Against this background, the announced availability of the NSF/Cen-
sus data series provided an opportunity to replicate the preceding analysis.
Such a replication was seen as a way of reviewing the perfect markets hy-
pothesis one more time, and of testing the sensitivity of the empirical
model to alternative data. If the empirical results using the NSF/Census
data set were very similar to the Compustat/Patent Office results, it would
encourage further work on the linkages between the data sets (National
Science Foundation 1985).
The current work attempts to replicate the previous (Compustat)
study by substituting the R&D expenditure data from the NSF/Census
database for the R&D expenditure data in the Compustat database over
the 1975–1982 time period. Census Bureau employees performed the
substitutions and provided results of the computer runs to the authors.
Whereas the prior work covered 303 firms from 1972 to 1982, the pres-
ent study uses a sample of 158 to 188 firms over the 1975–1982 time pe-
riod. The matching procedure created by substituting the NSF/Census
data for the Compustat data reduced the original 303-firm sample to
158 to 188 firms by eliminating (nonmanufacturing) firms not engaged
in R&D activities.^3 This procedure permitted comparisons to be made at
several different levels:
1.Annual comparisons of the aggregate R&D data across the whole
sample.
2.Annual comparisons of industry-level data across several subsamples.
3.Examination of the sensitivity of the models to the substitution of
NSF/Census data for Compustat data.
4.Examination of the impact of the compressed time frame on results.While there are several results that are worthy of detailed discus-
sion, the overall results suggest that the substitution of the NSF/Census
data had little effect on the model structure, thus indicating a good link-
age between the NSF/Census data set and the more business-oriented
Compustat data.^4
Comparison of R&D Expenditure Data
When the process of matching was finished, samples ranging in number
from 158 to 188 were obtained for the years 1975 through 1982. Exami-
nation of variances (Table 7.1) and means (Table 7.2) of the complete
sample shows that the R&D expenditure data are not statistically differ-
ent in six of the eight years, but that the 1978 and 1979 samples are sig-
nificantly different. The variance (Table 7.1) of the Compustat data is