Geographic Knowledge Spillovers in French Regions
Using Indirect and Feedbacks effets to Measure Geographic Knowledge Spillovers in French Regions ERCIM 2013
Thibault LAURENT (GREMAQ) Work in collaboration with In`es MOUSSA (LEREPS) University of Toulouse
[email protected]
14th December of 2013
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions
1 Introduction
2 Spatial Exploratory Data Analysis
3 Spatial Econometric Models for Panel Data
4 Conclusion
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Introduction
Context
Problem: relationship between Patents and R&D. Spatial unit: the N = 94 French metropolitan NUTS 3 regions (department). Time period: from 1995 to 2008 (T = 14). Objective: measure geographic knowledge spillovers e.g. distinguish and measure direct and indirect effects. Data source: EUROSTAT and French National R&D survey.
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Introduction
An illustration of spatial spillovers (1) A change in R&D in region i has a direct effect on patents of region i.
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Introduction
An illustration of spatial spillovers (2) It also has an indirect or spillover effect that produces a change in the neighborhood region j.
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Introduction
An illustration of spatial spillovers (3) feedback effects: impacts from region j going back to the unit i. Spillovers & as we move to regions located farther away from i.
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Exploratory Data Analysis
1 Introduction
2 Spatial Exploratory Data Analysis
3 Spatial Econometric Models for Panel Data
4 Conclusion
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Exploratory Data Analysis
Dependent variable yi,t Patent count from the European Patent Office (EPO) built from information on addresses of inventors.
Data source: EUROSTAT. R package spacetime (Pebezma, 2012).
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Exploratory Data Analysis
Some remarks on the dependent variable (1) Strong inequality: 70% of the patents have been deposited in 20% of the departments. log(yit ) is not a Poisson distribution.
R packages ggplot2 (Wickham, 2009) and GeoXp (Laurent et al., 2012).
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Exploratory Data Analysis
Some remarks on the dependent variable (2) Strong spatial autocorrelation (WN : spatial contiguity matrix)
Local indicators of spatial association (Anselin, 1995).
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Exploratory Data Analysis
Independent variable: R&Di,t−1 Internal R&D summed up across all establishments in each region (data source: French National R&D survey).
We consider three levels of R&D: R&Di,t−1 ∗ 1i∈G1 , with G1 (n1 =22) includes departments with high values of patents (73% of all the patents). R&Di,t−1 ∗ 1i∈G2 , with G2 (n2 =36) includes departments with low values of patents (21% of all the patents). R&Di,t−1 ∗ 1i∈G3 , with G3 (n3 =36) includes departments with very low values of patents (5% of all the patents).
Thibault Laurent
Micromap plot (Carr et al., 2000). Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Exploratory Data Analysis
Independent variables: POPi,t and GDPi,t (data source: EUROSTAT) These variables give measures of urbanization externalities. POP is population density taken in logarithm.
Thibault Laurent
GDP is GDP per capita taken in logarithm.
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Econometric Models for Panel Data
1 Introduction
2 Spatial Exploratory Data Analysis
3 Spatial Econometric Models for Panel Data
4 Conclusion
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Econometric Models for Panel Data
Fixed effect spatial lag model (1) y = ρ(IT ⊗ WN )y + X β + (ιT ⊗ IN )α + (ιN ⊗ IT )µ + ε, y = (y1,1995 , . . . , y94,1995 , . . . , y94,2008 )0 , vector of size NT . ρ is the spatial autoregressive coefficient and WN , the (raw-standardized) spatial contiguity matrix. X is the matrix of size NT × p of dependent variables. α (time effect), a vector of size T ; µ (individual effect), a vector of size N. ιT (resp. ιN ) a column vector of ones of dimension T (resp. N), IN (resp. IN ) an N × N (T × T ) identity matrix. εi ∼ N(0, σε2 ) Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Econometric Models for Panel Data
Fixed effect spatial lag model (2)
yi,t = ρ
N X
wij yjt + β1 log(1 + R&Di,t−1 ∗ 1i∈G1 ) +
j=1
β2 log(1 + R&Di,t−1 ∗ 1i∈G2 ) + β3 log(1 + R&Di,t−1 ∗ 1i∈G3 ) + β4 log(PIBi,t ) + β5 log(GDPi,t ) + αi + µt + εit , β1 β2 β3 β4 β5 ρ
Coefficient 0.350326 0.136184 -0.043099 0.764799 0.515059 0.108092
Asymptot t-stat 4.6650 2.6647 -1.4539 1.7323 1.9224 2.6636 Thibault Laurent
z-probability 3.087e-06 (***) 0.007705 (**) 0.145985 0.083220 (.) 0.054560 (.) 0.007731 (**)
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Econometric Models for Panel Data
Some tests
Wald χ2 test {β2 − β1 = 0}, {β3 − β1 = 0} and {β3 − β2 = 0} are all rejected. LM test for residuals autocorrelation is not significant (LM=3.686). Haussman test: check that fixed effect is preferred to random effect. Significance of individual and time effects to check
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Econometric Models for Panel Data
Estimation for the impact measures
Lesage and Pace (2009): Sr (W) = (INT − ρW)−1 INT βr with W = (IN ⊗ W )
β1 β2 β3 β4 β5
Direct 0.3511 0.1365 -0.0432 0.7666 0.5163
t-stat 4.838 (***) 2.612 (**) -1.350 1.696 (.) 1.806 (.)
indirect 0.04158 0.01616 -0.00511 0.09078 0.06114
Thibault Laurent
t-stat 2.28 (**) 1.753 (.) -1.119 1.340 1.467
total 0.3927 0.1526 -0.0483 0.8574 0.5774
t-stat 4.759 (***) 2.578 (**) -1.340 1.689 (.) 1.814 (.)
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Spatial Econometric Models for Panel Data
Spatial Partitioning of direct, indirect and total impacts W-Order
Variable R&Di,t−1 ∗ 1i∈G1
W0 W1 W2 W3 W4 P4
q=0
Wq
W-Order
Variable R&Di,t−1 ∗ 1i∈G2
W0 W1 W2 W3 W4 P4
q=0
Wq
Direct
indirect
total
0.35032 (***) 0 0.000838 0.000030 0.000005 0.3511 (***)
0 0.03792 (*) 0.00326 0.00041 0.00004 0.04158 (**)
0.35032 (***) 0.03792 (**) 0.00410 0.00044 0.00005 0.3927 (***)
Direct
indirect
total
0.13617 (**) 0 0.000325 0.000015 0.000002 0.1365 (**)
0 0.01474 (.) 0.00127 0.00016 0.00002 0.01616 (.)
0.13617 (**) 0.01474 (.) 0.00160 0.00017 0.00002 0.1526 (***)
Thibault Laurent
Geographic Knowledge Spillovers in French Regions
Geographic Knowledge Spillovers in French Regions Conclusion
Conclusion We had to distinguish regions depending on the level of innovation. For region with high level of innovation: Direct effect of R&D is strongly significant. Spillovers are significants. Feedback effects are non significant (ρ is not high enough).
Perspectives: add other measures of agglomeration economies (Ellison-Glaeser index), etc. try a Spatial Durbin Model models for panel data and compare the effects with Spatial Lag model.
Thibault Laurent
Geographic Knowledge Spillovers in French Regions