Spatial Econometrics
Lecture 8: Interpretation of parameters in spatial models
Andrzej Torój
Institute of Econometrics – Department of Applied Econometrics
Outline
1 Interpretation of parameters Single-source models Multi-source models
2 Statistical inference for spatial multipliers and average effects Inference in models with SAR component
Inference in models without SAR component
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Plan prezentacji
1 Interpretation of parameters
2 Statistical inference for spatial multipliers and average effects
Marginal effects of growth in regressors in the linear model (1)
Consider the SAR model (and multi-source models containing ρWy):
y = ρWy + Xβ + ε
If ρ = 0 then the marginal impact of the variable xk on y is simply the parameter βk, i.e. the respective element of β vector (hence the interpretation of parameters in a linear model):
y = Xβ + ε
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Marginal effects of growth in regressors in the linear model (1)
Consider the SAR model (and multi-source models containing ρWy):
y = ρWy + Xβ + ε
If ρ = 0 then the marginal impact of the variable xk on y is simply the parameter βk, i.e. the respective element of β vector (hence the interpretation of parameters in a linear model):
y = Xβ + ε
Marginal effects of growth in regressors in the linear model (2)
We can consider such an impact for every pair of observations in the sample by differentiating the vectory with respect to xk(each of them of length N – number of observations, hence the result is a matrix sized N × N):
Sk= ∂x∂y
k =∂[x∂xkβk]
k = βkIN=
∆xk,1= 1
↓ βk
∆xk,2= 1
↓
0 · · ·
∆xk,N= 1
↓
0 → ∆y1
0 βk · · · 0 → ∆y2
.. .
..
. . .. ... →
.. .
0 0 · · · βk → ∆yN
diagonal matrix: change ofxkfor i -th observation has no impact ony for any other observation
spherical matrix (equality of diagonal elements): the marginal impact is equal across observations
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Marginal effects of growth in regressors in SAR model (1)
Consider the SAR model (and multi-source models containing ρWy):
y = ρWy + Xβ + ε
If ρ 6= 0, the computation of marginal effect requires first to solve this equation for y:
y = (I − ρW)−1Xβ + (I − ρW)−1ε
Marginal effects of growth in regressors in SAR model (1)
Consider the SAR model (and multi-source models containing ρWy):
y = ρWy + Xβ + ε
If ρ 6= 0, the computation of marginal effect requires first to solve this equation for y:
y = (I − ρW)−1Xβ + (I − ρW)−1ε
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Marginal effects of growth in regressors in SAR model (2)
The effect of differentiation is, again, the N × N matrix, but this time...
Sk=∂x∂y
k =∂[(I−ρW)−1xkβk]
∂xk = βk(I − ρW)−1=
∆xk,1= 1
↓ m1,1
∆xk,2= 1
↓
m1,2 · · ·
∆xk,N= 1
↓
m1,N → ∆y1
m2,1 m2,2 · · · m2,N → ∆y2
.. .
..
. . .. ... →
.. .
mN,1 mN,2 · · · mN,N → ∆yN
this matrix is neigher diagonal (dependence between observations: change inxk
for i -th observation can have an impact ony for other observations)...
...nor spherical (various direct impacts across observations).
Interactions between neighbours – complication of the multipliers (1)
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Interactions between neighbours – complication of the multipliers (2)
Cat exerts an impact on Pat (β1). Pat carries forward this impact to Mat (β1·ρ2), Mat to Pat (β1ρ2·ρ1), Pat to Mat again (β1ρ2ρ1·ρ2)...
In other words (in spatial modelling terms):
Sk = βk(I − ρW)−1 = βkI + βkρW + βkρ2W2+ . . . βkI: effect without taking space into consideration (diagonal, spherical)
βkρW: effect of first-order neighbourhood (this effect still keeps the diagonal of Sk equal, because of zero diagonal in W) βkρ2W2+ . . .: the diagonal of Sk is no longer equal
(second-order neighbourhood effects, and further-orders, operate with different strength for individual observations)
Interactions between neighbours – complication of the multipliers (2)
Cat exerts an impact on Pat (β1). Pat carries forward this impact to Mat (β1·ρ2), Mat to Pat (β1ρ2·ρ1), Pat to Mat again (β1ρ2ρ1·ρ2)...
In other words (in spatial modelling terms):
Sk = βk(I − ρW)−1 = βkI + βkρW + βkρ2W2+ . . . βkI: effect without taking space into consideration (diagonal, spherical)
βkρW: effect of first-order neighbourhood (this effect still keeps the diagonal of Sk equal, because of zero diagonal in W) βkρ2W2+ . . .: the diagonal of Sk is no longer equal
(second-order neighbourhood effects, and further-orders, operate with different strength for individual observations)
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Interpreting the parameters of SEM and SLX models
Much simpler:
SEM: y = Xβ +
M
z }| {
(I −λW)−1u β – like in the linear model
λ – through the lens of matrix M = [mi ,j]:
∆uj = 1 → ∆yi= mi ,j (hardly ever useful in practice...) SLX: y = Xβ +WXθ+ ε
β and θ – through the marginal effects:
Sk = ∂x∂y
k = ∂[xk(β∂xkI+θkW)]
k = βkI + θkW ...and further as in SAR
Interpreting the parameters of SEM and SLX models
Much simpler:
SEM: y = Xβ +
M
z }| {
(I −λW)−1u β – like in the linear model
λ – through the lens of matrix M = [mi ,j]:
∆uj = 1 → ∆yi= mi ,j (hardly ever useful in practice...) SLX: y = Xβ +WXθ+ ε
β and θ – through the marginal effects:
Sk = ∂x∂y
k = ∂[xk(β∂xkI+θkW)]
k = βkI + θkW ...and further as in SAR
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Interpreting the parameters of SEM and SLX models
Much simpler:
SEM: y = Xβ +
M
z }| {
(I −λW)−1u β – like in the linear model
λ – through the lens of matrix M = [mi ,j]:
∆uj = 1 → ∆yi= mi ,j (hardly ever useful in practice...) SLX: y = Xβ +WXθ+ ε
β and θ – through the marginal effects:
Sk = ∂x∂y
k = ∂[xk(β∂xkI+θkW)]
k = βkI + θkW ...and further as in SAR
Direct, indirect and total effects (1)
Presentation of the entire matrix Sk is burdensome and normally useless. There are synthetic measures instead.
Direct effectfor the variable xk: EkD = tr (SNk).
Impact of unit increase in xkin a given region on y inthe same region (averaged over regions).
Total effectfor the variable xk: EkT = N1ΣNi =1
ΣNj =1mi ,j . Average column sum over the matrix Sj.
Impact of unit increase in xk in a given region on y inall regions jointly (averaged over the regions where the impulse can potentially occur).
Indirect effect for the variable xk: EkI = EkT− EkD.
Impact of unit increase in xj in a given region on y inall other regions jointly (averaged over the regions where the impulse can potentially occur).
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Direct, indirect and total effects (1)
Presentation of the entire matrix Sk is burdensome and normally useless. There are synthetic measures instead.
Direct effectfor the variable xk: EkD = tr (SNk).
Impact of unit increase in xkin a given region on y inthe same region (averaged over regions).
Total effectfor the variable xk: EkT = N1ΣNi =1
ΣNj =1mi ,j . Average column sum over the matrix Sj.
Impact of unit increase in xk in a given region on y inall regions jointly (averaged over the regions where the impulse can potentially occur).
Indirect effect for the variable xk: EkI = EkT− EkD.
Impact of unit increase in xj in a given region on y inall other regions jointly (averaged over the regions where the impulse can potentially occur).
Direct, indirect and total effects (1)
Presentation of the entire matrix Sk is burdensome and normally useless. There are synthetic measures instead.
Direct effectfor the variable xk: EkD = tr (SNk).
Impact of unit increase in xkin a given region on y inthe same region (averaged over regions).
Total effectfor the variable xk: EkT = N1ΣNi =1
ΣNj =1mi ,j . Average column sum over the matrix Sj.
Impact of unit increase in xk in a given region on y inall regions jointly (averaged over the regions where the impulse can potentially occur).
Indirect effect for the variable xk: EkI = EkT− EkD.
Impact of unit increase in xj in a given region on y inall other regions jointly (averaged over the regions where the impulse can potentially occur).
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Direct, indirect and total effects (1)
Presentation of the entire matrix Sk is burdensome and normally useless. There are synthetic measures instead.
Direct effectfor the variable xk: EkD = tr (SNk).
Impact of unit increase in xkin a given region on y inthe same region (averaged over regions).
Total effectfor the variable xk: EkT = N1ΣNi =1
ΣNj =1mi ,j . Average column sum over the matrix Sj.
Impact of unit increase in xk in a given region on y inall regions jointly (averaged over the regions where the impulse can potentially occur).
Indirect effect for the variable xk: EkI = EkT− EkD.
Impact of unit increase in xj in a given region on y inall other regions jointly (averaged over the regions where the impulse can potentially occur).
Direct, indirect and total effects (2)
Usually, i.e. with ρ ∈ (0; 1), it holds that βk < EkD < EkT. The standard interpretation of βk as in the linear model is erroneous, a similar interpretation is valid at most for EkT (up to averaging over regions).
These effects are computed using post-estimation command impacts
impacts <- impacts(model, listw = W)
Effects from SLX model are only computed for one estimation method (command lmSLX) and, in such a case, the argument listw is not required.
It does not make sense to compute impacts form SEM model (why?).
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Interpretation of parameters in multi-source models
SARAR: y = (I −ρW)−1Xβ +
M∗
z }| {
(I −ρW)−1(I −λW)−1u β and ρ – jointly as in SAR model (impacts in R) ρ and λ – jointly through the lens of M∗ (like SEM) SDM:
y = (I −ρW)−1Xβ + (I −ρW)−1WXθ+ (I −ρW)−1ε β, θ, ρ – marginal effects: Sk =∂x∂y
k =
∂{xk[βk(I−ρW)−1+θk(I−ρW)−1W]}
∂xk = (I − ρW)−1[βkI + θkW]
...and further as in SAR model (impacts in R) SDEM: y = Xβ +WXθ+ (I −λW)−1u
β and θ – like in SLX (impacts in R) λ – like in SEM
Interpretation of parameters in multi-source models
SARAR: y = (I −ρW)−1Xβ +
M∗
z }| {
(I −ρW)−1(I −λW)−1u β and ρ – jointly as in SAR model (impacts in R) ρ and λ – jointly through the lens of M∗ (like SEM) SDM:
y = (I −ρW)−1Xβ + (I −ρW)−1WXθ+ (I −ρW)−1ε β, θ, ρ – marginal effects: Sk =∂x∂y
k =
∂{xk[βk(I−ρW)−1+θk(I−ρW)−1W]}
∂xk = (I − ρW)−1[βkI + θkW]
...and further as in SAR model (impacts in R) SDEM: y = Xβ +WXθ+ (I −λW)−1u
β and θ – like in SLX (impacts in R) λ – like in SEM
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Interpretation of parameters in multi-source models
SARAR: y = (I −ρW)−1Xβ +
M∗
z }| {
(I −ρW)−1(I −λW)−1u β and ρ – jointly as in SAR model (impacts in R) ρ and λ – jointly through the lens of M∗ (like SEM) SDM:
y = (I −ρW)−1Xβ + (I −ρW)−1WXθ+ (I −ρW)−1ε β, θ, ρ – marginal effects: Sk =∂x∂y
k =
∂{xk[βk(I−ρW)−1+θk(I−ρW)−1W]}
∂xk = (I − ρW)−1[βkI + θkW]
...and further as in SAR model (impacts in R) SDEM: y = Xβ +WXθ+ (I −λW)−1u
β and θ – like in SLX (impacts in R) λ – like in SEM
Plan prezentacji
1 Interpretation of parameters
2 Statistical inference for spatial multipliers and average effects
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Inference in models with SAR component
SAR / SARAR: ˆSk = (I − ˆρW)−1βˆk SDM: ˆSk = (I − ˆρW)−1h ˆβkI + ˆθkWi
In both cases, multiplier matrices are non-linear functions of the estimated parameters (with respect to ˆρ).
The assessment of the related statistical uncertainty (and the resulting uncertainty of average effects) requires to use approximations or numerical methods.
Inference in models with SAR component
SAR / SARAR: ˆSk = (I − ˆρW)−1βˆk SDM: ˆSk = (I − ˆρW)−1h ˆβkI + ˆθkWi
In both cases, multiplier matrices are non-linear functions of the estimated parameters (with respect to ˆρ).
The assessment of the related statistical uncertainty (and the resulting uncertainty of average effects) requires to use approximations or numerical methods.
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Inference in models with SAR component
SAR / SARAR: ˆSk = (I − ˆρW)−1βˆk SDM: ˆSk = (I − ˆρW)−1h ˆβkI + ˆθkWi
In both cases, multiplier matrices are non-linear functions of the estimated parameters (with respect to ˆρ).
The assessment of the related statistical uncertainty (and the resulting uncertainty of average effects) requires to use approximations or numerical methods.
Method 1: delta approximation (SAR model) (1)
Let ∇k = ∂ ˆS
∗ k
h ∂ ˆρ ∂ ˆβk i, where ˆS∗k denotes a vertical vector of length N · N composed of elements from matrix ˆSk. The matrix of derivatives ∇k is sized (N · N) × 2 (derivatives of N · N elements with respect to 2 parameters).
∂ ˆSk
∂ ˆβk = ∂(I− ˆρW)
−1βˆk
∂ ˆβk = (I − ˆρW)−1 ≡ M
∂ ˆSk
∂ ˆρ = ∂(I− ˆρW)
−1βˆk
∂ ˆρ = − (I − ˆρW)−1(−W) (I − ˆρW)−1 ≡ MWM
From matrix calculus:
∂U(x )−1
∂x = −U (x )−1 ∂U(x)∂x U (x)−1
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Method 1: delta approximation (SAR model) (1)
Let ∇k = ∂ ˆS
∗ k
h ∂ ˆρ ∂ ˆβk i, where ˆS∗k denotes a vertical vector of length N · N composed of elements from matrix ˆSk. The matrix of derivatives ∇k is sized (N · N) × 2 (derivatives of N · N elements with respect to 2 parameters).
∂ ˆSk
∂ ˆβk = ∂(I− ˆρW)
−1βˆk
∂ ˆβk = (I − ˆρW)−1 ≡ M
∂ ˆSk
∂ ˆρ = ∂(I− ˆρW)
−1βˆk
∂ ˆρ = − (I − ˆρW)−1(−W) (I − ˆρW)−1 ≡ MWM
From matrix calculus:
∂U(x )−1
∂x = −U (x )−1 ∂U(x)∂x U (x)−1
Method 1: delta approximation (SAR model) (1)
Let ∇k = ∂ ˆS
∗ k
h ∂ ˆρ ∂ ˆβk i, where ˆS∗k denotes a vertical vector of length N · N composed of elements from matrix ˆSk. The matrix of derivatives ∇k is sized (N · N) × 2 (derivatives of N · N elements with respect to 2 parameters).
∂ ˆSk
∂ ˆβk = ∂(I− ˆρW)
−1βˆk
∂ ˆβk = (I − ˆρW)−1 ≡ M
∂ ˆSk
∂ ˆρ = ∂(I− ˆρW)
−1βˆk
∂ ˆρ = − (I − ˆρW)−1(−W) (I − ˆρW)−1 ≡ MWM
From matrix calculus:
∂U(x )−1
∂x = −U (x )−1 ∂U(x)∂x U (x)−1
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Method 1: delta approximation (SAR model) (2)
Let Σk =
Var ( ˆρ) Cov
ˆ ρ, ˆβk
. Var ˆβk
be a respective sub-matrix 2 × 2 from the variance-covariance matrix of the parameter estimates.
1-st order Taylor expansion of the variance-covariance formula for ˆS∗k yields the following matrix sized (N · N) × (N · N):
Var ˆS∗k
≈ ∇Tk · (Σk)−1· ∇k
Its diagonal elements can be rearranged back into the matrix sized N × N, and their square roots approximate the standard errors ˆSk.
Method 1: delta approximation (SAR model) (2)
Let Σk =
Var ( ˆρ) Cov
ˆ ρ, ˆβk
. Var ˆβk
be a respective sub-matrix 2 × 2 from the variance-covariance matrix of the parameter estimates.
1-st order Taylor expansion of the variance-covariance formula for ˆS∗k yields the following matrix sized (N · N) × (N · N):
Var ˆS∗k
≈ ∇Tk · (Σk)−1· ∇k
Its diagonal elements can be rearranged back into the matrix sized N × N, and their square roots approximate the standard errors ˆSk.
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Method 1: delta approximation (SAR model) (2)
Let Σk =
Var ( ˆρ) Cov
ˆ ρ, ˆβk
. Var ˆβk
be a respective sub-matrix 2 × 2 from the variance-covariance matrix of the parameter estimates.
1-st order Taylor expansion of the variance-covariance formula for ˆS∗k yields the following matrix sized (N · N) × (N · N):
Var ˆS∗k
≈ ∇Tk · (Σk)−1· ∇k
Its diagonal elements can be rearranged back into the matrix sized N × N, and their square roots approximate the standard errors ˆSk.
Method 2: bootstrap (1)
Speaking of bootstrap methods, we usually mean theirnon-parametric variant:
Draw independently with replacement N regions from the population of N regions.
For each draw r = 1, ..., R compute ˆS(r )k .
From the obtained distribution ˆS(r )k we derive the desired quantiles for each element of the matrix (e.g. of order 0.025 and 0.975, yielding 95-percent confidence interval).
This will not work well in spatial models!
Observations are not independent. This sampling scheme tears apart the network of dependencies summarized by ˆρ, which will be subject to our inference later on.
One potential solution is the block bootstrap method, as in dynamic time series models (observation + lag).
Quite difficult in dynamic spatial panels (two blocking dimensions: time and space!).
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Method 2: bootstrap (1)
Speaking of bootstrap methods, we usually mean theirnon-parametric variant:
Draw independently with replacement N regions from the population of N regions.
For each draw r = 1, ..., R compute ˆS(r )k .
From the obtained distribution ˆS(r )k we derive the desired quantiles for each element of the matrix (e.g. of order 0.025 and 0.975, yielding 95-percent confidence interval).
This will not work well in spatial models!
Observations are not independent. This sampling scheme tears apart the network of dependencies summarized by ˆρ, which will be subject to our inference later on.
One potential solution is the block bootstrap method, as in dynamic time series models (observation + lag).
Quite difficult in dynamic spatial panels (two blocking dimensions: time and space!).
Method 2: bootstrap (2)
Parametric bootstrap: having estimated the model parameters ˆρ, ˆβ...:
Generate R draws of the error vector ε(r ): i.i.d., N 0, ˆσ2 For every r , reconstructy(r )= f
ε(r ),X, W, ˆβ, ˆρ . Estimate
h ˆ ρ(r ), ˆβ(r )
i
= f (y(r ),X, W) and, based on this result, compute Sˆ(r )k .
Equivalently (and more simply in practice):
Draw R timesh ˆ ρ(r ), ˆβ(r )
i
∼ MVNh ˆ ρ, ˆβ
i , Cov
h ˆ ρ, ˆβ
i
and, for each r , compute ˆS(r )k .
From the obtained distribution ˆS(r )k we derive the desired quantiles for each element of the matrix (e.g. of order 0.025 and 0.975, yielding 95-percent confidence interval).
Limitation: we assume that our model is statistically correct!
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Method 2: bootstrap (2)
Parametric bootstrap: having estimated the model parameters ˆρ, ˆβ...:
Generate R draws of the error vector ε(r ): i.i.d., N 0, ˆσ2 For every r , reconstructy(r )= f
ε(r ),X, W, ˆβ, ˆρ . Estimate
h ˆ ρ(r ), ˆβ(r )
i
= f (y(r ),X, W) and, based on this result, compute Sˆ(r )k .
Equivalently (and more simply in practice):
Draw R timesh ˆ ρ(r ), ˆβ(r )
i
∼ MVNh ˆ ρ, ˆβ
i , Cov
h ˆ ρ, ˆβ
i
and, for each r , compute ˆS(r )k .
From the obtained distribution ˆS(r )k we derive the desired quantiles for each element of the matrix (e.g. of order 0.025 and 0.975, yielding 95-percent confidence interval).
Limitation: we assume that our model is statistically correct!
Method 2: bootstrap (2)
Parametric bootstrap: having estimated the model parameters ˆρ, ˆβ...:
Generate R draws of the error vector ε(r ): i.i.d., N 0, ˆσ2 For every r , reconstructy(r )= f
ε(r ),X, W, ˆβ, ˆρ . Estimate
h ˆ ρ(r ), ˆβ(r )
i
= f (y(r ),X, W) and, based on this result, compute Sˆ(r )k .
Equivalently (and more simply in practice):
Draw R timesh ˆ ρ(r ), ˆβ(r )
i
∼ MVNh ˆ ρ, ˆβ
i , Cov
h ˆ ρ, ˆβ
i
and, for each r , compute ˆS(r )k .
From the obtained distribution ˆS(r )k we derive the desired quantiles for each element of the matrix (e.g. of order 0.025 and 0.975, yielding 95-percent confidence interval).
Limitation: we assume that our model is statistically correct!
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Method 2: bootstrap (2)
Parametric bootstrap: having estimated the model parameters ˆρ, ˆβ...:
Generate R draws of the error vector ε(r ): i.i.d., N 0, ˆσ2 For every r , reconstructy(r )= f
ε(r ),X, W, ˆβ, ˆρ . Estimate
h ˆ ρ(r ), ˆβ(r )
i
= f (y(r ),X, W) and, based on this result, compute Sˆ(r )k .
Equivalently (and more simply in practice):
Draw R timesh ˆ ρ(r ), ˆβ(r )
i
∼ MVNh ˆ ρ, ˆβ
i , Cov
h ˆ ρ, ˆβ
i
and, for each r , compute ˆS(r )k .
From the obtained distribution ˆS(r )k we derive the desired quantiles for each element of the matrix (e.g. of order 0.025 and 0.975, yielding 95-percent confidence interval).
Limitation: we assume that our model is statistically correct!
Evaluating the uncertainty around average effects
R delivers an off-the-shelf bootstrap procecure related to the average direct, indirect and total effects.
Using the command impacts, we should supply the following additional arguments:
R = 200 (number of iterations R in the bootstrap procedure) zstats = TRUE
For models with the SAR component (SAR, SDM, SARAR, GNS), following the use of impacts, one can report the confidence intervals for the three types of effects, at a given confidence level (using the command
HPDinterval.lagImpact), and verify whether they contain zero.
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Inference in models without SAR component
Both methods can be applied to SLX and SDEM models. But senseless to use, since an an exact analytical formula for the distribution of linear functions of linear model parameters exists (derived e.g. inGreene, 2003)
In such cases, delta method provides the exact formula instead of an approximation, because Sk= βkI + θkWis a linear function with respect to βk and θk.
Bootstrap method also ensures the correct inference (though it is computationally inefficient).
For the SLX and SDEM models, summary(impacts) will test three hypotheses on the basis of the analytical distribution (p-values are printed):
H0: EkD = 0 H0: EkI = 0 H0: EkT = 0
Interaction variables
Caution! Throughout this lecture (regardless of the model and inference method) we assumed that regressor matrix X does not contain interaction variables (of type x1· x2).
Otherwise the multipliers for variables x1 and x2 should be derived differently. I recommend this as an exercise.
Spatial multipliers for a given region depend on:
size of the non-spatial impact of X on y (β) intensity of the spatial processes (ρ, θ)
degree of a region’s integration into the network (W) in interaction variable models: level of other predictors in the analysed region and other regions (X)
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics
Interaction variables
Caution! Throughout this lecture (regardless of the model and inference method) we assumed that regressor matrix X does not contain interaction variables (of type x1· x2).
Otherwise the multipliers for variables x1 and x2 should be derived differently. I recommend this as an exercise.
Spatial multipliers for a given region depend on:
size of the non-spatial impact of X on y (β) intensity of the spatial processes (ρ, θ)
degree of a region’s integration into the network (W) in interaction variable models: level of other predictors in the analysed region and other regions (X)
Homework 7
Select one explanatory variable from the ones discussed in HW 5 and 6.
For the best model (or, if that was the SEM model – for any other model) compute the direct, indirect and total effect for this variable.
Test the significance of this indirect effect.
Illustrate on a map the impact of a unit change of the indicated regressor in the selected region on the dependent variable across all regions. Comment on the results.
The PDF file should contain the calculation results with all interpretations and the map.
Andrzej Torój Institute of Econometrics – Department of Applied Econometrics