Spatial Economterics Using SMLE

Motivation
The xsmle command

Examples
References
xsmle - A Command to Estimate Spatial Panel Data Models in

Stata
Federico Belotti? ,
Gordon Hughes , Andrea Piano Mortari?
? CEIS, University of Rome Tor Vergata

School of Economics, University of Edinburg.
2013 Italian Stata Users Group Meeting

Florence, November 14 2013
Belotti, Hughes, Piano Mortari xsmle - Spatial Panel Data Models in Stata
Outline
1 Motivation
2 The xsmle command

Options common to all spatial models
The weighting matrix
Models specific options
xsmle postestimation command
3 Examples
Estimation
Postestimation
Testing
Motivation
The xsmle command
Examples
References
A general specification for Spatial Panel models:
n
X K
X K X
X n
yit = + yit 1 + wij yjt + xitk k + wij xjtk k +
j=1 k=1 k=1 j=1
+ i + t + it (1)
Xn
it = mij it + it i = 1, ..., n t = 1, ..., T (2)
j=1
Static Models ( = 0 ) and Dynamic Models ( 6= 0, Yu et al. (2008))

if = 0 and ! Spatial Autoregressive Model with Auto Regressive disturbances
(SAC)
if = 0 ! Spatial Durbin Model (SDM)
if = 0 and = 0 ! Spatial Autoregressive Model (SAR)
if = 0 and = 0 ! Spatial Error Model (SEM)
Pn
if = 0, = 0, and i = j=1 wij i + i ! Generalised Spatial Panel
Random Eects model (GSPRE)
Motivation
The xsmle command
Examples
References
A number of spatial-related routines have been written by users and available

through SSC. A non-comprehensive list includes:
Data management and visualization
shp2dta by K. Crow
spmat by D.M. Drukker et al
spwmatrix by P.W. Jeanty
spmap by M. Pisati
geocode3 by S. Bernhard
Cross sectional data
spreg: SAR, SEM, SAC via ML or GS2SLS by D. M. Drukker et al

spivreg: SAC via GS2SLS by D. M. Drukker et al
spmlreg: SAR, SEM, SDM, SAC via ML by P.W. Jeanty
spatreg: SAR, SEM via ML by M. Pisati
spautoreg: SAR, SEM, SDM, SAC via ML or GS2SLS by E.A. Shehata
Panel data
spreg*xt suite SAR, SEM, SDM, SAC via LS, GLS, GMM or GS2SLS by E.A. Shehata
Motivation
The xsmle command
Examples
References
DGP - 250 replications
n
X
yit = wij yjt + 0.3x1it + 0.7x2it + i + t + it n = 1, . . . , 188 t = 1, . . . , 5
j=1
where the nuisance parameters i (i = 1, . . . , n) are drawn from an iid standard

Gaussian random variable. To allow for dependence between the unit-specific
eects and the regressors, we generate the latter as follows
xkit = 0.4i + (1 0.42 )1/2 zkit ,
where k = 1, 2 and the zkit is an iid standard Gaussian random variable.

= 0.3 = 0.5 = 0.7
bias MSE bias MSE bias MSE
xsmle -0.0013 0.0020 -0.0016 0.0014 -0.0016 0.0007
spregfext 0.1473 0.0255 0.1972 0.0408 0.1859 0.0352
xtivreg2 0.0174 0.0091 0.0153 0.0063 0.0112 0.0033
Motivation Options common to all spatial models
The xsmle command The weighting matrix
Examples Models specific options
References xsmle postestimation command
xsmle fits (balanced) Spatial Panel data models via maximum likelihood (ML)
Requirements:
(At least) Stata Version 10
The n n matrix of spatial weights. xsmle will deal with the longitudinal
dimension automatically
Data must be tsset or xtset
The basic xsmle syntax is the following

xsmle depvar indepvars if in weight , options
The default model is the random-eects SAR model

Only aweight are allowed but the declared weight variable must be
constant within each panel unit
The mi prefix is allowed
Factor variables are allowed
Options common to all spatial models
model(name) specifies the spatial model to be estimated. May be sar for

the Spatial-AutoRegressive model, sdm for the Spatial Durbin Model, sem
for the Spatial-Error Model, sac for the Spatial-Autoregressive with
Spatially Autocorrelated Errors Model, gspre for the Generalised Spatial
Random Eects Model.
re use the random eects estimator; the default. This option cannot be
specified when model(sac).
fe use the fixed eects estimator. This option cannot be specified when
model(gspre).
type(type options [, leeyu]) specifies fixed-eects type; only for fe
estimators. May be ind for spatial fixed eects, time for time fixed eects
or both for both spatial and time fixed eects. Suboption leeyu allows to
transform the data according to Lee and Yu (2010) approach and can be
used only when type(ind).
noconstant suppresses the constant term in the model. Only for re

estimators.
noeffects suppresses the computation of direct, indirect and total eects.
nsim(#) sets the number of simulations for the LeSage and Pace (2009)
procedure to compute the standard errors of the direct, indirect and total
eects.
constraints(constraints list) applies specified linear constraints.
from(init specs) specifies initial values for the coefficients.
level(#) sets confidence level for confidence intervals; default is
level(95).
postscore save observation-by-observation scores in the estimation results
list.
posthessian save the Hessian corresponding to the full set of coefficients
in the estimation results list.
hausman performs the Hausman test.
Variance estimation
This section describes the arguments of the vce(vcetype) option.
oim observed information matrix.
opg outer product of the gradient vectors.
robust clustered sandwich estimator where clustvar is the panelvar.
cluster clustvar clustered sandwich estimator.
dkraay(#) Driscoll-Kraay robust estimator. Where # is the maximum lag
used in the calculation.
In xsmle the spatial weighting matrix can be
a Stata matrix
a spmat object
In both cases the matrix can be standardized or not.

e.g.
a Stata matrix can be created using matrix define, imported from Mata
using st matrix(string scalar name, real matrix) or imported from GIS
softwares like GeoDa using
spwmatrix gal using path to gal file, wname(name of the matrix)
spmat objects are created by spmat
spmat import name of the object using path to file
SAR model
wmatrix(name) specifies the weight matrix for the spatial-autoregressive

term.
dlag includes (time) lagged dependent variable in the model.
SDM model

term.
dmatrix(name) specifies the weight matrix for the spatially lagged
regressors; default is to use the matrix specified in wmat(name).
durbin(dvarlist) specifies the regressors that have to be spatially lagged;
default is to lag all independent variables specified in varlist.
dlag includes (time) lagged dependent variable in the model.
SEM model
ematrix(name) specifies the weight matrix for the spatial-autocorrelated

error term.
SAC model

term.
error term.
GSPRE model
wmatrix(name) specifies the weight matrix for the spatial-autocorrelated

random-eects.
error term.
error(#) defines the structure of the model. # is equal to 1 when
6= 6= 0, # is equal to 2 when = 0, # is equal to 3 when = 0, # is
equal to 4 when = .
Postestimation command allows to post-estimate spatial fixed or random

eects. The methods implemented in this command are the panel data
extensions of those available in Drukker, Prucha, and Raciborski (2011)

predict type newvar if in , statistic
where statistic includes:
rform the default, calculates predicted values from the reduced-form
equation: yit = (In W ) 1 (xit + i )
limited predicted values based on the limited information set. This
option is available only when model(sac).
naive predicted values based on the observed values of
yit = Wyit + xit + i
xb calculates the linear prediction including the fixed or random eect
xit + i .
a estimates i , the fixed or random-eect. In the case of fixed-eects
models, this statistic is allowed only when type(ind)
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
DGP - Fixed eects SDM
n
X n
X
yit = 0.3 wij yjt + 0.5x1it 0.3x2it 0.2x3it + 0.3 wij x1it +
j=1 j=1
n
X n
X
+ 0.6 wij x2it + 0.9 wij x3it + i + t + it n = 1, ..., 188 t = 1, ..., 5
j=1 j=1
where the nuisance parameters i (i = 1, . . . , n) are drawn from an iid standard

Gaussian random variable. To allow for dependence between the unit-specific
eects and the regressors, we generate the latter as follows
xkit = 0.4i + (1 0.42 )1/2 zkit ,
where k = 1, 2, 3, z1it is standard Gaussian, z2it is N(0, 1.52 ) and z3it is N(0, 22 ).
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
.. *** load a dta dataset containing the spatial contiguity matrix

. use ASL_contiguity_mat_ns.dta, clear
. *** get an spmat objects from dta

. spmat dta W W*, replace
. *** Summarize the spmat obj

. spmat summarize W, links
Summary of spatial-weighting object W

--------------------------------------
Matrix | Description
---------------+----------------------
Dimensions | 188 x 188
Stored as | 188 x 188
Links |
total | 906
min | 1
mean | 4.819149
max | 13
--------------------------------------
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
. ** Fixed-effects Durbin model (correctly specified, row normalized W)

. xsmle y x1 x2 x3, wmat(W) model(sdm) fe type(ind) nsim(500) nolog
Warning: All regressors will be spatially lagged
SDM with spatial fixed-effects Number of obs = 940
Group variable: id Number of groups = 188

Time variable: t Panel length = 5
R-sq: within = 0.5727

between = 0.3663
overall = 0.4554
Mean of fixed-effects = -0.0137
Log-likelihood = -1230.7734
------------------------------------------------------------------------------
y | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Main |
x1 | .5186041 .0364303 14.24 0.000 .4472019 .5900062
x2 | -.2946314 .0236541 -12.46 0.000 -.3409925 -.2482702
x3 | -.1923373 .0192912 -9.97 0.000 -.2301474 -.1545272
-------------+----------------------------------------------------------------
Wx |
x1 | .3772047 .075502 5.00 0.000 .2292235 .5251859
x2 | .5765484 .0449332 12.83 0.000 .4884809 .6646159
x3 | .8692021 .0372769 23.32 0.000 .7961408 .9422634
-------------+----------------------------------------------------------------
Spatial |
rho | .2519025 .0374278 6.73 0.000 .1785454 .3252596
-------------+----------------------------------------------------------------
Variance |
sigma2_e | .7915998 .0366863 21.58 0.000 .7196959 .8635037
-------------+----------------------------------------------------------------
[CONTINUES]
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
In a spatial setting, the eect of an explanatory variable change in a particular

unit aects not only that unit but also its neighbors (LeSage and Pace, 2009).
2 3
k w12 k w1n k
@Y 6 w21 k w2n k 7
= (I W ) 1 6 4
k 7
5
@xnk
wn1 k wn2 k k
If we have only 2 units and 1 regressor:

1 0
SAR and SAC ! (I W ) 1
0 1

1 0
SEM !
0 1

1 1 w12 1
SDM ! (I W )
w21 1 1
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
[CONTINUES]
-------------+----------------------------------------------------------------
Direct |
x1 | .5481382 .0362326 15.13 0.000 .4771237 .6191527
x2 | -.2642811 .0231199 -11.43 0.000 -.3095953 -.2189669
x3 | -.1422518 .0176968 -8.04 0.000 -.1769369 -.1075668
-------------+----------------------------------------------------------------
Indirect |
x1 | .6480929 .090572 7.16 0.000 .470575 .8256108
x2 | .6450951 .0599307 10.76 0.000 .5276331 .7625571
x3 | 1.050599 .058257 18.03 0.000 .9364176 1.164781
-------------+----------------------------------------------------------------
Total |
x1 | 1.196231 .1038425 11.52 0.000 .9927034 1.399759
x2 | .380814 .0677252 5.62 0.000 .2480751 .513553
x3 | .9083474 .0660288 13.76 0.000 .7789334 1.037761
------------------------------------------------------------------------------
. estimates store sdm_fe
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
** Fixed-effects Durbin model (correctly specified, row normalized W)

. xsmle y x1 x2 x3, wmat(W) model(sdm) re type(ind) nsim(500) nolog noeff
Warning: Option type(ind) will be ignored
SDM with random-effects Number of obs = 940


between = 0.4543
overall = 0.4936
------------------------------------------------------------------------------
-------------+----------------------------------------------------------------
Main |
x1 | .6230976 .0408605 15.25 0.000 .5430126 .7031826
x2 | -.2439834 .0264129 -9.24 0.000 -.2957518 -.192215
x3 | -.1688081 .0211584 -7.98 0.000 -.2102778 -.1273385
_cons | -.0169191 .0811545 -0.21 0.835 -.1759791 .1421409
-------------+----------------------------------------------------------------
Wx |
x1 | .3706183 .0824133 4.50 0.000 .2090911 .5321454
x2 | .557779 .0493092 11.31 0.000 .4611347 .6544234
x3 | .8845199 .0411496 21.50 0.000 .8038681 .9651717
-------------+----------------------------------------------------------------
Spatial |
rho | .2472432 .0376366 6.57 0.000 .1734769 .3210096
-------------+----------------------------------------------------------------
Variance |
lgt_theta | -.3920581 .1040247 -3.77 0.000 -.5959428 -.1881735
sigma_e | 1.005536 .0528831 19.01 0.000 .9018867 1.109185
------------------------------------------------------------------------------
. estimates store sdm_re
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
hausman sdm_fe sdm_re, eq(1:1 2:2 3:3)
---- Coefficients ----

| (b) (B) (b-B) sqrt(diag(V_b-V_B))
| sdm_fe sdm_re Difference S.E.
-------------+----------------------------------------------------------------
comp1 |
x1 | .5186041 .6230976 -.1044935 .
x2 | -.2946314 -.2439834 -.050648 .
x3 | -.1923373 -.1688081 -.0235292 .
-------------+----------------------------------------------------------------
comp2 |
x1 | .3772047 .3706183 .0065864 .
x2 | .5765484 .557779 .0187694 .
x3 | .8692021 .8845199 -.0153178 .
-------------+----------------------------------------------------------------
comp3 |
rho | .2519025 .2472432 .0046593 .
------------------------------------------------------------------------------
b = consistent under Ho and Ha; obtained from xsmle
B = inconsistent under Ha, efficient under Ho; obtained from xsmle
Test: Ho: difference in coefficients not systematic

chi2(7) = (b-B)[(V_b-V_B)^(-1)](b-B)
= -47.08 chi2<0 ==> model fitted on these
data fails to meet the asymptotic
assumptions of the Hausman test;
see suest for a generalized test
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
. ** Fixed-effects Durbin model (correctly specified, row normalized W)

. xsmle y x1 x2 x3, wmat(W) model(sdm) fe type(ind) hausman noeff nolog
... estimating random-effects model to perform Hausman test
SDM with spatial fixed-effects Number of obs = 940


between = 0.3663
overall = 0.4554
------------------------------------------------------------------------------
-------------+----------------------------------------------------------------
Main |
x1 | .5186041 .0364303 14.24 0.000 .4472019 .5900062
x2 | -.2946314 .0236541 -12.46 0.000 -.3409925 -.2482702
x3 | -.1923373 .0192912 -9.97 0.000 -.2301474 -.1545272
-------------+----------------------------------------------------------------
Wx |
x1 | .3772047 .075502 5.00 0.000 .2292235 .5251859
x2 | .5765484 .0449332 12.83 0.000 .4884809 .6646159
x3 | .8692021 .0372769 23.32 0.000 .7961408 .9422634
-------------+----------------------------------------------------------------
Spatial |
rho | .2519025 .0374278 6.73 0.000 .1785454 .3252596
-------------+----------------------------------------------------------------
Variance |
sigma2_e | .7915998 .0366863 21.58 0.000 .7196959 .8635037
------------------------------------------------------------------------------
Ho: difference in coeffs not systematic chi2(7) = 89.58 Prob>=chi2 = 0.0000
------------------------------------------------------------------------------
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
. ********* Postestimation
. predict yhat, rform
. predict yhat1, naive
. predict alphahat, a
. sum alpha alphahat
Variable | Obs Mean Std. Dev. Min Max

-------------+--------------------------------------------------------
alpha | 940 .037577 1.058726 -2.261747 3.343453
alphahat | 940 -.013692 1.182919 -2.688471 4.028156
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
Using matrix notation the SDM ( = 0) may be derived from a SEM model
y = X +u
{
u = Wu +
hence
u(1 W) =
y(1 W) = X (1 W) +
y = Wy + X WX +
y = Wy + X + WX +
and test the following constraints

1 = 0 and 6= 0 ) the model is a SAR
2 = ) the model is a SEM.
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
** Test for SAR

. test [Wx]x1 = [Wx]x2 = [Wx]x3 = 0
( 1) [Wx]x1 - [Wx]x2 = 0
( 2) [Wx]x1 - [Wx]x3 = 0
( 3) [Wx]x1 = 0
chi2( 3) = 740.80
Prob > chi2 = 0.0000
.
. ** Test for SEM
. testnl ([Wx]x1 = -[Spatial]rho*[Main]x1) ([Wx]x2 = -[Spatial]rho*[Main]x2) ([
> Wx]x3 = -[Spatial]rho*[Main]x3)
(1) [Wx]x1 = -[Spatial]rho*[Main]x1

chi2(3) = 545.31
Prob > chi2 = 0.0000
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
** SAC Model
. xsmle y x1 x2 x3, wmat(W) emat(W) model(sac) fe type(ind) noeff nolog
SAC with spatial fixed-effects Number of obs = 940


between = 0.0011
overall = 0.0912
------------------------------------------------------------------------------
-------------+----------------------------------------------------------------
Main |
x1 | .3212791 .0341734 9.40 0.000 .2543005 .3882577
x2 | -.3135993 .0232111 -13.51 0.000 -.3590923 -.2681064
x3 | -.2997975 .0178884 -16.76 0.000 -.334858 -.2647369
-------------+----------------------------------------------------------------
Spatial |
rho | -.6676721 .0542468 -12.31 0.000 -.7739939 -.5613504
lambda | .8426981 .0209346 40.25 0.000 .801667 .8837293
-------------+----------------------------------------------------------------
Variance |
sigma2_e | 1.001782 .0440957 22.72 0.000 .9153562 1.088208
------------------------------------------------------------------------------
Motivation
Estimation
The xsmle command
Postestimation
Examples
Testing
References
SDM vs SAC Testing
. qui xsmle y x1 x2 x3, wmat(W) model(sdm) fe type(ind) noeff nolog

. estat ic
-----------------------------------------------------------------------------
Model | Obs ll(null) ll(model) df AIC BIC
-------------+---------------------------------------------------------------
. | 940 . -1230.773 8 2477.547 2516.314
-----------------------------------------------------------------------------
Note: N=Obs used in calculating BIC; see [R] BIC note
. qui xsmle y x1 x2 x3, wmat(W) emat(W) model(sac) fe type(ind) noeff nolog

. estat ic
-----------------------------------------------------------------------------
Model | Obs ll(null) ll(model) df AIC BIC
-------------+---------------------------------------------------------------
. | 940 . -1386.086 6 2784.172 2813.247
-----------------------------------------------------------------------------
Note: N=Obs used in calculating BIC; see [R] BIC note
Motivation
The xsmle command
Examples
References
Lee, L. F. and Yu, J. (2010). Estimation of spatial autoregressive panel data

models with fixed eects. Journal of Econometrics, 154(2):165185.
LeSage, J. P. and Pace, R. K. (2009). Introduction to Spatial Econometrics.
Taylor & Francis.
Yu, J., de Jong, R., and Lee, L. F. (2008). Quasi-maximum likelihood
estimators for spatial dynamic panel data with fixed eects when both n and
t are large. Journal of Econometrics, 146:118134.

Spatial Economterics Using SMLE

Uploaded by

Spatial Economterics Using SMLE

Uploaded by

Motivation

The xsmle command

xsmle - A Command to Estimate Spatial Panel Data Models in

? CEIS, University of Rome Tor Vergata

2013 Italian Stata Users Group Meeting

2 The xsmle command

A general specification for Spatial Panel models:

Static Models ( = 0 ) and Dynamic Models ( 6= 0, Yu et al. (2008))

A number of spatial-related routines have been written by users and available

Cross sectional data

spreg: SAR, SEM, SAC via ML or GS2SLS by D. M. Drukker et al

DGP - 250 replications

where the nuisance parameters i (i = 1, . . . , n) are drawn from an iid standard

xkit = 0.4i + (1 0.42 )1/2 zkit ,

where k = 1, 2 and the zkit is an iid standard Gaussian random variable.

The basic xsmle syntax is the following

The default model is the random-eects SAR model

Options common to all spatial models

model(name) specifies the spatial model to be estimated. May be sar for

noconstant suppresses the constant term in the model. Only for re

In xsmle the spatial weighting matrix can be

In both cases the matrix can be standardized or not.

spwmatrix gal using path to gal file, wname(name of the matrix)

spmat objects are created by spmat

spmat import name of the object using path to file

wmatrix(name) specifies the weight matrix for the spatial-autoregressive

wmatrix(name) specifies the weight matrix for the spatial-autoregressive

ematrix(name) specifies the weight matrix for the spatial-autocorrelated

wmatrix(name) specifies the weight matrix for the spatial-autoregressive

wmatrix(name) specifies the weight matrix for the spatial-autocorrelated

Postestimation command allows to post-estimate spatial fixed or random

DGP - Fixed eects SDM

where the nuisance parameters i (i = 1, . . . , n) are drawn from an iid standard

xkit = 0.4i + (1 0.42 )1/2 zkit ,

.. *** load a dta dataset containing the spatial contiguity matrix

. *** get an spmat objects from dta

. *** Summarize the spmat obj

Summary of spatial-weighting object W

. ** Fixed-effects Durbin model (correctly specified, row normalized W)

SDM with spatial fixed-effects Number of obs = 940

Group variable: id Number of groups = 188

R-sq: within = 0.5727

Mean of fixed-effects = -0.0137

In a spatial setting, the eect of an explanatory variable change in a particular

If we have only 2 units and 1 regressor:

** Fixed-effects Durbin model (correctly specified, row normalized W)

SDM with random-effects Number of obs = 940

Group variable: id Number of groups = 188

R-sq: within = 0.5666

. estimates store sdm_re

hausman sdm_fe sdm_re, eq(1:1 2:2 3:3)

---- Coefficients ----

Test: Ho: difference in coefficients not systematic

. ** Fixed-effects Durbin model (correctly specified, row normalized W)

Group variable: id Number of groups = 188

R-sq: within = 0.5727

Mean of fixed-effects = -0.0137

. predict yhat1, naive

. sum alpha alphahat

Variable | Obs Mean Std. Dev. Min Max

and test the following constraints

** Test for SAR

(1) [Wx]x1 = -[Spatial]rho*[Main]x1

SAC with spatial fixed-effects Number of obs = 940

Group variable: id Number of groups = 188

R-sq: within = 0.2652

Mean of fixed-effects = -0.0117

SDM vs SAC Testing

. qui xsmle y x1 x2 x3, wmat(W) model(sdm) fe type(ind) noeff nolog

. qui xsmle y x1 x2 x3, wmat(W) emat(W) model(sac) fe type(ind) noeff nolog

Lee, L. F. and Yu, J. (2010). Estimation of spatial autoregressive panel data

You might also like