What matters to the MNEs’ location choice in a host country?

: A random coefficients discrete choice model for the LSA-FSA framework

March 9, 2006

In Hyeock (Ian) Lee

Business Economics & Public Policy

KelleySchool of Business

IndianaUniversity

  1. Motivations
  • Rugman and Verbeke (1992): The LSA-FSA framework

- MNEs’ location choice is the result from the interaction between a firm’s FSAs and the type of LSAs it faces, and the LSAs are different for each firm.

  • Dunning (1998): Key LSAs of four different types of FDIs

- Natural resource seeking

- Market seeking

- Efficiency seeking

- Strategic asset seeking

  • Firm-specific characteristics for inward FDI firms

- Industry

- Nationality

- Foreign ownership

- R&D expenditure

- R&D intensity

  • Research Questions

(1) What types of LSA matter to the MNEs’ location choice in a host country?

(2) What types of FSA matter to the MNEs’ location choice in a host country?

(3) What types of interactions between LSA and FSA matter to the MNEs’ location choice in a host country?

  • Pre-conditions for the empirical estimation of LSA-FSA framework

- Data on location-specific advantages

- Data on firm-specific advantages

- Development of relevant econometrics techniques

BLP (1995) & Nevo (2000): A random coefficients discrete choice model

  1. Conceptual Framework
  • The main objective of BLP’s (1995) approach, well-organized by Nevo (2000) later, is to estimate demand elasticity, how quantity demanded responds to price, based on an individual’s utility maximization.
  • In this paper, Nevo’s approach will be applied to the issue of estimating a demand function for differentiated locations by FDIs in a host country based on the assumption of a firm’s profit maximization.

Firm’s location choice = f {observed characteristics of a region (Xj),

unobserved characteristics of a region (ξj),

observed characteristics of an individual firm (Di),

unobserved characteristics of an individual firm (vi) }

+ error term (εij)

  • We have the number of FDI firms for 5 years on 234 different regions in Korea, for a total of 1,170 data observations. We also have data on 4 characteristics of each region, and 5 characteristics of the firms in each year.

, where Xj1t : the average wage in region j in year t,

Xj2t : the gross regional product in region j in yeart,

Xj3t : the existence of industrial districts in region j in yeart,

Xj4t : the number of patents in region j in yeart,

< Issue 1> The Heterogeneity of Firms  Random Coefficients

, where Di : firm i’s observed characteristics,

vi : firm i’s unobservables on its αi and βi,

∏ : a matrix of how parameters depend on firm i’s observables,

Σ : a matrix of how parameters depend on firm i’s unobservables.

By combining these two equations above,

< Issue 2 > Discrete Choice  Multinomial Logit Model

Conditional on the characteristics (X, ξ), each firm will decide its location to maximize its profit, locating either one or zero times each year.

If we use the Type I extreme-value distribution for εijt, then this is the multinomial logit model (McFadden, 1973).

< Issue 3 > An Endogeneity Problem  Instrumental Variables (Z)

The location choice of a firm is a function of the gross regional product, and this gross regional product is a function of the number of firms in each region.

< Issue 4 > GMM estimationMin ||s – S||

If can be inverted to produce the vector , then we can construct the unobservables of a region j as follows.

Our “moment equation” is .

Therefore, the GMM estimator from the moment expression is

, where Φ is a consistent estimator of E[Z’ωω’Z]

  1. Data and its sources
  • General description

- Location choices of inward FDI firms in Korea

- Time coverage: 2000 – 2004

- The manufacturing industry: KSIC 15 – 37

- Three types of sub-national regions in Korea

(1) Province level: 16 regions  80 observations

(2) County level: 234 regions  1,170 observations

(3) Cluster level: 96 regions (by KIET)  480 observations

  • Dependent variable

- The number of inward FDI firms (1,212 data observations)

- Location share of each region (normalized)

- Log (location share of each region)

 INSC (InvestmentNotificationStatisticsCenter) Database

  • Independent variables

- Location-specific characteristics of each region: Xj

(1) Natural resource seeking: per employment wage (in million won) ‘-’

(2) Market seeking: gross regional product (in million won)* ‘+’

(3) Efficiency seeking: the Industrial Complex for FDIs (dummy) ‘+’/‘-’

(4) Strategic asset seeking: the number of patents (per 1,000 people) ‘+’

 Statistics on regional economies in Korea, NSO

Statistics on the manufacturing industries in Korea, NSO

- Firm-specific characteristics for inward FDI firms: Di

(1) Industry: high-tech/low-tech (by OECD and KIET, dummy) ‘+’

(2) Nationality: OECD / non-OECD countries (dummy) ‘+’

(3) Foreign ownership (%) ‘+’

(4) R&D expenditure (in 100 million won)‘+’/‘-’

(5) R&D intensity (%)‘+’/‘-’

 Survey information by MOCIE on 113 FDI firms in Korea Stock Exchange and 125 FDI firms in KOSDAQ.

  1. Modification of Nevo’s Inputs

1. Creation of data sets

  • ps2.mat

- id: ID variable

- id_demo: ID variable for v and demogr

- s_jt: Dependent variable

- x1: Independent variables for the linear part of the estimation

- x2: Independent variables for the non-linear part of the estimation

- v: Random draws for the estimation

- demogr: Draws of demographic variables

  • iv.mat: Instrumental variables

2. Modification of Nevo’s codes

  • rc_dc.m: A main script file that reads in the data and calls the other functions  Pay attention on the dimension of each matrix
  • gmmobjg.m: Computes the GMM objective function and its gradient
  • meanval.m: Computes the mean utility level
  • mufunc.m: Computes the non-linear part of the utility
  • mktsh.m: Computes the market share for each product
  • ind_sh.m: Computes the individual probabilities of choosing each brand
  • jacob.m: Computes the Jacobian of the implicit function that defines the mean utility
  • var_cov.m: Computes the var-cov matrix of the estimates
  • cd_dum.m: Creates a set of dummy variables
  1. Some Pitfalls in Nevo’s Approach
  • s_jt should not be zero.
  • Obsolete functions in Matlab v.7: FMINU  FMINUNC
  • Choice of appropriate IVs
  • Importance of initial values
  • Sensitivity to data scaling

1