Spatial Stochastic Frontier Models On spatial dependence in stochastic frontier models with an application to European regional economic performance Thomas de Graa Department of spatial economics, Vrije Universiteit Amsterdam http://www.thomasdegraaff.nl/ Package development at https://github.com/Thdegraaff/SpatFrontier/ or http://www.thomasdegraaff.nl/SpatFrontier/ May 25, 2018
Regional economic performance Regional economic performance GDP per capita across Europe in 2007EU27 = 100 (source: cambridge econometrics database 50 100 150 200 250 300 Huge dierences between and within countries Dierent use of endowments Dierent relative location within the network
Related literature Related literature Starting point: Weinstein (Technometrics, 1964) Econometric literature: Aigner et al. (1977); Meeusen & van den Broeck (1977). Statistical literature: Azzalini (1985); Azzalini & Capitanio (1999); Dominguez-Molina et al. (2003)
Related literature Related literature Starting point: Weinstein (Technometrics, 1964) Econometric literature: Aigner et al. (1977); Meeusen & van den Broeck (1977). Statistical literature: Azzalini (1985); Azzalini & Capitanio (1999); Dominguez-Molina et al. (2003) Regional stochastic frontiers: Puig-Junoy (2001); Alvarez (2007)
The stochastic production frontier model Stochastic production frontiers In general, regional production,y, can be modeled as: y = f (X; β)TE, Assume Cobb-Douglas and TE = exp(−µ), then: ln y = ln(X)β − µ + ν =Observed residuals , Where usually: µ ∼ N+ 0, σ2 µ so µ ≥ 0 ν ∼ N 0, σ2 ν and ν and µ are independent of each other
The stochastic production frontier model Tradional (econometric) approach For likelihood purposes one usually considers the composite stochastic variable = ν − µ en then nd the marginal density of : f ( ) = ∞ 0 f (µ, )dµ = 2 σ φ σ Φ − λ σ Technical eciencies can be retrieved by f (µ| ).
The stochastic production frontier model Imposing spatial structure A typology Note that A = (I − λW) and B = (I − ρW) 1 Spatial dependence in technical eciency component µ ln y = ln(X)β −A−1µ + ν 2 Spatial dependence in error term ν: ln y = ln(X)β −u + A−1ν 3 Spatial lag separate from technical eciency component µ ln y = B−1 [ln(X)β] +B−1ν − µ 4 Spatial lag ln y = B−1 [ln(X)β] + B−1 [−µ + ν]
The stochastic production frontier model Alternative statistical approach Alternative specication, skew normal distribution: Z ∼ SN(α) = 2φ( )Φ(α ) Which has two types of genesis: 1 By convolution: Z = δ|U| + 1 − δ2V 2 By conditioning: Z = (V |U > 0)
The stochastic production frontier model Alternative statistical approach Alternative specication, skew normal distribution: Z ∼ SN(α) = 2φ( )Φ(α ) Which has two types of genesis: 1 By convolution: Z = δ|U| + 1 − δ2V 2 By conditioning: Z = (V |U > 0) And is closed under ane transformations: if Y = Xβ + σZ then Y ∼ SN(Xβ, σ, α)
The stochastic production frontier model Multivariate skew-normal distributions Via conditioning If we partition Ψ in component U and V of size 1 and n respectively, such that: Ψ = U V ∼ N 1+n(0, Σ∗), Σ∗ = 1 δt δ Σ then: Z = (V |U > 0) ∼ SNn(0, Σ, α), α = (1 − δtΣδ) 1 2 σ−1δ
The stochastic production frontier model Combining spatial structure with stochastic frontiers For the spatial error model If the model is: ln y = ln(X)β −µ + A−1ν then ∼ SN(ln y − ln(X)β, σ2(A t A)−1, α) For the spatial lag model If the model is: ln y = B−1 [ln(X)β] −µ + B−1ν then: ∼ SN(B−1(ln y − Xβ), σ2(B t B)−1, α)
Data Regional production function data Cambridge Econometrics Dataset for Employment Regional value added and capital stock from regional supply and use framework (Thissen et al. 20013a, 2013b and 2013c) Regional trade database between 256 European Nuts2 regions over 15 sectors (20002010) No Perpetual Inventory Method but value added for capital → Using country specic interest rates we can then nd capital stock (VAcapital = rK) Sample period: mean over 20002010 Spatial weights matrix W: 4 nearest neighbours
Conclusion & further research In conclusion & further research Introducing spatial structure in stochastic frontier models is possible and for some models relatively straightforward, but Multivariate distribution of error term should be explicitly taken into account Extension: Ψ ∼ U V ∼ Nd+n (0, Ω∗) , Ω∗ = Φ ∆ ∆ Ω . But then Z = (V |U > 0) becomes more complex and estimation requires Bayesian estimation in combination with simulation techniques