|
NIST GCR
02-829
|
(1) |
(2) Number of projects |
(3) Filtered projects a |
(4) Sample projects b |
(5) Sampling probability, percent |
(6) Number responding |
JOINT VENTURE | 118 |
81 |
36 |
44 |
29 |
No university involvement | 47 |
31 |
9 |
29 |
8 |
Universities involved as subcontractors | 42 |
28 |
9 |
32 |
8 |
Universities involved as research partner | 16 |
11 |
9 |
82 |
8 |
Universities
involved as both partner and subcontractor |
13 |
11 |
9 |
82 |
5 |
SINGLE APPLICANT | 234 |
111 |
18 |
16 |
18 |
No university involvement | 106 |
45 |
9 |
20 |
9 |
Universities involved as a subcontractor | 128 |
66 |
9 |
13 |
9 |
Total | 352 |
192 |
54 |
28 |
47 |
The ATP provided the name of a contact person in each of the 54 companies who was then contacted by telephone, explained the nature of the study, asked to participate in a survey, and assured that specific responses would remain confidential and reported only in summary form. Each agreed to participate in the survey. The respective category-specific survey was sent to each respondent. Each non-respondent was re-contacted up to three times on a weekly basis and urged on each occasion to complete and return the survey. Table 1 (column 6) shows the number of surveys received by category of university involvement. (15) The sample for analysis became 47, as shown at the bottom of Table 1. Seven did not respond.
We emphasize, again, that we are aware of the limitations of the self-reported data that we analyzed. While our survey instruments were pre-tested, the possibility that our primary data reflect the personal attitudes of the respondents as well as objective characterizations of their program is still present. Thus, efforts to generalize from our findings should be made with caution.
Reasons for the early termination of the 21 projects were investigated and ranged from the financial health of the participant(s) to lack of research success in the early part of the project: 11 were joint venture projects, and 10 were single company applicant projects. Joint ventures represent 34 percent of the population of ATP-funded projects, but they are 52 percent of terminated projects. Thus, joint ventures appear to have a higher probability of termination than single company projects. Of the 11 joint ventures that were terminated, three included a university as a research partner and two others included a university as a subcontractor. Four of the single company applicant projects included a university as a subcontractor. Thus, 9 of the 21 terminated projects involved a university in some research capacity.
To consider in a
more systematic manner the relationship between university involvement
in an ATP-funded project and the probability that the project will
terminate early, we estimated a probit model of termination probability
conditional on ATPs share of funding, involvement of a university,
type of project, size of the lead participant, and technology area.
A time variable denoting the year in which each project was initially
funded was also included.
To be precise, we
estimated the following model:
Pr
(project i terminates early) = F(Xi ß)
(1)
where F is the cumulative
normal probability function and Xi is a vector of variables
that characterizes project i.
The probit model estimates the probability of an event as a function of explanatory variables. An unobserved indicator variable is a linear combination of the explanatory variables and random standard normal error. When the indicator variable exceeds zero, the event is observed. Thus, the observed response variable is dichotomous, taking the value zero or one. Given the specification of the unobserved indicator variable, the model allows maximum likelihood estimates of the parameters linking the explanatory variable to the probability of the event being studied (Maddala, 1983, pp. 22-23). |
The probit estimates from alternative specifications of equation (1) are reported in Appendix A (Table A1), and the predicted probabilities as a function of key variables are shown in Table 2. Of particular interest is the nature of the relationship between university involvement and termination. Results imply that projects with university involvement as either a research partner or subcontractor have a lower probability of early termination. The probability of early termination decreases as ATPs share of funding increases, although the effect is barely significant, and only for the specification to simulate the results shown in Table 2. Termination rate does not vary across technology area,(16) but projects where the lead partner is of medium size are more likely to terminate early than do the others.
Table 2. Simulation of Probability of Termination of ATP Information Technology Projects Begun in 1991University
involved |
No
university involved |
|
Size of lead participant (50% ATP share) | ||
Small | 0.036 |
0.094 |
Medium | 0.189 |
0.344 |
Large | 0.042 |
0.106 |
Not-for-profit | 0.081 |
0.179 |
ATP share of funding (medium size, lead participant) | ||
Zero | 0.423 |
0.612 |
25 percent | 0.296 |
0.477 |
50 percent | 0.189 |
0.344 |
75 percent | 0.111 |
0.228 |
100 percent | 0.059 |
0.138 |
The top portion of Table 2 presents the calculated probabilities for a project terminating early by size of the lead participant., For this example (information technology projects begun in 1991), the calculated probability of early termination is lower for each size category when a university is involved in the project. Similarly (bottom portion of Table 2), the calculated probability of each termination is lower for each discrete level of ATP's share of funding when a university is involved in the project. Similar relationships exist across other research technology areas. In the population of ATP projects, university involvement is clearly associated with a lower probability of early termination. (17)
Perhaps university participation reduces the likelihood of early termination simply because the projects are more complex and thus project managers may have more difficulty seeing that the project will fail to reach the technical goals until late in the project. Also, more complex projects, even if they fail to achieve their ultimate objective, may still generate knowledge of potential utility to the award recipients.
ESTIMATION OF THE PROBABILITY OF RESPONSE TO THE SAMPLE SURVEY
Only two of the six
categories of university involvement listed in Table 1 (column
6) had a 100 percent response rate. Contact persons in joint venture
projects were less likely to respond, with the least responsive
category being joint venture projects with universities as both
partners and subcontractors (only five of nine surveys were returned).
The probability of survey response was examined using a probit
model to quantify the potential bias because of non-response.
The probit estimates
for a model of the probability of responding are reported in Appendix
A (Table A2). When all of the independent variables are included,
the results are not very significant. The only variable that is
even marginally informative about the probability of survey response
is the dummy for joint ventures with universities as both partner
and subcontractor, (18) which
are arguably the most complex arrangement contractually. Other
factors held constant, contact persons in joint ventures with universities
as research partners and as subcontractors have a lower probability
of response than other contact persons. The associated predicted
probabilities of response by selected technology areas and type
of university involvement are reported in Table 3.(19)
Table 3. Predicted Probability of Survey
Response
Project type | Predicted
probability |
Sample
probability |
Number
in Sample |
Number
of Responses |
JVUS in materials or information technology | 0.27 |
0.25 |
4 |
1 |
JVUS in manufacturing | 0.66 |
0.50 |
2 |
1 |
Non-JVUS in materials or information technology | 0.84 |
0.80 |
15 |
12 |
Non-JVUS in manufacturing | 0.98 |
1.00 |
5 |
5 |
All other projects | 1.00 |
1.00 |
28 |
28 |
In the results presented
later, response bias will be corrected in two ways: (a) by simply
including the dummy for joint ventures with universities as both
partner and subcontractor in estimations to test for response bias (20) and
(b) estimating a full two equation model using maximum likelihood,
where one equation is the equation of interest and the other is
the equation for response probability. The implication of the first
strategy will be that we cannot identify the direct effects of
being a joint venture with a university participating as a partner
and as a subcontractor separately from the impact on the probability
of survey response.
____________________
7. These
data have been analyzed in Link (1996). See also Hagedoorn,
Link, and Vonortas (2000).
8. This section of the Omnibus Trade and Competitiveness Act of 1988 is also known as the Technology Competitiveness Act.
9. The generic term "partner" is being used to refer to a university-industry relationship where the university is either a subcontractor to a single company or to a joint venture or where the university is a research partner in a joint venture, which means that the university is a formal member of the joint venture. To refer to this latter case, we describe the university as a "research partner."
10. Since December 1997, single applicant, large company participants must provide for at least 60 percent of direct and indirect project costs.
11. Participants in joint venture projects must provide for at least half the total costs of the project while single applicant, non-large company, participants are at a minimum responsible for indirect costs.
12. Expected
project duration is agreed upon at the time ATP funds the project.
13. Variability
in these probabilities reflects the fact that the sample size is
constant at nine and that the size of the population of appropriate
projects to sample, by category type, varies (column 4).
14. Copies of the survey instruments are in Appendix B.
15. Because there are multiple dimensions of ATP-funded projects, we do not claim that our sample of 47 respondents is representative of the filtered population or of the whole population in all dimensions. We offer our sample as one sample to consider, and possibly to generalize about, given the stated filtering and selection process.
16. This conclusion needs to be qualified slightly. Because no projects in discrete manufacturing terminated early, these projects could not be included in the models estimated in the first 2 columns of Table A1 (where technology dummies are used). Clearly projects in this technology have a lower early termination rate than projects in the other technology areas.
17. The
information in Table A1 is used
to calculate a hazard rate for the probability that a project does
not terminate early for use in the subsequent statistical analyses
of a sample of ATP-funded projects to control for possible sample
selection bias. To anticipate the use of this variable in later
survey question equations it is important to note that its inclusion
in an ordered probit or tobit is not really econometrically correct
if it actually enters. That is, if the probability distribution
in the termination equation and the distribution in the survey
question equation are dependent, then the appropriate method is
to specify a full maximum likelihood model for the two random variables
and estimate jointly (such a model is outlined in the appendix
to Hall, Link, and Scott 2000). In fact, we found that the termination
hazard and the sample response hazard never entered significantly,
and that joint maximum likelihood estimates did not differ significantly
from our single equation estimates, which implies that sample selection
is unlikely to produce significant bias in our estimates. However,
our sample size is small, so the power of all these tests is low.
18. The
same university cannot be both a partner and a subcontractor in
a joint venture.
19. The sample size in Tables A2 and Table 3 is quite small (only 29 observations), because all projects with large lead participants or whose technology area was electronics, biotechnology, chemicals, energy, or the environment responded to the survey and hence these projects could not be used to estimate the probability equation (they had one or more characteristics that were perfect predictors). In later estimations, a response probability equation was used that does not depend on technology and is therefore defined for the whole sample.
20. As with our analysis of the probability of early termination, the results in Table A2 could be used to calculate a survey hazard rate to be used in the statistical analyses that follow. The survey hazard rate is the conditional probability density of responding to the survey. However, in practice, the only variable that predicted response or non-response in a simple probit model was joint venture projects with universities as both partner and subcontractor. We therefore used a simpler and more robust method to correct for response bias, by including the dummy for joint ventures with universities as both partner and subcontractor directly in the estimated model. Unlike the use of a hazard rate, this correction does not require normality of the response probability equation to be valid. In the case of a single dummy variable predictor, of course, the two approaches for converting any response bias would be equivalent if normality held.
Return to Contents or go to next section.
Date created: October
18, 2002
Last updated:
August 2, 2005
ATP website comments: webmaster-atp@nist.gov / Technical ATP inquiries: InfoCoord.ATP@nist.gov. NIST is an agency of the U.S. Commerce Department |