, it seems im-
plausible that a firm that did no prior R&D would never spend money on
R&D. Clearly, some firms must launch R&D programs even though they pre-
viously had none. The COMP303 model can accommodate this event while
the others cannot.
- The change in sign between 1981 and 1982 may reflect the effects of the R&D
tax incentives associated with the 1981 tax reforms. This should be examined
as a separate issue.
- The dividends and new debt equations were changed by the reduced time
frame. In the dividends equation, new debt was no longer significant, a finding
that is consistent with the work of Dhrymes and Kurz (1967) for the
1947–1960 period. New debt, in turn, is influenced by capital expenditures,
but not by cash flow. The relationship between dividends and new debt
changes from negative to positive, although the relationship is not well-
behaved. It was positive in three years and negative in two, all five being statis-
tically significant. Thus, it appears that the new debt equation is highly sensitive
to the change in time frame. While a positive relationship between new debt
266 NOTES
