Expectations, network effects and timing of technology adoption: some empirical evidence from a sample of SMEs in Italy

Several surveys of equilibrium diffusion models exist in the literature (Stoneman 1983; Dosi 1991). For this short review, we follow Stoneman (2002).

Empirical research on network effects has been growing over the last decade. The table highlights only a selection of empirical contributions and the indicators that have been used as proxies for the network effect. It does not aim at being exhaustive.

Additional information reveals that client server applications (81%), intranet and extranet developments (30% and 32% respectively), and emails (22%) were the most important drivers of firms’ decision to adopt new LAN standards and equipment.

We are aware of the importance of investigating the overall attitude of firms towards technological innovations and the strategy they follow in their adoption choices with respect to LAN standards, when analysing the process of adoption of new technologies. A thorough analysis of these issues goes beyond the scope of this paper, but constitutes the object of a research we are currently carrying out (see Corrocher and Fontana 2006).

We are specifically interested in the timing of adoption of Fast Ethernet, and not in the process of adoption of high speed standards in general. Therefore, we consider adopters of FDDI and ATM as non adopters.

This classification is clearly inspired by Rogers (2003), who identifies five ‘ideal types’ of adopters (Innovators, Early adopters, Early majority, Late majority and Laggards). However, while Roger’s distinction is done on the basis of sociological and behavioural attitudes toward innovation, our distinction is not driven by the same concerns.

We thank an anonymous referee for pointing out that this estimation strategy was more appropriate given the features of our data.

See Stewart (1983) for a description of the method and Tedds (2006) for an application.

This result holds when using alternative proxies for size such as the number of employees, the number of connected network nodes, and the number of connected company sites. Bigger firms have larger networks in place and are more likely to experience congestion problems. Upgrading to Fast Ethernet is a way of reducing congestion.

Swcost is excluded to ensure some variability between the two steps of the model, in order to reduce simultaneity problems leading to possible spurious significance of sample selection effects (Battisti and Stoneman 2005) and because switching costs are traditionally considered to influence the probability to adopt rather than the timing of adoption. The same applies for Size, though it is included in the regression. Including this variable provides us with a simple test of its validity as instrument by checking whether it is significant in the selection equation but not in the equation of the second stage. Thanks to an anonymous referee for this suggestion.

The estimates reported are computed by the two stages approach in which the mills ratio from the probit regression is used in the second stage. Standard errors are corrected via bootstrapping. In particular we resample (with replacement) from the entire sample and repeat the procedure 200 times.

At the bottom of Table 5 we report the McKelvey & Zavoina R² as a measure of goodness of fit and a sigma statistics. The R² is constructed as: \({Var\left({\hat{{y}}-y^{\ast}}\right)}/{Var\left({y^{\ast}} \right)}\) using the variance of the latent variable y ^* and the variance of the latent predicted variable \(\hat{{y}}\) . The sigma statistics is the equivalent of the standard error of estimate in a OLS regression.

It can be argued that this variable just captures the impact of budget constraints on adoption choices. This is not plausible in our case. If some respondents wait for the price to decline because they cannot afford the product at a high price, we should expect them to be particularly sensitive to the economic variables influencing adoption. If the variable PRICEEXP captures this sensitiveness instead of the impact of expectations, it should be positively and significantly correlated to other ‘budget related’ variables. Among these variables, cost of adoption is the most relevant. In the questionnaire, respondents’ evaluation of adoption costs is captured by two different questions. One question asks to rank on a four point scale how important high costs of adoption are as a barrier for the decision to upgrade their existing technology. Answers to this question are coded into the variable (COST_BAR). Another question asks how important minimising the capital cost of upgrading is as an objective when deciding to invest in networking. In this case, the corresponding variable is (UP_COST_OB). It can be noted that PRICEEXP is negatively and significantly (albeit weakly) correlated with COST_BAR (−0.1632), and positively although not significantly correlated with UP_COST_OB (0.0034). This seems to contradict the hypothesis that PRICEEXP captures the impact of budget constraints on adoption choice.

In carrying out this estimation we follow Demoussis and Giannakopulos (2006).

This implementation deviates from the two-step standard Heckman procedure in the sense that the effects of sample selection do not follow the traditional (simple linear) approach. We have not adjusted the standard errors for the estimated Multinominal Logistic coefficients. Making this type of adjustment requires quite a complex procedure (see van de Ven and van Praag (1981) for a description and Gooroochurn and Hansley (2006) for an application) that goes beyond the scope of our sensitivity analysis, which aims instead at providing a preliminary exploration of the effect of selection on timing of adoption.