Research and Advances
Computing Applications

Anticipating Internet Diffusion

Understanding the sociotechnical aspects of Internet growth helps anticipate Internet diffusion, even in countries with vastly different socioeconomic conditions and telecommunication infrastructures.
  1. Introduction
  2. Structural Mechanics of Internet Growth
  3. Feedback Model Calibration and Predictive Validity
  4. Significance of Sociotechnical Drivers
  5. A Different Growth Environment
  6. Conclusion
  7. References
  8. Authors
  9. Figures
  10. Tables

The Internet has meant new opportunities for communication and coordination among individuals, businesses, and nations. Although the U.S. has lead the way in the diffusion of this new technology, diffusion has also become a worldwide phenomenon. Consider the following statistics from such sources as the International Telecommunication Union ( Europe is expected to surpass the U.S. in e-commerce spending on consumer products this year; 50% of the worldwide online community is outside the U.S.; by 2005, non-U.S. Web users are expected to include 700 million of the world’s total one billion users; in India, there is immediate demand for 500,000 more Internet connections; and there will soon be more than 24.3 million Internet users in Latin America. Despite the Internet’s global reach, its diffusion is uneven across countries [3, 11]. Anticipating the growth of the Internet is therefore an important but challenging task for all its various stakeholders.

While descriptive statistics are widely available, models explaining Internet growth appear only infrequently in the literature. Such models are useful, however, providing insight into the mechanics of that growth for policymakers, network operators, equipment vendors, and others on both the demand and supply sides of the phenomenon. A model of Bitnet growth was one such early effort [4]. A more recent source [8] tested three separate models—Logistic, Gompertz, and Exponential—to explain the growth of Internet hosts within the U.S. Providing insight not readily gleaned from descriptive statistics alone, they were based on contagion effects from diffusion-of-innovation theories and assumed that nonadopters of an innovation are increasingly likely to imitate adopters over time [9]. Notably, [8] also found that the Logistic and Gompertz models had less predictive validity compared to the Exponential model. The Logistic and Gompertz models are based on contagion; the Exponential model is memoryless. The authors inferred that contagion effects alone might not completely explain Internet growth. The models ignore external factors, including government policy and sponsorship, as well as technological developments [8]. In short, studies prior to our research confirm the importance of contagion effects but suggest that contagion alone reflects an incomplete understanding of the mechanics of Internet growth.

We build on earlier contagion models, motivated by the observation that Internet diffusion is both a social and a technical phenomenon. Technology plays a key role, but social factors (such as literacy, economic development, regulatory climate, and social norms) also influence the diffusion of the technology [3, 6]. Thus, our approach has been to try to explicitly capture the social and technical drivers in a cause-effect structure. If the behavior of this structure is validated, it would enhance our understanding of the mechanics of Internet growth. It would then be possible to test what-if scenarios of growth, making the model useful for planning and policy setting.

Selecting from among the various approaches for representing processes, we picked the systems dynamics (SD) methodology [1] for several reasons: it has helped model processes in such diverse areas as environmental policy, technology management, and organizational change; it is well-suited to synthesizing individual cause-effect relationships into an overall causal structure; and its models can be simulated to study what-if scenarios. The SD premise is that system structure causes system behavior. The major structural building block is the loop; behavior results from interaction among a system’s feedback loops. Developing an SD model involves identifying the system’s feedback loop structure, then validating it by comparing simulated with observed behavior.

Back to Top

Structural Mechanics of Internet Growth

Figure 1 uses standard SD symbols to show the structural components and causal model of Internet growth. A positive (negative) link polarity indicates that, other variables remaining constant, an increase in the cause results in an increase in the effect. A double bar on a link indicates a delayed effect. A sequence of links ending back at the originating cause gives rise to a causal loop. An even number of negative links in a loop results in a positive (negative) feedback loop. A positive (negative) feedback loop reinforces change in any variable in the loop.

Of the four major structural components outlined in Figure 1, “contagion effects” captures the contagion mechanism used in prior models; growth in the number of adopters is driven by innovator and imitator fractions that change over time due to feedback. The socioeconomic-factors component represents one aspect of feedback. A correlation between economic development and telecommunications infrastructure has been documented in the literature [6, 10]. Similarly, measures of social development are known to have an effect on diffusion [10]. We use the human development index [5], combining such social development measures as an external factor moderating contagion. The double bars on the links emanating from “economic activity level” reflect the fact that socioeconomic processes are usually slow. The level of competition in the telecommunications sector in any country, including the U.S. and India, is also known to be a significant driver of diffusion [10], as it affects price and quality of services.

We represent technology using three aggregate parameters in view of the model’s broad scope. Starting with simple text-based email and file transfers, Internet applications have progressed from static Web pages, to dynamic Web pages, to Web-enabled database access, to streaming audio. Application variety is clearly an important driver of Internet diffusion. However, application diversity also poses a challenge for interoperability. Although standards improve interoperability, they represent a moving target since application variety is constantly increasing. In any case, interoperability influences diffusion, with lower levels resulting in slower adoption. Perceptions about potential security threats also affect diffusion. While encryption algorithms and firewalls reduce such risk, increasing numbers of users and application variety increase the security risk. Figure 1 reflects the feedback effects of application variety, interoperability, and security risk on contagion through the component labeled “technical factors.”

The third component in Figure 1 is the Internet infrastructure itself and its price-performance characteristics as experienced by users. Lower access prices spur adoption, but the level of competition allowed by regulation and the size of the market as represented by actual adopters are important determinants of the number of providers and, hence, price. The number of actual adopters determines traffic load, which, in conjunction with infrastructure capacity, determines infrastructure-utilization levels—an important determinant of network performance. Congested networks lead to user dissatisfaction, slowing the adoption process.

The structure in Figure 1 explicitly integrates contagion mechanics and major external drivers into an integrated feedback model of Internet growth through the formalisms of SD. It was implemented using the SD stock-flow constructs [1]; the resulting system of difference equations was simulated with empirical data to calibrate the model.

Back to Top

Feedback Model Calibration and Predictive Validity

The variable labeled “actual adopters” in Figure 1 is measured using proxies (such as number of hosts and subscribers) that are publicly reported with some regularity. We applied the feedback model separately to Internet growth data from the U.S. and from India, two countries with vastly different socioeconomic conditions and telecommunication infrastructures. We wanted to see whether this integrated feedback model would yield a closer approximation of observed growth compared to contagion effects alone and whether it is robust enough to apply to markedly different socioeconomic contexts. Confirmation on both counts would suggest that the feedback mechanics identified here do indeed underlie the diffusion process and may apply to a range of countries, even those as different as the U.S. and India.

We obtained data about the number of U.S. hosts from the Internet Engineering Task Force (see ftp. We augmented the data used in [8] with additional data from the IETF covering 1997 to January 2000; the model was calibrated using data from 1981 to 1995. Figure 2 plots actual vs. predicted values; a correlation greater than 0.96 between the two sets of values indicates good model fit.

Back to Top

Significance of Sociotechnical Drivers

To test predictive validity, we simulated the calibrated model to predict the number of hosts in the U.S. from January 1996 to January 2000; the table compares predicted and actual values. We assessed prediction quality using the mean absolute difference between actual and predicted values, averaged over the data points. The table shows that prediction quality improved significantly when contagion models were augmented with external factors; for example, the two contagion models reported in [8]—Logistic and Gompertz—had mean absolute percent errors (MAPEs) of 47% and 35% respectively for the prediction period 1995–1997. Even the Exponential model, which did not assume contagion effects but had the best predictive validity, had a MAPE of 14%. By comparison, our feedback loop model had a MAPE of 4.4% during 1996–2000.

Improved prediction aside, it is interesting to note the external effects that resulted in the best least squares fit, since they provide insight into the growth process beyond just contagion effects. The two external factors that most improved fit for U.S. data were “application variety” and “perceptions of security threat level.” We found that increasing application variety in late 1992 improved fit substantially beyond what would be obtained through contagion alone; 1992 coincides with some significant Internet-related events. The Internet was opened to commercial use in 1988, and the Web was born around 1991. With the release of easy-to-use browsers around 1993, the Web and related applications began to proliferate. Our model reflects such significant technical advances and confirms their effect on the mechanics of Internet growth. It is clear from the literature that although no single event can be viewed as triggering concern over security, security has become an increasingly common concern. We found that further improvement in fit resulted from the increasing security threat level in late 1994. Although the timing of this increase cannot be tied to any single event, it is roughly consistent with the period in the mid-1990s when security concerns increased significantly.

Back to Top

A Different Growth Environment

We next applied the feedback model to India. The contagion mechanism was the same as in the U.S., but external drivers were vastly different. Physical infrastructure levels are much lower; there is less competition among service providers due to regulatory policy, and the human development index, reflecting the level of social development, is much lower.

Internet service in India began in 1994 with Videsh Sanchar Nigam Ltd., the government-controlled international telecom service provider, as the sole service provider. The access price was high (as much as 2% of a typical worker’s monthly salary) and service poor (including transmission cut-offs). Early penetration was almost nonexistent; only 4,000 subscribers had signed up by 1996. Then growth began to take off; for example, during 1997, it was exponential, reaching an annual rate of approximately 60%. During 1997–1998, growth slowed significantly. In 1998, the Indian government deregulated Internet service. By 1999, the number of Internet service providers had increased, and competition resulted in improved service levels and falling access costs. These factors caused rapid growth (50%) in Internet subscriptions after mid-1999 [2].

Given that Internet service was introduced in 1994, the feedback model for India had to be calibrated using quarterly data during 1996–1998 (only 12 data points); the findings should therefore be viewed as preliminary. The proxy for actual adopters here is the number of subscribers; Figure 3 shows the plot of actual vs. simulated values of number of subscribers for 1996–1998. Therefore, despite fewer data points, we conclude that the model fits Indian data reasonably well. We used the calibrated model to predict the number of subscribers for the next eight quarters (1999–2001). The table compares the model’s predicted numbers with actual values.

Although prediction errors were not as good as in the U.S. case (6.92% vs. 4.4%), they were still much better than those obtained with the best Exponential model for the U.S. (14%). The parameter changes that needed to be made in order to fit the model to Indian data corroborate what is known about the Indian context; the most significant items include:

  • Internet diffusion. The effect of Internet diffusion on economic development had to be weaker than in the U.S. model. We attribute this difference to the Indian economy’s difficulty absorbing the benefits of Internet growth as rapidly or as extensively as in the U.S.
  • Application variety. There was less application variety than in the U.S. model, because many applications, including e-banking, travel reservations, and telework, were not yet popular in India.
  • Price elasticity. The price elasticity of demand was greater than in the U.S. model. Internet charges represent a substantial expense by Indian economic standards, resulting in greater price sensitivity.
  • Diffusion parameters. Surprisingly, diffusion parameters had to be made more insensitive to infrastructure performance compared to the U.S. model. We surmise that other external drivers are more dominant, drowning out the effects of performance at this time.
  • Deregulation mechanism. Compared to the U.S. model, the effect of the telecom deregulation mechanism in the India model on competition was much more gradual. We attribute this to the inexperience of India’s telecom regulatory organization.

Back to Top


Internet diffusion is of interest to many constituencies. In view of the good fit of our model in two different settings—the U.S. and India—we can make several observations about the mechanics of growth based on the major feedback loops that should be of interest to each of them, especially policymakers. The basic positive feedback loop in Figure 1 is the contagion mechanism, whereby innovators start the process gradually and imitators then accelerate growth. An interesting aspect of Figure 1 is how interacting feedback loops stemming from the external factors moderate the contagion process; for example, a negative feedback loop connects access price to the number of actual adopters, that is, higher prices choke off growth, especially in developing economies, as reflected in the Indian model. The positive feedback loop connecting infrastructure capacity and actual adopters indicates that Internet growth is inhibited if capacity does not keep pace with demand. Both these infrastructure parameters—price and capacity—are strongly affected by the level of competition, which in turn is governed by telecom regulation.

Most developing countries have been deregulating their telecommunications sectors since at least the 1980s, but the pace of deregulation might need to be accelerated to stimulate Internet growth; for example, the Telecom Regulatory Authority of India, the Indian counterpart of the U.S. Federal Communications Commission, has been slow in recommending ways to move to more flat-pricing schemes, and the international gateway market—previously a VSNL monopoly—was opened to competition only in 2000.

On the socioeconomic side, note the positive feedback loop connecting the human development index and actual adopters via the variables “innovator fraction” and “information value.” This feedback loop indicates that lower levels of development will continue to inhibit Internet growth. As long as other policies (such as those dealing with education, transportation, and energy) inhibit development, the economy will not be able to absorb the benefits of the Internet, thus stifling Internet growth. Though not a developing country, Singapore is a relevant example in this context; in the early 1980s it decided to emphasize technology education in its schools, recognizing that people represent the country’s major resource and that it needs a technically qualified work force to take advantage of the worldwide digital revolution.

The diffusion of the Internet is indeed a sociotechnical phenomenon. The balance between technology drivers and a society’s ability to absorb the benefits of technology represents the control valve pacing that growth.

Back to Top

Back to Top

Back to Top


F1 Figure 1. Feedback loop structure of Internet growth.

F2 Figure 2. Fitted growth curve based on 1981–1995 U.S. data.

F3 Figure 3. Fitted growth curve based on 1995–1998 India data.

Back to Top


UT1 Table. Predicted and actual adoption.

Back to top

    1. Coyle, R. System Dynamics Modeling: A Practical Approach. Chapman & Hall, London, UK, 1996.

    2. Dataquest. Internet: Action time ahead. Dataquest Indian Ed. (July 15, 2000);

    3. Elie, M. The Internet and global development. In Proceedings of INET98 (Geneva, Switzerland, July 21–24). Internet Society, Reston, VA, 1998; see

    4. Gurbaxani, V. Diffusion of computing networks: The case of Bitnet. Commun. ACM 33, 12 (Dec. 1990), 65–75.

    5. Ivanova, I., Arcelus, F., and Srinivasan, G. An assessment of the measurement properties of the human development index. Soc. Indicat. Res. 46, 2, (1999), 157–179.

    6. Kraemer, K., Gurbaxani, V., and King, J. Economic development, government policy, and the diffusion of computing in Asia-Pacific countries. Public Admin. Rev. 52, 2 (1992) 146–157.

    7. Press, L. Tracking the global diffusion of the Internet. Commun. ACM. 40, 11 (Nov. 1997), 11–18.

    8. Rai, A., Ravichandran, T., and Samaddar, S. How to anticipate the Internet's global diffusion. Commun. ACM 41, 10 (Oct. 1998), 97–104.

    9. Rogers, E. Diffusion of Innovations. Free Press, New York, 1995.

    10. Saunders, R., Warford J., and Wellenius B. Telecommunications and Economic Development. The Johns Hopkins University Press, Baltimore, MD, 1994.

    11. Wolcott, P., Press, L., McHenry, W., Goodman, S., and Foster, W. A framework for assessing the global diffusion of the Internet. J. AIS (Nov. 2001).

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More