Innovation in academe: the diffusion of information technologies

Abstract

This study investigates the diffusion of two early Information technologies across 1348 institutions of higher education: (1) the adoption of Because It's Time Network (BITNET), a precursor to the Internet as we know it today and (2) the adoption of the Domain Name System (DNS) with its registration of domain names, an essential feature of the modern Internet. We find that the time paths of adoption for both generally exhibit the typical S shape found for other innovations. We identify factors likely responsible for the patterns observed and in the process extend the scope of the diffusion literature by incorporating insights from the optimization behaviour of nonprofits. Using a proportional hazards framework, we find that faster adoption occurred among institutions focused on research and doctoral education as well as among select liberal arts colleges relative to nonselect colleges. Faster adoption also occurred for larger institutions, suggesting that they benefited from economies of scale. Adoption was slower for institutions having a larger percent of female faculty members. Also, there is some evidence to suggest that public institutions were faster to innovate than private institutions, while institutions in the South tended to innovate more slowly than institutions located in other regions of the country.

Keywords:

• innovation
• diffusion
• information technology
• higher education

JEL Classification::

• O33
• C41
• L31

Acknowledgements

This research is funded by a grant from the Andrew W. Mellon Foundation titled ‘The Diffusion of Information Technology across Institutions of Higher Education: Effects on Productivity by Type of Institution and Gender’. The authors are grateful to the valuable comments received from two anonymous referees and to Kelly Wilkin and Josh Leesman for their research assistance.

Notes

¹ Strictly speaking, ARPANET was not the Internet until it switched to the Transmission Control Protocol/Internet Protocol (TCP/IP) standard in 1983 (BITNET History, on-line).

² Somewhat later in 1986, the National Science Foundation Network (NSFNET), which was created by the NSF, began connecting NSF-supported university-based supercomputer centers with ARPANET (NSF, on-line).

³ An institution could conceivably connect to the system even if it did not have its own mainframe via a ‘dumb’ terminal and a high-speed telephone connection to a participating institution.

⁴ By 1992 to 1993, the number of academic organizations connected to the Internet actually exceeded the number participating in BITNET and by 1993, the number connected to BITNET began to fall.

⁵ Data and graphs depicting the exponential growth of the number of hosts (as well as domains) on the Internet are available in Zakon (2005).

⁶ This includes further development of electronic mail, E-mail (with its capacity to send attachments), the development of the global hypertext system that makes it possible to jump from one document to another simply by clicking on hot spots, and the introduction of Web browsers, such as Mosaic and NETSCAPE, which facilitate searches, access to data and access to and the visibility of one's own work. Here, only the diffusion of DNS is studied because data on the adoption dates of these other innovations by academic institutions are not readily available.

⁷ Not all Internet servers are part of the Web.

⁸ The universe of institutions was initially formed by a careful review of years of institutional data available in Integrated Postsecondary Education Data System (IPEDS). Specialized institutions such as engineering schools were excluded from the start. Institutions were not dropped if they expanded their program offerings and in the process changed their name from ‘College’ to ‘University’. Details are provided in the Appendix along with a discussion of the issues that had to be resolved when matching the adoption data with the reference set of institutions.

⁹ Early studies by Kellerman (1986) and Gurbaxani (1990) also found an S-curve shape for the diffusion of BITNET.

¹⁰ See the Appendix.

¹¹ See, for example, Coleman et al. (1957), Mahajan and Peterson (1985), Labson and Gooday (1994), Karshenas and Stoneman (1995), Gruber (1998), Geroski (2000), Stoneman (2002), Hall and Khan (2003), Hall (2004), Rodgers (2003), Lee and Lee (2009) and Young (2009).

¹² This approach has been very popular in the marketing and sociological literature on diffusion.

¹³ To our knowledge, only one prior study, Getz et al. (1997), directly examined the adoption of innovations in higher education, but they did not explicitly consider the preferences or constraints facing nonprofit educational institutions. Instead, they used the profit-maximizing framework specified by Karshenas and Stoneman (1993).

¹⁴ Indicators of prestige include, for example, research funding, rankings of graduate departments and indices of undergraduate selectivity.

¹⁵ James (1990) assumes that smaller class size, especially at advanced instructional levels, and lower teaching loads increase faculty satisfaction.

¹⁶ In recent years, the states’ contribution has fallen in many cases. As a result, many public institutions have had large increases in tuition rates and have relied more heavily on garnering external grants and contracts.

¹⁷ For example, the Wesleyan University website states that ‘Our faculty brings their scholarship and creative work to their teaching and engage their students as creators of knowledge’. (www.wesleyan.edu/academics/).

¹⁸ Work by Bodenhorn (1997) and Robinson et al. (2001) confirm that faculty at select liberal arts colleges make significant contributions to the respective literatures in economics and the geosciences.

¹⁹ We also investigated using a top-40 designation as the indicator of a select college. Notably, while the direction of the effects is the same, the results are stronger in terms of statistical significance when using the larger group of institutions.

²⁰ In the case of BITNET, we assume a 6-month lag for the latest adopter to learn from prior adopters and to implement the new technology. In the case of DNS, a 3-month lag is assumed, since implementation was considerably easier than for BITNET.

²¹ In many instances, the cost of the complementary investments may dominate the cost of acquiring the new technology. For example, Brynjolfsson (2000) argues that the full cost of adopting an information system based on networked personal computers is about 10 times the cost of the hardware.

²² Missing values for this variable resulted in a loss of 40 observations in the empirical work. An alternative measure of university size initially considered in the empirical analysis was the full-time equivalent enrollment count. Although similar results were obtained using this measure as with the number of faculty, the results using the faculty counts were stronger and these results are reported in this article.

²³ Institutions located in such areas may also have access to more and better information concerning the benefits from registering a domain name and thus may adopt faster than institutions located elsewhere.

²⁴ The starting date is the month that the first adoption occurs. If the data series ends before an institution has adopted the technology, then the elapsed time is calculated to this date and the observation is censored.

²⁵ The proportional hazards model assumes that the hazard functions for different values of the covariates are parallel over time. Weibull and Gompertz models are the two common ways of specifying the underlying baseline hazard. Karshenas and Stoneman (1983, p. 521) report that their results support ‘the findings of other empirical studies that within the class of proportional hazard models, the assumed form of the baseline hazard does not significantly affect the parameter estimates.’

²⁶ For nonlinear models, the former statistic is akin to the F-test that is usually reported for linear regression models. The usual coefficient of determination, R ², cannot be calculated in a nonlinear model; furthermore, there is a lack of consensus as to what is the best way to determine a comparable measure. The pseudo R ² reported here is derived from the likelihood-ratio chi-square statistic proposed by Cox and Snell (1989) and recommended by Allison (1995, p. 248).

²⁷ A hazard ratio can also be converted to a probability by dividing the hazard ratio by the factor (1+ hazard ratio).

²⁸ Following the earlier work of Karshenas and Stoneman (1983), we do not include quadratic or cubic terms for Cum _t in the models estimated. While including such terms would likely be appropriate when estimating the cumulative time path of adoptions, this is not the intent of the present study.

²⁹ For example, while in Model 2, a 1% increase in institutional size increases the hazard rate by 161%, in Model 4, the increase is only 22%.

³⁰ The positive effects of Med, Masters and Urban on the hazard are now also statistically significant when the number of prior adopters is controlled for.

³¹ It also means that we violate the assumption of proportional hazards. Including the significant time-dependent effects, however, corrects for the possible biases in the parameters estimated (Allison, 1995, p. 157).

³² We first estimated a model with all variables that are not time-varying interacted with time and tested for the significance of the time interactions. Model 1 reported here does not differ significantly from the ‘unrestricted’ model that contained all possible time interactions (χ ² = 4.75, 4 df).

³³ The topic is sufficiently important that the editor-in-chief of Science began the practice of dedicating journal space to scientific research in education (Alberts, 2009, p. 15).

³⁴ This included, among others, Martin Dodge, Chris Condon and Ira Fuchs.

³⁵ We deleted one institution from our event history dataset that was listed as pending with a node but for which no given adoption date could be ascertained. Our assumption is that the node was added after the BITNET list was formulated since the nodes list contained data only through the first half of 1991.

³⁶ We found 61 cases where we suspect that the date is later than the actual date the institution adopted a domain name.

³⁷ For the Claremont Colleges, a closely linked group of schools, we assigned all the members the date of the system's domain name and not the date that each individual institution registered its own domain name.