5. Data and econometric method
　The number of IC&SP carriers in Japan is too small in number for a detailed econometric study using the firm as the unit of analysis. Moreover, such an analysis is complicated by the fact that the carrier may be vertical integrated in some cities and countries and contract for transportation services in other cities and countries. Therefore, we chose the parcel as our unit of analysis, which allows us to investigate both inter-firm and intra-firm organizational heterogeneity. Also, because of difficulty in tracking parcels originating in other countries and shipped into Japan, we chose to focus on individual parcels shipped from Japan.
　We collected data on a total of 995 sample parcels shipped from 37 different origin cities in Japan to 160 destination cities in 42 countries during February and March 1998. Parcels were distributed among 14 IC&SP carriers (see Table 1). Information network tracking data was either self-reported by the carriers, recorded directly based on test parcels sent by researchers , or reported by shippers who assisted us in our research. After omitting observations with missing data we have between 815 and 903 observations depending on the dependent variable analyzed. Since the market share of IC&SP carriers was unavailable, we attempted to distribute the parcels evenly among carriers, though the number of packages per carrier ultimately varied from 49 to 200 per carrier. The variables for which we collected data are described below.
　Dependent variables: Since the customer contacts a carrier that performs the function of a freight forwarder in all cases, our analysis investigates whether the freight forwarder vertically integrates into any of the three transportation segments. DomTrk quantifies whether or not domestic trucking for picking up parcels is vertically integrated (coded 1) or contracted out (coded 0). IntAir quantifies whether or not international air service is vertically integrated (coded 1) or contracted out (coded 0). ForTrk quantifies whether or not foreign trucking is vertically integrated (coded 1) or contracted out (coded 0).
　Data for these variables was collected by interviewing each of the participating IC&SP carriers and then mapping its organizational choices for domestic trucking, international air, and foreign trucking on a parcel by parcel basis according to the parcel's origination and destination cities. All types of equity relationships were coded as vertically integrated. The absence of an equity relation was coded as contracting out.
　Independent variables: Our principal hypothesis is that integration is positively correlated with investments in the firm-specific information network for each transportation segment. Unfortunately, we were unable to obtain data on each firm's investment in its proprietary information network. As a proxy for such investments we developed a survey that collected data on the type and availability of real-time parcel information tracked by a carrier's information network for each transportation segment (the survey is displayed in the appendix). We then created an index for each transportation segment that counts the pieces of information available to the freight forwarder (which contracted with the shipper to transport the individual package) from each transportation segment. These indices, described below, assume that the level of proprietary information network investment is positively correlated with the amount of real-time information available from each transportation segment, except where noted. Thus, we assume that the amount of parcel tracking information is a proxy for the level of specific investment.
　The variable kDomTrk is an index which increases with the amount of parcel data available from domestic trucking, questions 1 through 7 on the survey. Data includes: (Q1) whether or not the package has been picked up, (Q2) name of the pick-up driver, (Q3) time of pick-up, (Q4) place of pick-up, (Q5) whether or not the package arrives at the local terminal, (Q6) time of local terminal arrival, and (Q7) name of local terminal. For each piece of information available on the information network, we add one to the index. Thus, the index ranges between 0, which indicates no data is available on the carrier's information network, and 7, which indicates all of the information is available on the carrier's information network. We expect vertical integration in domestic trucking is more likely with higher levels of kDomTrk and thus predict a positive coefficient.
　The variable kIntAir is an index which increases with the amount of parcel data available from international air carriage, questions 10 through 17 on the survey. Data includes: (Q10) whether or not the parcel was loaded onto an airplane, (Q11) the loading time, (Q12) whether or not the airplane departed, (Q13) the departure time, (Q14) the city the airplane departed from, (Q15) the cities that package visits during air transit, (Q16) whether or not the airplane landed at the destination airport, and (Q17) time of arrival. The index ranges between 0 and 8. We expect vertical integration in international air service is more likely with higher levels of kIntAir and thus predict a positive coefficient.
　The variable kForTrk is an index which increases with the amount of parcel data available from foreign trucking, questions 20 through 26 on the survey. Data includes: (Q20) whether or not the delivery truck departed from the local terminal, (Q21) the time the delivery truck departed from the local terminal, (Q22) name of the local delivery terminal, (Q23) whether or not the parcel has been delivered, (Q24) time of delivery, (Q25) location of delivery, and (Q26) name of recipient. The index ranges between 0 and 7. We expect vertical integration in foreign trucking is more likely with higher levels of kForTrk and thus predict a positive coefficient.
　Control variables: We employ different control variables depending on the dependent variable. Doc is a binary variable and coded 1 for parcel containing documents only, and coded 0 otherwise. Parcels containing documents typically are more time sensitive than small packages. Moreover, documents are lighter in weight and larger in number than other air freight. Thus, following the theoretical prediction developed by Nickerson and Silverman (1997) that time-sensitive and complex flows of freight introduce contractual hazards most efficiently organized via hierarchy, we predict a positive with respect to each of the dependent variables.
　JCities is a dummy variable that is coded 1 for parcel origin cities of Tokyo, Osaka, and Nagoya else coded 0. IC&SP Interviewees state that these cities, the largest in Japan, are responsible for much of IC&SP demand. Indeed, more than 90% of our parcel level data with originated from these cities reflects this fact. If economies of scale are present for domestic trucking and determine the organizational form, it would lead to a positive coefficient. However, the fact that the market for domestic trucking in these cities is substantial provides for a "thick" market of independent domestic trucking carriers that would lessen the need for vertical integration, which would lead to a negative coefficient.
　FinCities is a dummy variable that is coded 1 for parcel destination cities of Chicago, Hong Kong, London, Los Angeles, New York, San Francisco, and Singapore. These cities are financial centers and are destination cities for many IC&SP parcels. More than 60% our parcels were delivered to these cities. We expect economies of scale, if present for either international air or foreign trucking, would lead to positive coefficients. Alternatively, the fact that the market for international air to or foreign trucking these cities is substantial again would provide for a "thick" market of independent airlines and foreign trucking carrier that would lessen the need for vertical integration into both transportation segments, which would lead to negative coefficients.
　MarkSize is a variable that measures cumulative annual weight (in billions of kilograms) of international air freight, a freight segment including parcels larger than courier and small packages, from Japan to each destination country in the most recent year for which we have data, 1995. We employ this data as our best available proxy for country specific volume. If economies of scale in either international air or foreign trucking to a particular country are present then we would expect coefficients for this variable to be positive.
　Method: As described above, our model has three dependent variables: whether freight forwarding is vertically integrated or not into domestic trucking, international air, or foreign trucking. We employ a Probit estimation procedure for each dependent variable independently. This method assumes that error terms are uncorrelated. We model the benefit (inverse cost differential) of internal governance as unobserved variables DomTrk*, IntAir*, and ForTrk* such that:




where F( ・) denotes the normal distribution function.
　　Table 2 displays descriptive statistics for all variables. The correlation coefficients ar generally small to moderate in magnitude, which suggests that multicollinearity does not raise a problem for our estimation.




6. Results and Discussion
　Table 3 reports the Probit results for two sets of estimations: baseline models with only control variables and fully specified models with control and independent variables. First, consider the result with control variables only. Both constant and coefficient for Doc and JCities are statistically significant with respect to DomTrk. The positive sign for Doc indicates the parcels containing documents are more likely to be shipped by an integrated carrier. The negative sign for JCities supports that interpretation that the market for domestic trucking services is thick in Japan's financial centers, which reduces the need for freight forwarders to vertically integrate. With 903 usable observations, the model correctly predicts organizational choice 67.7% of the time but provides little explanatory power (R2 = 0.062).
　The constant and coefficients for Doc and FinCities are statistically significant with respect to IntAir while MarkSize is insignificant. The positive sign for Doc indicates that integration is the more likely organizational form when parcels contain documents. The negative sign for FinCities supports that interpretation that the market for international air services between Japan and foreign financial centers is sufficiently thick so as to reduce the need for freight forwarders to vertically integrate. With 902 usable observations, the model correctly predicts organizational choice 68.7% of the time but provides only modest explanatory power (R2 = 0.169).
　Finally, the constant and Doc are statistically significant with respect to ForTrk while FinCities and MarkSize are insignificant. Again, the positive sign for Doc indicates that integration is the more likely organizational form when parcels contain documents. With 822 usable observations, the model correctly predicts organizational choice 69.0% of the time but provides little explanatory power (R2 = 0.090).
　The addition of independent variables improves the models' predictive power especial for domestic and foreign trucking segments. For DomTrk, we find as predicted that the coefficient for kDomTrk is positive and significant; the greater the amount of real-time parcel information available from domestic trucking the greater the likelihood of vertical integration. Coefficients for the constant, Doc, and JCities retain the same sign as in the baseline model and remain significant. With 896 usable observations, the model correctly predicts organizational choice 72.3% of the time and provides explanatory power (R2 = 0.201)-a substantial improvement over the baseline model.
　For IntAir, we find as predicted that the coefficient for kIntAir is positive and significant; the greater the amount of real-time parcel information available from international air service the greater the likelihood of vertical integration. Coefficients for the constant, Doc, and FinCities retain the same sign as in the baseline model and remain significant. With 893 usable observations, the model correctly predicts organizational choice 64.0% of the time and provides only modest explanatory power (R2 = 0.147). Although kIntAir is statistically significant, the loss of a degree of freedom and a reduction in the number of data lead to no improvement in the model's explanatory power compared to the baseline model.
　For ForTrk, we find as predicted that the coefficient for kForTrk is positive and significant; the greater the amount of real-time parcel information available from foreign trucking the greater the likelihood of vertical integration. Coefficients for the constant, Doc, FinCities, and MarkSize retain the same sign and level of significance as in the baseline model. With 815 usable observations, the model correctly predicts organizational choice 69.4% of the time-a minor improvement over the baseline model-and provides explanatory power (R2 = 0.157)-a substantial improvement over the baseline model.
　The empirical results provide support for our theory. The greater the amount of real-time information available in each transportation segment-our proxy for the level of specific investment in a proprietary information network-the greater the likelihood that a freight forwarder integrates into that segment. Indeed, specific investment in an information network appears to be critically important for the organizational choice of domestic and foreign trucking services and at least somewhat important for international air services. Integration reduces the cost of hold-up and maladapation hazards that may emerge when specific investments are made in a dynamic and changing environment.
　This finding is important given the fact that we have controlled for the effect of temporal specificity (Pirrong 1993, and Nickerson and Silverman 1997) found in other TCE studies of logistics. Following Nickerson and Silverman (1997), we expect that the choice of logistics technology-hub-and-spoke v. point-to-point-introduces additional contracting hazards. In addition, documents, a highly time-sensitive freight which also requires the most complex coordination of pick-up and delivery operations, are found to correspond a greater likelihood of integration in each segment, as is shown by the Doc variable. This finding is consistent with Nickerson and Silverman's (1997) conclusion that integration is more likely for hub and spoke logistics technology when frieght is highly time-sensitive and on-time delivery is important.
　Other control variables, while not perfect measures, do have currency. The control variable of total market size of air freight suggests that economies of scale (and possibly scope) are not available in the transportation segments-none of the empirical findings suggest vertical integration hinges on such economies. On the contrary, thick markets, at least in some instances, lead to contracting out transportation segments, ceteris paribus. Thus, high-volume geographic markets by themselves either have no effect on the integration decision or increases the likelihood of contracting out.
　We note that our analysis is limited in a number of ways. A binary integration variable for each transportation segment was constructed even though more than two organizational alternatives are observed. In our paper, integration includes ownership, partial equity positions, and equity joint ventures. Additional variation in organizational choice might be explained if each organizational alternative was treated separately. Importantly, our analysis does not encompass any performance measures. Organizational choices are likely to have performance implications such as lower governance cost and/or higher service (i.e., shorter or more reliably delivery). In future research we plan to investigate the performance implications of organizational choice for IC&SP service. We also note that our analysis, which controls for financial centers, does not control for nation-level institutional factors for destination countries. Williamson (1991) has argued that such factors may act as shift parameters that influence the choice of organization. In future research we plan to incorporate variation in country-specific institutional factors in our analysis to control for such influences. Finally, the econometric method employed assumes errors between dependent variables are uncorrelated, which may not be an accurate assumption. In future research we plan to explore the use of alternative econometric methods to correct for correlated errors (Chib and Greenberg 1996).




7. Conclusion
　The information network assets for real-time data acquisition in the transportation segments of IC&SP service and the transmission of parcel information to the originating freight forwarder require specific investments. Our study supports the proposition that the greater the specific investment in an information network in each transportation segment, as proxied by the amount of real-time information available from each segment, the greater the likelihood of vertical integration in each segment. This result suggests that the efficiencies obtained by multinational and multi-transportation segment integration arise not from scale economies, but from the governance structure chosen in response to the nature of investments in an information network. The analyses also show that integration in each transportation segment is more likely for parcels containing documents because they are lighter in weight, larger in number, and typically most time sensitive than small packages and other air freight.
　Our paper extends TCE analysis of the logistics industry in several ways. First, the organization of international courier and small package service and investment in complementary information networks have received little academic attention. Our paper not only is one of the first to investigate this topic but also provides an empirical analysis of a contractual hazard not previously studied by TCE research of logistics-investment in information networks. Second, unlike most TCE studies that analyze a single transaction, this paper analyzed a constellation of transactions involved the value chain of parcel delivery. Finally, no prior TCE study has analyzed freight-level data. This unit of analysis allows us to investigate why one firm may choose different organizational forms depending on the transportation routes.