10.

2478/v10103-012-0040-8

GRAYNA TRZPIOT*

Spatial Quantile Regression

Abstract
In a number of applications, a crucial problem consists in describing and
analyzing the influence of a vector Xi of covariates on some real-valued
response variable Yi. In the present context, where the observations are made
over a collection of sites, this study is more difficult, due to the complexity of the
possible spatial dependence among the various sites. In this paper, instead of
spatial mean regression, we thus consider the spatial quantile regression
functions. Quantile regression has been considered in a spatial context. The
main aim of this paper is to incorporate quantile regression and spatial
econometric modeling. Substantial variation exists across quantiles, suggesting
that ordinary regression is insufficient on its own. Quantile estimates of
a spatial-lag model show considerable spatial dependence in the different parts
of the distribution.

1. Introduction

1.1 Linear Regression - introduction

Linear regression is the standard tool for many empirical studies. When
the relationship between a dependent variable, y, and a set of explanatory
variables, X, can be written as y = X + u , a simple ordinary least squares
*

Ph.D., Professor at the University of Economics in Katowice

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

266

Grayna Trzpiot

(OLS) regression of y on X can provide unbiased estimates of the parameters, ,

and a predicted value, y = X .
This heavy reliance on linear regression models has been carried over to
the analysis of spatial data. The most commonly used spatial model adds
a weighted average of nearby values for the dependent variable to the list of
explanatory variables:

y = WY + X + u .

(1)

In this model, W is a spatial weight matrix that specifies the

relationships between observations. W is a spatial weight matrix with rows
that sum to one and zeros on the diagonals, and is a parameter measuring the
strength of the relationship.
The model can be useful when X does not fully account for the tendency
for the dependent variable to be highly correlated over space, so that nearby
values of Y provide significant explanatory power. The endogeneity of WY
poses challenges for estimation. Most empirical applications are based on
maximum likelihood estimation of the model under the assumption of normally
distributed errors. Other approaches are based on instrumental variables (IV)
estimation, usually with spatially lagged values of X (such as WX and WWX) as
instruments for WY. Several researchers have used the spatial AR model as the
basis for quantile regressions in which both and are allowed to vary across
quantiles.
Typical specifications of the spatial weight matrix are based on first-order
contiguity when the data are drawn from geographic units such as counties or
census tracts. Though the approach is used less commonly for point data, typical
specifications are similar in that the spatial weights are assumed to decline
rapidly with distance between observations.
Predicted values are then based on
(1) the structural model

Y ( ) = WY + X
(2) the reduced form

Y ( ) = ( I W ) 1 X
(3) or a decomposition into trend andsignal components

Y ( ) = X + W ( I W ) 1 X
Spatial effects generally appear as noise around a spatial trend that looks
much like the predicted values from an OLS regression of Y on X. The objective

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

Spatial Quantile Regression

267

of a regression analysis is to estimate the coefficients, and , and to obtain

predictions of y at given values of X.
Regression analysis is not well suited to explaining the distribution of
a variable. When the predicted values from a regression are y = X , then the
distribution of the predicted values simply mimics the distribution of the
variables in X. The implied effect of a change in one of the explanatory
variables is to cause a parallel shift of y by an amount determined by the
variables estimated coefficient. Though a parallel shift may be reasonable in
some cases, it is limitation that a research may not want to impose beforehand.

1.2. Quantile Regression

Quantile regression is a method for estimating functional relations
between variables for all portions of probability distribution. Typically
a response variable Y is some function of predictor variable X. Regression
application focus in estimating rates of changes in the mean of the response
variable distribution as some function of a set of predictor variables. In the other
words the function is defined for the expected value of Y conditional X, E(Y|X).
Regression analysis gave incomplete picture of the relationships between
variables especially for regression models with heterogeneous variances.
Quantile regression was developed as an extension of the linear model for
estimating rate of change in all parts of the distribution of response variables.
The estimates are semi parametric in the sense that no parametric distributional
form (eg. normal, Poisson, negative binominal, etc.) is assumed for the random
error part of the model , although a parametric form is assumed for the
deterministic portion of the model (eg. 0X0 + 1X1). The conditional quantiles
denoted by Qy(|X) are the inverse of the conditional cumulative distribution
function of the response variable Fy1 ( X) , where [0, 1] denotes quantile
rank.
The quantile model posits the th quantile of Y conditional on x to be,

Q( x ) = () + x() , 0 < < 1.

(2)

If () is a constant , the model reduces to the standard conditional

expectation model, E(Yx) = +x, with constant variance errors. When ()
depends on , the model allows the distribution of Y to depend on x in different
ways at different parts of the distribution. The traditional linear model can be
viewed as a summary of all the quantile effects; that is, Q ( x) d = E (Y x) .

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

268

Grayna Trzpiot

Under this interpretation, traditional analysis loses information due to its

aggregation of possibly disparate quantile effects. Many different quantile paths,
for example, can lead to k = 0. On the one hand, k = 0 can mean xk does not
matter does not affect the distribution of Y. But it can also mean there are
important but compensating quantile effects relating Y and x.
Quantile regression is much better suited to analyzing questions involving
changes in the distribution of a dependent variable. Quantile regressions allow
for separate effects of an explanatory variable on different points of the
dependent variable distribution. The coefficient estimates are then frequently
interpreted as being analogous to standard linear regression estimates, albeit for
different points in the distribution of the dependent variable (Trzpiot 2008, 2009
a, b, c, 2010, 2011 a, b).
It is less commonly recognized that quantile regression can produce
estimates of changes in the full distribution of the dependent variable when the
values of the explanatory variables change. The set of coefficients produced for
independent variables imply a change in the full distribution dependent
variables.
Special issues do not necessarily arise when estimating quantile
regressions using spatial data. Several researchers have proposed variants of the
spatial autoregressive (AR) model, y = WY + X + u , for quantile analysis.
These procedures treat WY as just another endogenous explanatory
variable. The spatial AR model may not necessarily be the best choice for spatial
modeling, particularly for large data sets comprising individual geographic
points rather than large zones or tracts. In situations where the distribution of the
dependent variable changes smoothly over space, a nonparametric procedure
may be a much better approach.

2. Distribution of the Dependent Variable

In general, the conditional quantile function for y given a set of variables
X can be written:

Qy ( X ) = X ( X ) , 0 < < 1.

(3)

Usually, we have limited our attention to a small number of values for the
quantile, . Focusing on that values provides useful information about the
distribution of the dependent variable given values of X, but it certainly does not
provide a complete picture of the full distribution of y.

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

Spatial Quantile Regression

269

One way to use quantile regression estimates to simulate the distribution

of the dependent variable is to draw randomly from possible values of and then
estimate a separate quantile regression for each value of . For example, we
might draw 1000 values of from a uniform distribution ranging from 0 to 1,
i.e., ~ U(0, 1). If we let J represent the number of draws from the U(0,1)
distribution, then we have:

Q y ( j X ) = X ( j X ) j = 1,, J

(4)

With J estimates of the conditional quantile in hand, a standard kernel

density function can be applied to X ( j X ) to estimate the density function
for the dependent variable. Since quantile estimates are generally fairly smooth
across, drawing multiple values of from a U(0,1) distribution is a very
inefficient way of constructing the density function. Using a limited range of
value for is more efficient. For example, we might restrict the estimates to , ,
or a still more limited set of values for that provides good coverage of the set of
permissible values for . Since quantile estimates are likely to have very high
variances at extreme values of such as 0,01 or 0,99, it generally is a good idea
to trim the extreme observations if a grid of values is used for .

3. The Effect of a Discrete Change in an Explanatory Variable

Quantile regression estimates can have interesting implications for the
distribution of y values even in a model with a single explanatory variable.
Consider a model with k explanatory variable in addition to the intercept. After
estimating quantile regressions for J quantiles, the predicted values for quantile
are simply j:

Q y ( j X ) = 0 ( j ) + 1 ( j ) x1 + .... + k ( j ) xk ,

j = 1,,J

(5)

When simplified the notation by replacing ( j X ) with ( j ) , but it

should be clear that the estimates depend on the observed values of X. Even in
the single-explanatory case where k = 1, the implied effect of changing from to
produces J separate values for

Q y ( j X , x1 = 0 ) = 0 ( j ) + 1 ( j ) 0 + .... + k ( j ) xk , j = 1,,J

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

(6)

270

Grayna Trzpiot

Q y ( j X , x1 = 0 ) = 0 ( j ) + 1 ( j )1 + .... + k ( j ) xk , j = 1,,J

(7)

With J quantiles and n observations, equation (6) and (7) imply nJ values
for the conditional quantile functions. Since 1 ( j ) is not constant, the
conditional quantile functions imply a full distribution of values for y even when
x1 is the only variable in the model.

4. Quantile version of the Spatial AR Model

The analysis up to this point has not been explicitly spatial. Although the
explanatory variables might include measures of access to various amenities
such as a citys central business districts, parks, or lakes, nothing yet is unique to
the analysis of spatial data. Several attempts have been made to adapt the
standard spatial autoregressive (AR) model for quantile regression.
The spatial AR model adds a weighted average of nearby values of the
dependent variable to the list of explanatory variables. The model is written
y = WY + X + u , where X is the nxk matrix of explanatory variables, Y is
the dependent variable, and W is an nxn matrix specifying the spatial
relationship between each value of Y and its neighbors.
Suppose the observations represent census tracts. If each tract is
contiguous to four other tracts, then Wij = for each of the four tracts that is
contiguous to observation i, and Wij = 0 for all other values of j. In this example,
each of the n elements of WY is simply the average, for each observation, of the
four neighboring values of Y. More generally, if observation i is contiguous to
other tracts, then Wij = 1/ni for the tracts that are contiguous to observation i, and
Wij = 0 otherwise. For point data, WY might form a weighted average of the
nearest K neighbors, or the weights might decline with distance. WY is clearly
an endogenous variable. Indeed, one interpretation of WY is that it is the set of
predicted values from kernel regressions of Y on the set of geographic
coordinates. For example, suppose we were to write yi = f (loi , lai ) + ui . If we
use a rectangular kernel with a very small window size e.g., the four closest
observations then the cross-validation version of the kernel regression
estimator is

y i =

1
ni

I
j =1

yj ,

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

(8)

Spatial Quantile Regression

271

where I j indicates that observation j is one of the nearest neighbors to

observation i, and indicates the number of observations that are being given
weight when constructing the estimate for observation i. Not surprising, adding
the predicted value of Y as an explanatory variable for Y often produces highly
significant results.
Although the spatial lag variable, WY, is formally equivalent to a kernel
regression, the approaches could hardly be more different in spirit.
The spatial AR model is based on an assumption that the researcher can
truly specify the full spatial relationship between all of the observations. After
specifying the entire path by which each of n observations can influence all of
the other observations, all that is left is to determine the strength of the
relationship by estimating .
In contrast, nonparametric and semi-parametric regressions involve far
less difficult. We could easily write the model in semiparametric form as

yi = f (loi , lai ) + X i + ui

(9)

or in the conditionally parametric form

yi = X i (loi , lai ) + ui .

(10)

The spatial AR model is based on the assumption that the researcher can
specify a simple parametric function that accounts for both the relationship
between X and Y and the entire spatial relationship between all observations.
Nonparametric approaches are based on an assumption that the researcher can
correctly specify the variables that influence Y, but they allow for local variation
in the marginal effect of X on Y. The spatial AR model and its variants may be
useful in situations where the objective is to estimate a causal relationship
between Y and neighboring values of the dependent variable.

4.1 Quantile Regression with an Endogenous Explanatory Variable

The spatial AR model is most commonly estimated by maximizing the
log-likelihood function that is implied under the assumption of normally
distributed errors. An alternative approach based on the generalized method of
moments method allows the model to be estimated using a variant of two-stage
least squares (2SLS). In the first stage, the endogenous variable, WY, is
regressed on a set of instruments.

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

272

Grayna Trzpiot

The predicted value of WY is then used as an explanatory variable in the

second stage regression. Most researchers use X, WX, and, sometimes,
additional orders of the spatial lags such as WWX as instruments for WY.
Though this method can work well when the goal is to estimate a standard
regression, quantile regression may be more complex because instrumental
variables are needed for WY when estimating a regression for each quantile, .
Two methods have been used to form the instrumental variables needed
these quantile regressions. The simpler version was proposed by Kim and Muller
(2004). Their approach is a straightforward extension of 2SLS.
For each value of , they first estimate a quantile regression for WY using
the set of instruments (e.g., X and WX) as explanatory
variables. The predicted

values from the quantile regression are WY ( ) . In the second stage, they
estimate another quantile regression for the same value of , this time with Y as
the dependent variable and X and WY ( ) as the explanatory variables.
Only 10 quantile regressions are needed to estimate the model for 5
quantiles (e.g., = 0,10, 0,25, 0,50, 0,75, 0,90). Zietz et al. (2008) and Liao and
Wang (2012) use this approach to estimate quantile versions of the spatial AR
model. They use bootstrap procedures to construct standard error estimates.
Though somewhat more complicated, the Chernozhukov and Hansen
(2006) approach may be more robust than the Kim and Muller (2004) approach
because it does not require that the same quantile be used in both stages of the
procedure.
An additional advantage is that Chernozhukov and Hansen present
a covariance matrix estimate that is easy to construct.

In the version describe here, the predicted values, WY , from an OLS

regression of WY on the instruments are used as the instrumental variable for
WY. This instrumental variable is then used as an explanatory
variable for

a series of quantile regressions Y WY of on X and WY The same quantile,

, is used for each of the regressions, while a grid of alternative values is used

for . The estimated value of is the value that produces the coefficient on WY
that is closest to zero. After finding the estimated values of are calculated
by a quantile regression of Y WY on X. The motivation behind this
estimator is a property of two-stage
least squares: when instruments are chosen

optimally, the coefficient on WY will be zero when both the actual variable,
WY, and the instrumental variable are included in a regression.
Standard error estimates are easy to construct for the Chernozhukov and
Hansen method. Let e represent the residuals from the quantile regression
of Y WY on X, and define I ( ei < h) /(2h) , where h is a constant
bandwidth.

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

Spatial Quantile Regression

273

Define i = f i WY i and Z i = f i X i . Then the covariance matrix for

= ( , ) is:

V () = J ( )1 S ( ) J ( ) 1

(11)

where

WY
J ( ) =
Z WY

W
Y
WY
W
Y X
S ( ) = (1 )
X WY
X X

X
and
Z X

As is the case for any instrumental variables (IV) estimator, the estimates
from either approach can be sensitive to the choice of instruments. However, an
important advantage of the IV approach over maximum likelihood estimation,
which is commonly used for the non-quantile version of the spatial AR model, is
that that there is no need to invert the nxn matrix ( I W ) 1 when estimating
the model.
It may still prove necessary to invert large matrices when constructing
predicted values for Y. Let Y ( ) denote the set of predicted values of the
dependent variable for quantile .
Three procedures are often used to construct Y ( ) :
quantile version of the structural model

Y ( ) = ( )WY + X ( )

(12)

quantile version of the reduced form

Y ( ) = ( I ( )W ) 1 X ( )

(13)

quantile version of a decomposition into trend andsignal components

Y ( ) = X ( ) + ( )W ( I ( )W ) 1 X ( )

(14)

Though the first procedure may be viewed as cheating because it uses

actual values of WY to predict Y, it is commonly used for standard linear
simultaneous equations models.
The second version follows directly from the original model specification:
the equation implies Y = ( I W ) 1 ( X + u ) , from which equation (13)
follows by setting u = 0. Finally, equation (14) is derived by noting that equation
(13) also provides a way to estimate WY for the expression given in equation (12):

WY ( ) = W ( I ( )W ) 1 X ( )

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

(15)

274

Grayna Trzpiot

5 Conditionally Parametric Quantile Regression

Nonparametric approaches can be adapted to quantile regression models.
In the case of a single explanatory variable, x, all that is necessary to make the
model nonparametric is to add a kernel weight function K((x-xt)/h) when
estimating a quantile regression for a target point xt. After estimating the
function for a series of target points, the estimates can then be interpolated to all
values of x. The nonparametric approach is a flexible way to add nonlinearity to
the estimated quantile regressions. Although nonparametric approaches can
potentially be applied to variables with many explanatory variables, the variance
of the estimated can become very high when there are more than two or three
explanatory variables.
The problem can be simplified significantly in the case of spatial data sets.
It is usually reasonable to assume that a simple linear model fits the data well in
small geographic areas. The difficulty is in specifying a global parametric model
that fits the data well across a large region. The spatial AR model allows for
local variation around a global parametric trend. In contrast, a conditionally
parametric approach allows for flexible trends by allowing the coefficients to
vary smoothly over space. Nonparametric models can be hard to summarize
because they produce separate estimates for every observation. As we have seen,
this is not a problem for quantile regression models because the easiest way to
interpret the results is to present sets of density estimates showing how the
distribution of the dependent variable changes as the values of individual
explanatory variable change. As a result, it is just as easy to estimate
nonparametric quantile regressions as it is to use a linear approach.

5.1 CPAR Quantile Regression for Spatial Data

Consider the conditional quantile function Qy ( X ) = X ( X ) , in
which the dependent variable is a linear function of a set of explanatory
variables, X. Now suppose that we want to allow the coefficients to vary over
space. Using lo and la to represent the geographic coordinates, we can write the
conditional quantile function as:

Qy ( X , lo, la) = X ( X , lo, la )

(16)

Although it is possible to include lo and la as explanatory variables in

which case they are part of the X matrix the more common approach is to keep

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

Spatial Quantile Regression

275

them separate. Also, note that lo and la can represent any geographic coordinate
system rather than just longitude and latitude.
All that is necessary to estimate a nonparametric version of equation (16)
is to specify a kernel weight function that indicates the weight given an
observation with coordinates (lo, la) when estimate the function at a target point
(lot, lat). One approach is to use a simple product kernel:

K ((lo lot ) / h1, (la lat ) / h2 )

(17)

A more commonly used alternative is to make the weights depend simply

on the straight-line distance between each observation and the target point, dt:

K ( d t / h)

(18)

The kernel weight function in equation (19) draws a circle around the
target point to form the weights. Although, equation 18 is slightly more general,
there is little difference between the two in practice. With J different quantiles,
the set of estimated coefficients for explanatory variable k, k , is an nxJ matrix.
With K explanatory variables in addition to the intercept, the nxJ matrix of
quantile predictions is

y = 0 + xk k
K

(20)

k =1

These predictions can be used to calculate density functions for predicted

values of the dependent variable for arbitrary values of the explanatory
variables. Suppose we want to evaluate the model at x1 = . Then the nxJ set of
predicted values is simply

y = 0 + x1 + xk k .
K

(21)

k =2

The calculations can be repeated for other values of and for other
explanatory variables. The results can then be summarized using estimated
kernel density functions.
Model is analogous to conditionally parametric (CPAR) local linear
regression. The estimation procedure involves estimating separate quantile
regressions for various target points, with more weight placed on observations
that are close to the target. Unlike a fully nonparametric approach, the CPAR

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

276

Grayna Trzpiot

approach produces coefficient estimates for the explanatory variables. But unlike
the spatial AR version of quantile regression, the estimated coefficients vary
over space. The CPAR approach is less sensitive to model misspecification than
the fully parametric spatial AR approach, and it accounts for local variation in an
overall spatial trend. The approach is well suited for quantile analysis in
situations where the distribution of the dependent variable is, for example,
highly skewed in some locations, tightly clustered in others, while all the time
varying smoothly over space. Moreover, the CPAR approach does not require
the specification of a large (n x n) spatial weight matrix, making it amenable to
large data sets.
We present results (Chambers, Pratesi, Salvati, Tzavidis 2005) obtained
for the estimation spatial distribution of the mean and median production of
olives per farm LES. The data are from Farm Structure Survey (2003). Z the
incidence matrix of dimensions 2508 farm per 42 LESs. The neighborhood
structure W is defined as follows: spatial weight wij, is 1 if area shares an edge
with j and 0 otherwise.
The median map is intensive to the presence a few big farms that raise the
medium level of production as a consequence the spatial distribution of the
median is more homogenous.
Figure 1 a) Mean b) median production of olives

Source: Chambers R. & all (2005).

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

Spatial Quantile Regression

277

6. Conclusions
In summary we can say that the classic paper for quantile regression is
Koenker and Bassett (1978). Koenker (2005) presents an extensive examination
of the econometric theory related to a wide variety of quantile models.
Buschinsky (1998) helped popularize the use of quantile regression analysis on
the distribution of wages. The spatial AR version of the quantile model relies on
approaches developed by Chernozhukov and Hansen (2006) and Kim and
Muller (2004). The approaches have been applied to studies of house prices by
Kostov (2009), Liao and Wang (2012) and Zeitz et al. (2008). The studies rely
on the IV approach for estimating the spatial AR model. Nonparametric versions
of quantile models relies heavily on Koenker work. Splines are also a potential
alternative to kernel smoothing; it was done in Koenker and Mizera (2004). The
use of nonparametric methods for spatial models has been forced by the
invention of new terms by geographers for procedures that have already been
used extensively in statistics and economics.

References
Buchinsky M. (1998), Recent Advances in Quantile Regression Models: A Practical Guideline for
Empirical Research, Journal of Human Resources, 33
Chambers R., Pratesi M., Salvati N., Tzavidis N. (2005), Small Area Estimation: spatial
information and M- quantile regression to estimate the average production of olive per farm, Wp
279, Dipatimento di Statistica e Matematica Applicata allEconomia, Universita di Pisa
Chernozhukov V., Hansen Ch.. (2006), Instrumental Quantile Regression Inference for Structural
and Treatment Effect Models, Journal of Econometrics, 132
Cleveland, W. S. (1994), Coplots, Nonparametric Regression, and Conditionally Parametric Fits.
In T.W. Anderson, K.T. Fant, and I. Olkin (Eds.), Multivariate Analysis and its Applications,
Hayward: Institute of Mathematical Statistics.
Kim T.-H., Muller Ch.. (2004), Two-Stage Quantile Regression when the First Stage is Based on
Quantile Regression, Econometrics Journal, 7
Koenker R., Basset B., (1978), Regression Quantiles, Econometrica, Vol 46
Koenker R., Ng P., (2005), Inequality Constrained Quantile Regression, The Indian Journal of
Statistics, Vol. 67
Koenker R., Hallock K. F. (2001), Quantile Regression, Journal of Economic Perspectives, 15
Koenker R., Mizera I. (2004). Penalized Triograms: Total Variation Regularization for Bivariate
Smoothing, Journal of the Royal Statistical Society: Series B, 66

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

278

Grayna Trzpiot

Kostov, P. (2009), A Spatial Quantile Regression Hedonic Model of Agricultural Land Prices,
Spatial Economic Analysis, 4
Liao, Wen-Chi and Xizhu Wang (2012), Hedonic House Prices and Spatial Quantile Regression,
Journal of Housing Economics, 21, 16-27
Mcmillan. D. P. (2010), Issues in Spatial Data Analysis, Journal of Regional Science, 119-141.
Trzpiot G., (2008), The Implementation of Quantile Regression Methodology in VaR Estimation
Studies and Researches of Faculty of Economics and Management University of Szczecin,
Trzpiot G., (2009 a), Quantile Regression Model versus Factor Model Estimation, in: Financial
Investments and Insurances, University of Economics in Wrocaw, Vol 60
Trzpiot G., (2009 b), Application weighted VaR in capital allocation, Polish Journal of
Environmental Studies, Olsztyn, Vol 18, 5B
Trzpiot G., (2009 c), Estimation methods for quantile regression, Economics Studies 53, Karol
Adamiecki University of Economics in Katowice
Trzpiot G., (2010), Quantile Regression Model of Return Rate Relation Volatility for Some
Warsaw Stock Exchange Indexes, (in Polish), Finances, Financial Markets and Insurances. Capital
Market, University of Szczecin, Vol 28, 61-76
Trzpiot G. (2011a). Bayesian Quantile Regression, Studia Ekonomiczne, Zeszyty Naukowe nr 65,
Uniwersytet Ekonomiczny w Katowicach, 33-44
Trzpiot G. (2011b). Some tests for quantile regression models, Acta Universitatis Lodziensis,
d, Folia Economica, 255, 125-135
Zeitz J., Zietz E. N., Sirmans G. S. (2008) Determinants of House Prices: A Quantile Regression
Approach, Journal of Real Estate Finance and Economics, 37

Streszczenie
PRZESTRZENNA REGRESJA KWANTYLOWA

W wielu zastosowaniach, podstawowym problemem jest opis i analiza wpywu

wektora skorelowanych zmiennych objaniajcych X na zmienna objanian Y.
W przypadku, gdy obserwacje badanych zmiennych s dodatkowo rozmieszczone
przestrzennie, zadanie jest jeszcze trudniejsze, poniewa mamy dodatkowe zalenoci,
wynikajce ze zmiennoci przestrzennej.
W tej pracy, w miejsce przestrzennej regresji wykorzystujcej redni,
rozpatrzymy przestrzenna regresj kwantylow. Regresja kwantylowa zostanie
omwiona w przestrzennym kontekcie. Gwnym celem pracy jest wskazanie na

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM

Spatial Quantile Regression

279

moliwoci powizania metodologii regresji kwantylowej i ekonometrycznego

modelowania przestrzennego. Dodatkowe zasoby informacji o zmiennoci otrzymujemy
badajc kwantyle, wychodzc poza tradycyjny opis klasycznej regresji. Estymacja
kwantylowa w modelu przestrzennym uwydatnia zalenoci przestrzenne dla rnych
fragmentw rozwaanych rozkadw.

Brought to you by | CAPES

Authenticated
Download Date | 12/9/14 1:31 AM