Proximity strategies in outsourcing relations: The role of geographical, cultural and relational proximity in the European automotive industry

Abstract

Several trends that affect the manufacturing of sophisticated goods – increasing international fragmentation of production, and lean and modular process technologies – have increased the importance of proximity in the supply chain. We use the case of the European automotive industry to simultaneously evaluate the relative importance of three dimensions: geographical, cultural, and relational proximity. Using a rich and novel data set, we find that carmakers value some aspects of each dimension independently in their sourcing strategy. The estimates indicate which proximity measures provide the largest (independent) benefits, but also that the positive effects the literature has attributed to some measures tend to reflect past relationships rather than predict new ones. In particular, co-location and a low cultural distance should be interpreted as outcomes of a sourcing strategy, not as predictors for sourcing success. Finally, we investigate to what extent firms from different countries follow different strategies, and which choices suppliers can make to boost their attractiveness as outsourcing partners.

APPENDIX

Robustness Analysis

Measurements of geographic distance

We have conducted a series of robustness checks, and report the relevant results in Table A1. For all different specifications, a linear distance variable was included throughout. This tends to reduce the point estimates on the PROX700KM and NO BORDER variables, but leaves all qualitative findings unchanged – see for example Specification 1.

Table A1 Robustness analysis

We have also performed the regressions including an additional discrete distance effect, at 100 km, which allows for multiple deliveries per day, or defined the cutoffs based on frequency distributions of distance rather than absolute distances. In all cases, the benefits of geographical proximity gradually decay as we move away from the assembly plant, at least if we do not control for relational proximity.

Alternative cluster variables

Next, we consider alternative measures of supplier clusters. In Specification 2 we require a minimum of five (instead of three) plants to qualify as a cluster; in Specification 3 we enlarge the radius around supplier plants that defines a cluster to 30 km (from 10 km). Both changes raise the number of firms in an average cluster, but in Specification 2 the total number of clusters is reduced, whereas it is raised in Specification 3. In both cases, the elasticities associated with any type of cluster increase, but the changes are very small. Using a more narrow definition of clusters leads to slightly lower elasticity estimates.

Definitions of cultural distance

We have also estimated the benchmark model using an alternative definition of bilateral headquarters’ distance, grounded on the psychic distance approach. We used the variable constructed by Håkanson & Ambos (2010: 201), defined as the “sum of factors (cultural or language differences, geographical distance etc.) that affect the flow and interpretation of information to and from a foreign country”. For better comparison, the variable is standardized relative to the sample mean. In Specification 4 we see that this alternative variable produces similar results, but it captures less variation in our sample. Interestingly, it seems to capture more of the cultural aspects embedded in the border effect, but it performs worse at measuring very close cultures, as in the case of shared nationality.

Correlation between contracts

The assumption that the residuals are uncorrelated might be violated if some important connection between observations is not modeled. The tests in Table 7 investigate this for two possible types of dependency, namely spatial correlation and historical linkages. Another potential violation of the independence assumption might occur if the supplier choice for one component depends on outcomes of contracts for other components of the same vehicle. We can accommodate such effects by using a clustered variance–covariance matrix that allows the covariance terms to differ from 0 within each set of contracts for the same car model. The estimates in Specification 5 of Table A1 have indeed larger errors, but this does not affect the interpretation of the results.

Endogeneity of location

One potential endogeneity problem is that suppliers could change the location of their plants to influence the outcome of contracting decisions. It would invalidate our treatment of the observed locations as predetermined supplier characteristics. Our analysis already mitigates this issue in two ways. First, by including contract-level fixed effects, we compare across suppliers only for a given model–component pair, which holds constant anything unusual about the contract, such as a remote assembly plant location, or a component with strong co-location requirements. Second, given that the sample includes for all suppliers many more contracts than production locations, they can locate their plants only near a few assembly plants.

Nonetheless, we can verify the sensitivity of the results by re-estimating the model excluding components that are more likely to suffer from this endogeneity problem. Components that are bulky, that generate high coordination costs, or that require a lot of face-to-face interactions are candidates. In Specifications 6 and 7, we re-estimate the benchmark specification, but exclude engineering-intensive or design-intensive components. Taking an extreme position, we eliminate half of the observations in the first regression and almost 40% in the second. The results change only slightly, except for the PROX 10KM variable. It confirms our finding from Table 7 that endogenous locations cannot be dismissed entirely as an explanation for the effect of co-location.

Construction of choice sets

From the start, we had to define unique components to construct a set of potential suppliers for each observed contract. All results so far are based on a detailed classification system that separates components according to their generic name, functionality, and area of application in the car. An alternative approach is to group components according to their generic name alone. This broader definition groups components in the same category irrespective of their function or application in the vehicle. The implication is that choice sets include more potential suppliers for the same component. Competing suppliers will be less alike, and contract-level fixed effects will absorb a smaller fraction of the variation. Results in Column 8 of Table A1 show that using generic component names does not materially change the estimates.

The IIA assumption

A well-known restriction of the conditional logit model is that it implies IIA. By construction, the relative probability of selecting one supplier instead of another is independent of the presence or characteristics of further potential suppliers. Especially when the model needs to describe the choices of a heterogeneous group of decision-makers over varying choice sets, this assumption might be overly restrictive. The inclusion of supplier fixed effects alleviates this concern somewhat.

A statistical test for the validity of the IIA assumption is readily available (Hausman & McFadden, 1984). One needs to estimate the model excluding suppliers one by one from all choice sets, and then test for significant differences in the coefficient estimates. While excluding the majority of suppliers did not pose a problem, for a few of the largest suppliers in the sample and a few non-European firms the predictions in the benchmark model changed slightly but in a statistically significant way. Our specifications with interaction terms, however, provided more robust results.

Supplier Choice Model

For each set of potential contracts r=1, 2, …, n (defined in the Methodology section), car assemblers choose which suppliers s=1, 2, …, nr they want to sign a contract with. We assume that the final choice of suppliers maximizes the profits of car assemblers. Denoting by S _r the set of options available in the supplier selection process, the optimization problem is to choose which of the potential suppliers to award the outsourcing contract and which to decline it, so that expected profits are highest, given the quantities q specified in the contract.²¹ Profits are modeled as a linear function of contract-specific characteristics α _r , which include the characteristics of car assemblers and assembly locations, a set of characteristics of suppliers and supplier locations β′x _rs , and a nuisance term ɛ _rs that captures unobserved factors plus measurement errors, assumed to be independently and identically distributed with a type I generalized extreme value distribution:

McFadden (1974) shows how the maximization of a random utility function can be linked to the conditional logit model. Woodward (1992) applies an equivalent profit maximization problem to the location choice of Japanese manufacturing start-ups in the United States. Similarly, we study the geography of the European automotive industry with a model that implies an underlying profit maximization for car assemblers choosing amongst different component suppliers. The joint probability function that corresponds to such choices is given by

where Y _i =y _i , i_{rs}=1, 2, …, N are the observed dichotomous yes-or-no choices of supplier, and m _r is the number of successful suppliers in each set of potential outsourcing contracts over which the conditioning takes place. The two outer summations in the denominator are over the sets M _r of all possible combinations of 0 and 1 to for all r=1, 2, …, n, such that Notice that the contract-specific scalar α _r , as any other constant term in the model, cancels out after conditioning, since it can be factorized away on both numerator and denominator. This implies that we are not able to retrieve its estimates. The parameter vector β is estimated by maximum likelihood estimation, as shown in Mehta and Patel (1995).

Our basic model specification includes K distinct measures of proximity and their average effects β′=(β₁, …, β _K ), k=1, 2, …, K, in addition to supplier fixed effects β_9s and country fixed effects for the locations of supplier plants β_10c:

We report the mean of elasticity for each dichotomous proximity variable. More specifically, we calculate the transformed point estimates g(β _k ), where g(·) is a transformation that retrieves the change Δy _i /y _i in the fitted outcome variable that is due to a change from 0 to 1 in a given variable x _ki .²²

Two-Step GMM Conditional Logit Estimator

We can test whether we have omitted an extensive spatial correlation structure by checking whether ρ=0 in the following spatial autoregressive AR(1) model:

In the second equation line, the inverse matrix has a full-blown MA(∞) representation that pre-multiplies both the error term and the explanatory variables:

We construct a spatial weight matrix using the inverse distance between the locations of suppliers, for each observation in the sample d _ij ⁻¹, as shown below:

If ρ is significantly different than 0, the errors are not uncorrelated as assumed, and our βs will be inconsistently estimated, owing to omitted variables. Whereas the first issue can be handled with bootstrap estimation, the problem of structural bias requires further treatment.

The variables WX are easily obtained from the sample, but WY is not available, as Y is a latent variable. Kelejian and Prucha (1998) propose replacing WY by an instrumented variable. Klier and McMillen (2008a) demonstrate how a related two-step GMM estimator can be obtained for use with large samples, in order to avoid the need to work with very large matrices (for us this means inverting a square 269,608 matrix, which exceeds our computational resources). In their method, the objective function of the standard 2 stage least squares (SLS)-GMM estimator is replaced by a one-shot guess at the orthogonality condition – a linear approximation at the point where ρ = 0.

The motivation is that the solution for the model Y=(I−ρW)⁻¹Xβ+(I−ρW)⁻¹ɛ is relatively close to the point In the full spatial model, a GMM estimator would require minimization of the condition that the regressors (I−ρW)⁻¹Xβ are orthogonal to the error term (I−ρW)⁻¹ɛ. In the linearized model, instead of using the full expression for the error term, it is approximated by applying a Taylor series expansion of first degree around :

Define Now the GMM method is to minimize the condition that a set of instruments Z is orthogonal to v, with respect to θ:

We use M=(Z′Z)⁻¹, which amounts to the two-stage least squares estimator. and can be easily obtained from the initial conditional logit regression. is, by definition, the gradient of the error term of the full model with respect to the parameters of interest, an endogenous term if not treated. Measured at the starting point it becomes

Furthermore, the residuals in the full model u≡(I−ρW)⁻¹ɛ can be retrieved by u=CHOICE−P, where CHOICE is the observed binary choice variable. Thus the above reduces to

P is the conditional logit probability, again with the assumption of one positive outcome within each stratum, written as, for each observation i:

where

and

After some algebra, the gradient terms become

where:

and

Assuming at the starting point that ρ=0, we do not have to bother about all the heteroskedastic terms, and the above gradients become much simpler, with Λ equal to 0 on the diagonal, and X _i ^**=X _i :

where

Now we have all the pieces needed to construct the variable v and run the 2SLS with instruments Z. In the first stage, we regress the endogenous gradient terms on all our exogenous variables plus a set of instrumental variables. In the second stage, the endogenous variables are replaced by the predicted values from the first stage.

Klier and McMillen (2008a) provide Monte Carlo results to show that this procedure (a logit variant) can deliver good results for parameter values ρ < 0.5. We estimate a ρ coefficient that is close to 0, positive, and statistically insignificant. The instruments we used are all variables from the standard model, some of the WX variables that vary sufficiently in space, and the plain latitude and longitude coordinates. The validity of instruments is typically tested using over-identifying restriction in the GMM condition, a procedure called a Sargan test. In the notes to Table 7 we list the p-values of this test, which does not reject the validity of our instruments.

The regression with the relational correlation variable uses essentially the same procedure, but instead of space, time is the dimension in the autoregressive matrix, called the lagging vector L now:

Because the time dimension of our sample does not exhibit sharply defined intervals, and because we want to estimate the broader effect of past collaboration, time is collapsed over all previous relationships between a certain OEM and a supplier into a single lagged period.

The estimates of the two-step GMM have a direct interpretation as marginal effects. However, the comparison with the results from the conditional logit model is not direct. The two-step GMM is an approximation method, and produces well-performing point estimates only if the sample is large enough, which is the case in our study. However, the relative magnitude of the point estimates is very much comparable, as discussed in the ‘Results’ section.