3. Methodology

View the most recent version.

3.1 Conceptual framework

There has been an increasing emphasis on the role of the agglomeration of economic activities. Over history, we have witnessed that industries tend to geographically concentrate in a few regions, and this pattern of concentration has not ceased to increase, even in an era of low transportation and communication costs. The degrees of regional concentration ofindustries are simply too great to be explained by historical accident or random process (Ellison and Glaeser 1997). The high concentrations of economic activities are thus believed to be driven by the advantages that regional-agglomeration economies have to offer. The three most widely acknowledged advantages of regional agglomeration are knowledge spillovers; specialized, skilled labour; and input sharing (Marshall 1920). While the effects of specialized, skilled labour and input sharing can be relatively easily measured with accuracy, estimating the effects of knowledge spillovers has resisted econometric scrutiny, mainly because of its unobservable nature in most cases.

There are two types of knowledge: one is explicit or codified knowledge that can be effectively expressed using symbolic forms of representation; the other is tacit knowledge that defies such representation (Reber 1995). The more easily explicit knowledge can be accessed, the more critical a role does tacit knowledge play in sustaining and enhancing the competitive position of the firm (Maskell and Malmberg 1999). Consequently, tacit knowledge now plays a greater role than ever at a time when explicit knowledge has become easier to obtain.

When implementing new technologies, plants face many kinds of uncertainties associated with the costs and benefits of technologies, adaptation difficulties and employee training. More information on these would not only reduce the uncertainties associated with adopting new technologies but would also enable plants to better assess risks and expectations. However, because some types of information associated with technology implementations are tacit—for example, detailed engineering characteristics or particular organizational changes to fully exploit technology capabilities—learning this type of knowledge depends on direct observation of early adopters, demonstration, word-of-mouth and other informal mechanisms. Therefore, the local presence of prior adopters may facilitate interplant-knowledge spillovers in the region. Furthermore, the feedback loop of knowledge spillovers from technology adopters facilitates, and is facilitated by, regional agglomeration (Case 1992; Jaffe, Trajtenberg and Henderson 1993; Powell and Brantley 1992; and von Hippel 1988).

Therefore, we can think of the local presence of prior adopters of technology as a measure of a source of information that is difficult to obtain from a distance. Consequently, we will refer to the positive impact resulting from the local presence of prior adopters of technology as knowledge spillovers from prior adopters to potential adopters.

3.2 Empirical framework

Can data on the pattern of technology adoption reveal whether a plant's adoption decision is affected by the presence of prior adopters of technology around it? To explain the methodology for estimating the impact of the local presence of prior adopters of technology, we first begin by modelling the probability of a plant's technology-adoption decision. Let us suppose that the true model governing the technology adoption of a plant is given by

where p indexes plant, i indexes industry, r indexes region, τ indexes technology and t indexes time. Adoption _{pτ irt}   is a binary variable indicating whether a plant p in industry i in region r adopts technology τ at time t .

Because knowledge spillover is unobservable, and thus needs to be inferred, the central task in estimating knowledge spillover lies in how accurately we can specify the channel of knowledge spillovers and how finely we can control for exogenous effects that influence a plant's probability of technology adoption. By estimating the impact of the local presence of technology adopters on other plants' probability of technology adoption, this paper attempts to identify the knowledge spillovers from prior adopters to potential adopters. The preliminary results show that the presence of prior adopters positively affects technology adoption by other plants. Detailed results are discussed in Chapter 5.

Can this positive effect of prior adopters on potential adopters be interpreted as evidence of some kind of knowledge spillover or information sharing between them? Ideally, if we know all the exogenous influences that affect a potential adopter's decision, then we would be able to infer the positive effect of prior adopters as capturing their own effects only. However, there are potential alternative hypotheses that the presence of prior adopters is positively correlated with the technology-adoption decision of potential adopters, even if there are no knowledge spillovers of any kind. We will now look at these alternative hypotheses.

The first alternative hypothesis is that the results are driven by unobserved local-area characteristics that are correlated with both the presence of technology adopters and the decision of potential adopters. Specifically, regions where there are agglomerations of economic activities would provide advantages that are favourable not only to do business but also to adopt new technologies. These things include such obvious factors as the presence of abundant skilled labour (i.e., scientists and engineers who would make the adoption and implementation of technology easier), the local presence of input suppliers and output consumers, the presence of universities or research institutions, tax policies and a good infrastructure. Furthermore, plants in the same region face the same exogenous local influences such as a local research-and- development (R&D) subsidy, tax incentives, or a business cycle that would affect any technology adoption. These location-specific characteristics would idiosyncratically affect the technology-adoption decision of all plants in that region, and some of these effects are positively correlated with the existing number of technology-user plants in the area. Therefore, distinguishing the effects of location-specific characteristics from the effects of the presence of prior adopters is imperative.

The second hypothesis is that the results are driven by unobserved effects operating at various levels that are correlated with both the presence of technology adopters and the decision of potential adopters in a region. Such unobserved effects may operate at the industry level, the technology level, or even at the interaction of industry × region, industry × technology or region × technology levels. For example, the adoption rate of any advanced manufacturing technologies in the Aircraft and aircraft parts industry, Standard Industrial Classification (SIC) 321, is 28% while it is only 4% in the Rubber hose and belting industry, SIC 152. Because of these kinds of industry-level fixed effects, a regional concentration of technology-intensive industries would be positively correlated with the number of technology users in the region. Another example of such unobserved effects would be a cost reduction in the adoption of, say, computer-aided design and engineering (CAD/CAE) that would increase the adoption rate of that technology. Similarly, a local tax incentive or R&D subsidy to a particular industry would increase the overall technology adoption for that industry within the region. Since there are many potential factors that may influence technology adoption at the plant level, it is essential to control for these unobserved effects that operate at various levels.

The third hypothesis is that results are driven by 'omitted' plant characteristics that influence the technology adoption decision. A theory on the differential capacity of firms to absorb, and make good use of, new technical information emphasizes differences in internal expertise, access to financial resources and organizational routines. These differences affect each firm's expected profitability—the incremental returns to investing in the new technology—that, in turn, gives rise to the observed uneven pattern of adoption (Cohen and Levinthal 1990, Dosi 1988, Malerba 1992, Nelson and Winter 1982). Also, plant learning—and ability to act on that information— will also vary by the level of organizational resources, scale of the production process, appropriateness of the new technology to that plant's core-production process and sources of information, all of which may have nothing to do with geography itself. For example, large plants or multi-product plants may well be run by more innovative entrepreneurs who tend to adopt more technologies. Therefore, there may be a plant-specific component to the error term in the specification that is correlated with included right-hand side variables. In addition, there is a possibility that some of the variables are potentially endogenous—for example, a plant introduces a new technology and then decides to locate to a region, or it relocates to a region for the purpose of adopting technologies.

In order to separately identify and estimate the effect of prior adopters on potential adopters from the above-mentioned unobserved effects and thus eliminate them as potential explanations of the results, the following methods are employed. First, the concern that omitted location variables may drive the results is addressed as follows. If the results hold up after inclusion of location- fixed effects, they cannot be driven by any effects that are common at the regional level, such as the presence of universities, location advantage, transportation, tax policies or regional influences. Therefore, we include location-fixed effects at the economic region level—at which both the dependent variable and the key variables are measured. ⁷ Furthermore, to make sure that the results are not capturing the agglomeration effects—such as the local presence of specialized, skilled labour, input suppliers, output consumers and the overall size of regional manufacturing activities at a finer geographical level than the economic-region level—variables capturing these effects at the census-division level are included. Consequently, these will make sure that the results are not driven by the agglomeration effects operating at the census-division level as well as any unobserved effects at the economic-region level.

Second, the concern about unobserved effects operating at various other levels is dealt with in the following ways. Industry-, technology- and time-fixed effects are included to control for the effects that are common to the industry, the technology and time. In addition, two variables measuring the average adoption rate of overall technologies by industry × region and the average adoption rate in a region by particular technology × industry are included in the specifications to further control for effects that operate at the industry × region and technology × region levels. ⁸

Third, the issue about the unobserved plant heterogeneity is handled by including an extensive set of plant characteristics, such as size, plant status, the number of commodities, ownership and age. Although plant-level heterogeneity would most ideally be controlled by plant-fixed effects, the small variation across the adoption pattern of 22 technologies within a plant does not allow the inclusion of plant-fixed effects. Therefore, variables capturing important plant-level heterogeneity that affects adoption decisions are used instead. The set of plant characteristics included here is the richest plant-level information used in the literature, in the author's opinion.

In addition to the extensive controls and fixed effects mentioned above, the effects of prior adopters are estimated separately, based on the functional, geographical and technological distance from the potential adopter. These separate estimations not only reveal how the effects are bound by the three dimensions, but also serve as a test to show that the results are not driven by any of the alternative hypotheses mentioned above. The reasons for this are as follows. First, the estimation of the effects of prior adopters of the same region, separately by the functional distance from the potential adopter, allows us to determine whether the results are driven by region × technology-fixed effects or if the effects are function/industry specific within each region × technology level. Second, the estimation of the effects of prior adopters of the same technology, separately by the geographic distance, allows us to discern whether the effects are driven by industry × technology-fixed effects or if they are geography specific within each industry × technology level. Finally, the estimation of the effects of prior adopters of the same region, by the technological distance from the potential adopter, allows us to analyse whether the effects are driven by the industry × region-fixed effects or if they are technology specific within each industry × region level. These separate estimations of the effects of prior adopters on potential adopters by the functional, geographical and technological distances confirm that the effects of prior adopters are not driven by any of the potential alternative effects mentioned above, but are very likely capturing the effects of the presence of prior adopters. The only remaining possibility of the spurious result is that the results are driven by fixed effects that operate at the level of region × industry × technology × time. Not only is it very unlikely to come up with fixed effects that operate at this detailed level, but the fact that the results are strongest when prior adopters are in a similar-but-not-the-same industry as the potential adopter provides very compelling evidence of the validity of the results obtained here.

The estimating equation for plant p 's adoption of technology τ at time t hence is

where F represents the logistic cumulative distribution. Logit model is used to capture the 'fat tail' of the distribution (i.e., there is a larger proportion of non-adopters of any technology at time t ). X_pirt is a vector of plant characteristics, Avg _ Ind _ Regioni _Rt is an average adoption rate of advanced technologies overall among plants in industry i in economic region R at time t , and Avg _ Ind _ Tech _{iτ t}   is an average adoption rate of technology τ in industry i across economic regions at time t . δ R is the location fixed effect, γ i is the industry fixed effect, Õτ is the technology fixed effect, and λ t is the time fixed effect. The variables in the estimating equation are explained in detail in the next chapter and Table 3.

7 . More discussions of the unit of geography are in Section 4.

8 . The inclusion of fixed effects at the interaction levels would completely capture all the effects. However, due to the variability of the sample, it is not allowed in this specification.