Working Paper 09-15 Departamento de Estadística

Statistics and Econometrics Series 05 Universidad Carlos III de Madrid

March 2009 Calle Madrid, 126
28903 Getafe (Spain)
Fax (34) 91 624-98-49

SMALL AREA ESTIMATION OF POVERTY INDICATORS

Molina, Isabel1 and Rao, J. N. K.2

Abstract

We propose to estimate non-linear small area population quantities by using Empirical

Best (EB) estimators based on a nested error model. EB estimators are obtained by
Monte Carlo approximation. We focus on poverty indicators as particular non-linear
quantities of interest, but the proposed methodology is applicable to general non-linear
quantities. Small sample properties of EB estimators are analyzed by model-based and
design-based simulation studies. Results show large reductions in mean squared error
relative to direct estimators and estimators obtained by simulated censuses. An
application is also given to estimate poverty incidences and poverty gaps in Spanish
provinces by sex with mean squared errors estimated by parametric bootstrap. In the
Spanish data, results show a significant reduction in coefficient of variation of the
proposed EB estimators over direct estimators for most domains.

Keywords: Empirical best estimator; Parametric bootstrap; Poverty mapping; Small

area estimation.

1
Departamento de Estadística, Universidad Carlos III de Madrid, C/ Madrid 126, 28903 Getafe (Madrid),
Spain. E-mail: isabel.molina@uc3m.es
2
School of Mathematics and Statistics, Carleton University, 125 Colonel By Drive, Ottawa, ON CANADA
K1S 5B6. E-mail: jrao@math.carleton.ca
 Small Area Estimation of Poverty Indicators∗
Isabel Molina†and J. N. K. Rao‡

Abstract
We propose to estimate non-linear small area population quantities
by using Empirical Best (EB) estimators based on a nested error model.
EB estimators are obtained by Monte Carlo approximation. We focus
on poverty indicators as particular non-linear quantities of interest, but
the proposed methodology is applicable to general non-linear quantities.
Small sample properties of EB estimators are analyzed by model-based
and design-based simulation studies. Results show large reductions in
mean squared error relative to direct estimators and estimators obtained
by simulated censuses. An application is also given to estimate poverty
incidences and poverty gaps in Spanish provinces by sex with mean squared
errors estimated by parametric bootstrap. In the Spanish data, results
show a significant reduction in coefficient of variation of the proposed EB
estimators over direct estimators for most domains.

Key words: Empirical best estimator; Parametric bootstrap; Poverty

mapping; Small area estimation.

1 Introduction
The first of the Millennium Development Goals established by the United Na-
tions is the eradication of extreme poverty and hunger. The availability of the
most possible accurate information concerning the living conditions of people is
a basic instrument for targeting policies and programs aiming at the reduction
of poverty. Thus, there are no doubts about the importance of the estimation of
poverty measures. In many cases the information collected from national surveys
is limited and allows estimation only for larger regions or larger population sub-
groups. Therefore, small area estimation techniques are required that “borrow
strength” across areas through linking models and auxiliary information such as
∗
Supported by the Spanish grants MTM2006-05693 and SEJ2007-64500, and by the Euro-
pean project num. 217565-FP7-SSH-2007-1.
†
Department of Statistics, Universidad Carlos III de Madrid. Address: C/Madrid 126,
28903 Getafe (Madrid), Spain, Tf: +34 916249887, Fax: +34 916249849, E-mail: is-
abel.molina@uc3m.es
‡
School of Mathematics and Statistics, Carleton University. E-mail: jrao@math.carleton.ca

1
censuses and administrative data; see Rao (2003) for a comprehensive account
of these techniques.
Many measures of poverty and inequality are non-linear functions of a welfare
variable for the population units. This makes many of the current small area
estimation methods, typically developed for the estimation of linear characteris-
tics, such as means, not applicable. Here we propose the use of empirical best
predictors (EBPs) obtained through Monte Carlo approximation. This method
provides estimators that are “best” in the sense of minimizing the mean squared
error (MSE) under assumed small area models. It is useful for the estimation
of any function of a welfare variable, when this variable or some transforma-
tion of it follows a linear model. We show by simulations that empirical best
predictors behave good in terms of bias and mean squared error. We also pro-
pose a parametric bootstrap method for MSE estimation and study its bias in
simulations.
In the U.S. the need for small area poverty estimates has given rise to the
SAIPE program (Small Area Income & Poverty Estimates) of the U.S. Census
Bureau; for further details see http://www.census.gov/hhes/www/saipe. The
main objective of this program is to provide updated estimates of income and
poverty statistics for the administration of federal programs and the allocation of
federal funds to local jurisdictions. The county level methodology, summarized
by Bell (1997), basically uses a Fay-Herriot area level model (Fay & Herriot, 1979)
to produce model-based county estimates of school-age children under poverty.
The World Bank (WB) has been releasing small area poverty and income
inequality estimates for some countries, using the methodology of Elbers et al.
(2003). This methodology is currently widely used, see e.g. Neri et al. (2005),
Ballini et al. (2006), Tarozzi and Deaton (2007) and Haslett and Jones (2005).
Elbers et al. (2003) assumed a unit level model that combines both census and
survey data. Using that model, they produce disaggregated maps that describe
the spatial distribution of poverty and inequality. Section 7 shows that this
method, when used to estimate small area means, is basically equal to a synthetic
regression estimator.
Measures of inequality include the Gini coefficient, the Sen index, the general
entropy and the Theil index (see e.g. Neri et al., 2005). Although the method
developed in this paper allows the estimation of these inequality measures as
well, for the sake of brevity we will focus on the estimation of a class of poverty
measures called FGT poverty measures due to Foster et al (1984), see Section 2,
and used in the WB method.
A common definition of poverty classifies a person as “under poverty” when
the selected welfare variable for this person is below a threshold called poverty
line, which is usually a given percentage of the median welfare for the population.
For instance, the Spanish National Statistical Institute (INE) established the
poverty line at 60% of the median per capita income. Here we assume that the
poverty line is fixed at some quantity established by the corresponding authority
and we avoid discussions concerning the definition of this threshold.

2
 The paper is organized as follows. Section 2 defines the family of FGT poverty
measures and Section 3 introduces two basic types of direct estimators, which
make use only of the sample data from the target area. Section 4 describes the
best prediction methodology for finite populations and Section 5 applies this
methodology to the estimation of FGT poverty measures through the use of a
nested error linear regression model. Section 6 describes a parametric bootstrap
method for MSE estimation. Section 7 makes a theoretical comparison of the
different methods in the context of estimating small area means. Sections 8 and
9 present the results of simulation studies on the performance of the proposed
method relative to the WB method and direct estimation, in terms of bias and
MSE. Performance of bootstrap MSE estimator is also studied. Finally, in Sec-
tion 10, the proposed method is applied to Spanish data to estimate poverty
incidences and poverty gaps in Spanish provinces by gender.

2 FGT poverty measures for small areas

Consider a finite population of size N that is partitioned into D small areas
of sizes N1 , . . . , ND . Let Edj be a suitable quantitative measure of welfare for
individual j in small area d, such as income or expenditure, and let z be the given
poverty line; that is, the threshold for Edj under which a person is considered
as “under poverty”. The family of poverty measures introduced by Foster et al.
(1984) and called throughout the paper as FGT poverty measures, for each small
area d, is defined as
Nd  α
1 X z − Edj
Fαd = I(Edj < z), α = 0, 1, 2, d = 1, . . . , D,
Nd j=1 z

where I(Edj < z) = 1 if Edj < z (person under poverty) and I(Edj < z) = 0
if Edj ≥ z (person not under poverty). For α = 0 we get the proportion of
individuals under poverty in small area d, also called poverty incidence or head
count ratio. The FGT measure for α = 1 is called poverty gap, and measures
the area mean of the relative distance to non-poverty (the poverty gap) of each
individual. For α = 2 the measure is called poverty severity. This measure
squares the poverty gaps, and thus emphasizes extreme poverty.

3 Direct estimators of poverty measures

In the inference process, a random sample of size n < N is drawn from the
population according to a specified sampling design. Let Ω denote the set of
indexes of the population units. Let s be the set of units selected in the sample
and r the set of indexes of the units that are not selected (with size N − n).
The restrictions of Ω, s, N and n to area d are denoted by Ωd , sd , Nd and nd
respectively, where n = n1 + · · · + nD . The unweighted sample FGT poverty

3
measures are given by
 α
1 X z − Edj
fαd = I(Edj < z), α = 0, 1, 2, d = 1, . . . , D. (1)
nd j∈s z
d

A direct estimator for a small area, as a sample estimator, uses only the
sample data from the target small area and it is usually design-based. Let wdj
be the sampling weight (inverse of the probability of inclusion) of individual j
from area d. Direct estimators of the FGT measures are given by
 α
w 1 X z − Edj
fαd = wdj I(Edj < z), α = 0, 1, 2, d = 1, . . . , D, (2)
N̂d j∈s z
d

P
where N̂d = j∈sd wdj is the direct estimator of the population size of small area
d, Nd . If the sampling weights wdj do not depend on the unit j, for example
wdj = nd /Nd under simple random sampling within areas, then (2) reduces to
the unweighted mean (1).
The limited sample sizes nd within some of the areas prevent the use of es-
timators such as (1) or (2). Indeed, a common definition of poverty classifies
a person as “under poverty” when the selected welfare variable for this person
is below 60% of the median. Under this definition the outcome of being un-
der poverty is likely to have low frequency. Then, to obtain reliable estimators
for small domains or geographical areas it is necessary to appeal to small area
techniques (Rao, 2003). These techniques improve the estimation procedures by
using models that establish some relationships between the areas, based on aux-
iliary information (census and/or administrative variables) related to the welfare
variables of interest. These models provide “indirect” estimators that make use
of related data from other areas, and which might reduce drastically the esti-
mation errors as long as model assumptions hold. Model checking should be an
integral part of indirect estimation methods.

4 Best prediction under a finite population

Section 4 introduces the best predictor (BP) of a function of a random vector
in a finite population. Application of the BP methodology for estimating FGT
poverty measures in small areas is described in Section 5.
Consider a random vector y = (Y1 , . . . , YN )′ containing the values of a random
variable associated with N units of a finite population. Let ys be the sub-vector
of y corresponding to sample elements and yr the sub-vector of out-of-sample
elements; that is, y = (ys′ , yr′ )′ . The target is to predict the value of a real-
valued function δ = h(y) of the random vector y using the sample data ys . For
a particular predictor δ̂, the mean squared error is defined as

MSE(δ̂) = Ey {(δ̂ − δ)2 }, (3)

4
where Ey denotes expectation with respect to the joint distribution of the pop-
ulation vector y. The BP of δ is the function of ys that minimizes (3). Consider
the conditional expectation δ 0 = Eyr (δ|ys ), where the expectation is taken with
respect to the conditional distribution of yr given ys and the result is a function
of sample data ys . Subtracting and adding δ 0 in the mean squared error, we
obtain

MSE(δ̂) = Ey {(δ̂ − δ 0 )2 } + 2 Ey {(δ̂ − δ 0 )(δ 0 − δ)} + Ey {δ 0 − δ)2 }.

In this expression, the last term does not depend on δ̂. For the second term,
observe that
h n oi
Ey {(δ̂ − δ 0 )(δ 0 − δ)} = Eys Eyr (δ̂ − δ 0 )(δ 0 − δ)|ys
h i
0
 0
= Eys (δ̂ − δ ) δ − Eyr (δ|ys ) = 0.

Thus, the BP is the value δ̂ that minimizes Ey {(δ̂ − δ 0 )2 }. Since this quantity is
non-negative and its minimum value is zero, the BP is

δ̂ B = δ 0 = Eyr (δ|ys ). (4)

Typically, the joint distribution of y depends on some unknown model pa-

rameters. Then an empirical BP (EBP) of δ can be obtained by replacing all
unknown parameters by suitable estimators and then evaluating the expectation
(4) from the estimated distribution.

Remark 1. Suppose that y follows a Normal distribution with mean vector µ =

Xβ, for a known matrix X, and positive definite covariance matrix V. Let the
target quantity δ be a linear function of y, δ = a′ y. In this case, the best linear
unbiased predictor (BLUP) of δ, obtained by Royall (1976), is equal to the BP
when the true β is replaced by the generalized least squares estimator.

5 Empirical best prediction of FGT poverty mea-

sures
In this section we describe how to obtain BPs of FGT poverty measures for small
areas. For a given α ∈ {0, 1, 2}, let us define the random variables
 α
z − Edj
Fαdj = I(Edj < z), j = 1, . . . , Nd , d = 1, . . . , D.
z

Then the FGT poverty measure for area d is the mean

Nd
1 X
Fαd = Fαdj , d = 1, . . . , D. (5)
Nd j=1

5
 Suppose that there is a one-to-one transformation Ydj = T (Edj ) of the welfare
variables, Edj , such that the vector y containing the values of the transformed
variables Ydj for all the population units satisfies y ∼ N (µ, V). Then we can
express the random variables Fαdj in terms of Ydj as
α
z − T −1 (Ydj )

Fαdj = I{T −1 (Ydj ) < z} = hα (Ydj ), j = 1, . . . , Nd .
z

Thus, the FGT poverty measure (5) is a non-linear function of the vector y. By
the results of Section 4, taking δ = Fαd the BP of Fαd is
B
F̂αd = Eyr (Fαd |ys ). (6)

Using the decomposition of the mean (5) in terms of sample and out-of-sample
elements, we have ( )
1 X X
Fαd = Fαdj + Fαdj , (7)
Nd j∈s j∈r
d d

where rd denotes the set of out-of-sample elements belonging to area d. Now

taking conditional expectation of (7) and introducing the conditional expectation
inside the sum, the BP becomes
( )
B 1 X X
B
F̂αd = Fαdj + F̂αdj , (8)
Nd j∈s j∈r
d d

B
where F̂αdj = Eyr (Fαdj |ys ) is the BP of the out-of-sample variable Fαdj = hα (Ydj ),
which is defined as
Z
B
F̂αdj = Eyr [hα (Ydj )|ys ] = hα (Ydj )f (Ydj |ys ) dYdj , j ∈ rd , (9)
IR

where f (Ydj |ys ) is the conditional density of Ydj given the data vector ys . The
expectation in (9) cannot be calculated explicitly due to the complexity of hα .
However, since y = (ys′ , yr′ )′ is Normally distributed with mean vector µ =
(µ′s , µ′r )′ and covariance matrix partitioned conformably as
 
Vs Vsr
V= ,
Vrs Vr

the conditional distribution of yr given ys is

yr |ys ∼ N (µr|s , Vr|s ), (10)

where
µr|s = µr + Vrs Vs−1 (ys − µs ), Vr|s = Vr − Vrs Vs−1 Vsr . (11)
Then, we propose to use an empirical approximation by Monte Carlo simulation
(ℓ)
of a large number L of vectors yr generated from (10). Let Ydj be the value

6
of the out-of-sample observation Ydj , j ∈ rd , obtained in the ℓ-th simulation,
ℓ = 1, . . . , L. A Monte Carlo approximation to the best predictor of Ydj is then
given by
L
B 1X (ℓ)
F̂αdj = Eyr [hα (Ydj )|ys ] ≈ hα (Ydj ), j ∈ rd . (12)
L ℓ=1
In practice, the mean vector µ and the covariance matrix V usually depend
on an unknown vector of parameters θ. Thus the conditional density f (Ydj |ys )
depends on θ and we make this explicit in the notation using f (Ydj |ys , θ). We
can take an estimator θ̂ of θ such as the maximum likelihood (ML) estimator or
the residual ML (REML) estimator. Then the expectation can be approximated
(ℓ)
by generating values Ydj from the estimated density f (Ydj |ys , θ̂). The resulting
EB
predictor, denoted F̂αdj , is called the empirical best predictor (EBP). Finally,
the EBP of the poverty measure Fαd is given by
" #
EB 1 X X
EB
F̂αd = Fαdj + F̂αdj .
Nd j∈s j∈r
d d

In this paper, we consider the nested error linear regression model for the
Ydj (Battese et al., 1988). This model relates the transformed variables Ydj (e.g.,
log-earnings) to a vector of p explanatory variables xdj for all areas, and includes
a random area-specific effect ud along with the usual residual errors edj :

Ydj = x′dj β + ud + edj , ud ∼ iid N (0, σu2 ),

edj ∼ iid N (0, σe2 ), j = 1, . . . , Nd , d = 1, . . . , D, (13)

where the area effects ud and the errors edj are independent. Defining vectors
and matrices obtained by stacking the elements for area d as

yd = col (Ydj ), ed = col (edj ), Xd = col (x′dj ),

1≤j≤Nd 1≤j≤Nd 1≤j≤Nd

the vectors yd , d = 1, . . . , D, under model (13), are independent with

yd ∼ N (µd , Vd ),

where
µd = Xd β and Vd = σu2 1Nd 1′Nd + σe2 INd .
We assume that the population model (13) holds for the sample, i.e., sample
selection bias is absent (Pfeffermann et al. 1998).
Consider the decomposition of yd into sample and out-of-sample elements
′ ′ ′
yd = (ydr , yds ) , and the corresponding decomposition of µd = E(yd ) and Vd =
Var(yd ). Then the distribution of the out-of-sample vector ydr given the sample
data yds is
ydr |yds ∼ N (µdr|s , Vdr|s ), (14)

7
where the conditional mean vector and covariance matrix are given by
−1
µdr|s = Xdr β + σu2 1Nd −nd 1′nd Vds (ys − Xs β) (15)
Vdr|s = σu2 (1 − γd )1Nd −nd 1′Nd −nd + σe2 INd −nd . (16)

Note that ydr |yds and ydr |ys have the same distribution (14) due to the inde-
pendence of the vectors yd .
Observe that the application of the Monte Carlo approximation (12) involves
simulation of D multivariate Normal vectors of sizes Nd − nd , d = 1, . . . , D, from
(14). Then this process has to be repeated L times, something computationally
very intensive. This can be avoided in the following way. Observe that the
matrix Vdr|s corresponds to the covariance matrix of a vector ydr generated by
the model
ydr = µdr|s + vd 1Nd −nd + ǫdr , (17)
with new random effects vd and errors ǫdr that are independent and satisfy

vd ∼ N (0, σu2 (1 − γd )) and ǫdr ∼ N (0Nd −nd , σe2 INd −nd ).

Using these relations, instead of generating a multivariate normal vector of size

Nd −nd , we just need to generate univariate normal variables vd ∼ N (0, σu2 (1−γd ))
and ǫdj ∼ N (0, σe2 ) independently, for j ∈ rd , and then obtain the responses Ydj
from (17) using the known value of µdr|s . As mentioned before, in practice all the
unknown model parameters β, σu2 and σe2 are replaced by suitable estimators, and
then the variables Ydj are generated from the corresponding estimated normal
distributions.
Summarizing, the proposed EBP method to estimate poverty measures works
in the following way:

(a) Fit model (13) to the initial (transformed) data ys .

(ℓ)
(b) Draw L out-of-sample vectors yr , ℓ = 1, . . . , L from (14), or equivalently
from (17), but with the unknown parameters replaced by the estimators
obtained in (a).
(ℓ)
(c) With the L generated vectors yr , ℓ = 1, . . . , L, and using the sample data
ys , compute EBPs of the poverty measures from (8) using the Monte Carlo
approximation (12).

Instead of introducing the expectation inside the sum as in (8), we can approx-
imate by Monte Carlo directly the expectation of the poverty measure (6). In
this way, the proposed procedure can be used to predict any other function h(y)
and it requires only the assumption of a model for some transformation of the
welfare variable Edj .

8
6 Parametric bootstrap for MSE estimation
Here we propose to use an extension of the parametric bootstrap method for
finite populations (González-Manteiga et al., 2008) to estimate the MSE of the
EB
empirical best predictors F̂αd , d = 1, . . . , D. For a given α, this method works
as follows:

1. Fit model (13) to sample data ys and obtain model parameter estimates
β̂, σ̂u2 and σ̂e2 .

2. Generate bootstrap random area effects as u∗d ∼ iid N (0, σ̂u2 ), d = 1, . . . , D.

3. Generate, independently of the random effects u∗d , bootstrap random errors

e∗dj ∼ iid N (0, σ̂e2 ), j = 1, . . . , Nd , d = 1, . . . , D.

4. Construct a bootstrap population using the estimated model,

Ydj∗ = x′dj β̂ + u∗d + e∗dj , j = 1, . . . , Nd , d = 1, . . . , D, (18)

and calculate the FGT measures for this population, that is, calculate first
∗
Fαdj = hα (Ydj∗ ), j = 1, . . . , Nd , and then take the small area means as

Nd
∗ 1 X ∗
Fαd = Fαdj , d = 1, . . . , D.
Nd j=1

5. Take the elements Ydj∗ with indices contained in the sample s, denoted ys∗ .
Fit the model again to ys∗ obtaining new model parameter estimates β̂ ∗ ,
σ̂u2∗ and σ̂e2∗ .

6. Using the bootstrap sample data ys∗ and the known matrix X, apply the EB
EBP ∗
method as described in Section 5 and calculate bootstrap EBPs, F̂αd ,
d = 1, . . . , D.

Observe that the bootstrap elements, given the original sample data, preserve
properties of the original population model. Random effects and errors are iid
with

E∗ (u∗d ) = 0, V ar∗ (u∗d ) = σ̂u2 , E∗ (e∗dj ) = 0, V ar∗ (e∗dj ) = σ̂e2 , (19)

where E∗ and V ar∗ denote expectation and variance with respect to the distri-
bution defined by the bootstrap model (18) given sample data ys . Consider the
vector y∗ with all the population bootstrap elements defined analogously to y
for the original population. Then the mean vector and covariance matrix of this
bootstrap vector are

E∗ (y∗ ) = Xβ̂, V ar∗ (y∗ ) = σ̂u2 ZZ′ + σ̂e2 IN .

9
Thus, the distribution of the bootstrap population y∗ (given sample data ys )
EB
imitates that of the original population y. Then an estimator of MSE(F̂αd ) is
the bootstrap MSE of the bootstrap EBP, defined as
EBP ∗ EBP ∗ ∗ 2
MSE∗ (F̂αd ) = E∗ {(F̂αd − Fαd ) }.

In practice, this quantity can be approximated through a Monte Carlo procedure,

by repeating steps 2–6 a large number of times, B, and then taking the mean over
∗(b) EB∗(b)
the B replicates. More specifically, let Fαd and F̂αd be the poverty measure
and its corresponding EBP for the bootstrap replicate b, for b = 1, . . . , B. Then,
the estimator of the MSE is calculated as
B
1 X EB∗(b) ∗(b)
EB
mse(F̂αd )= (F̂ − Fαd )2 . (20)
B b=1 αd

It is possible to obtain a better MSE estimator, in terms of relative bias, by

using a double bootstrap method (Hall and Maiti, 2006) applied to the trans-
formed sample data.

7 Estimation of small area means

In this section we restrict ourselves to the estimation of the small area means
Nd
1 X
Ȳd = Ydj , d = 1, . . . , D,
Nd j=1

since means are common target quantities which also deserve attention; more-
over, means allow for the study of some theoretical properties of estimators.
Consider first the case of population elements Ydj following a linear model
without area effects,

Ydj = x′dj β + edj , edj ∼ iid N (0, σe2 ), j = 1, . . . , Nd , d = 1, . . . , D. (21)

Let us assume for simplicity of exposition that all model parameters are known.
Taking the average of (21) over the elements in area d, we can express the true
mean as
Ȳd = x̄′d β + ēd ,
where x̄d = Nd−1 N
P d ′ −1
PN d
j=1 xdj and ēd = Nd j=1 edj . Common small area es-
timators derived under model (21) are the synthetic estimators, obtained by
predicting all the Ydj and then taking the area mean, that is, taking
Nd
ˆ SY N 1 X
Ȳd = Ŷdj ,
Nd j=1

10
as the synthetic estimator, where Ŷdj = x′dj β is the predictor of Ydj for j =
1, . . . , Nd . However, under model (21), the BP of Ȳd is obtained by predicting
only the out-of-sample observations and keeping the sample data, i.e.,
( )
1
ȲˆdB =
X X
Ydj + Ŷdj .
Nd j∈s j∈r
d d

Let us compare the MSEs of the synthetic and the BP estimators. Writing
the synthetic estimator as ȲˆdSY N = x̄′d β, we obtain

ȲˆdSY N − Ȳd = x̄′d β − (x̄′d β + ēd ) = ēd ,

and then the MSE becomes

V ar(edj ) σ2
MSE(ȲˆdSY N ) = E{(ȲˆdSY N − Ȳd )2 } = E(ē2d ) = = e.
Nd Nd
However, observe that the difference between the BP and the true mean is equal
to
1 X
ȲˆdB − Ȳd = edj ,
Nd j∈r
d

which implies that the MSE of ȲˆdB is given by

σe2 σe2
 
ˆ ˆ nd
B B 2
MSE(Ȳd ) = E{(Ȳd − Ȳd ) } = 1− < = MSE(ȲˆdSY N ).
Nd Nd Nd

Thus, under model (21) with known parameters, the BP has always smaller MSE
due to the more efficient use of available information, namely the sample data.
When the sampling fraction nd /Nd is negligible both estimators have a similar
MSE.
Now consider the case of extra area variation not explained by the auxiliary
variables; that is, the true model is (13), but we fit the model (21). The true
mean for area d is then given by

Ȳd = x̄′d β + ud + ēd .

Then the MSE of the synthetic estimator under the true model is
σe2
MSE(ȲˆdSY N ) = E[(ud + ēd )2 ] = σu2 + . (22)
Nd
For the BP, the MSE under the true model is
σe2
 
ˆ B 2 nd
MSE(Ȳd ) = σu + 1− ,
Nd Nd

which is again smaller than MSE(ȲˆdSY N ).

11
 The method of Elbers et al. (2003), henceforth called the ELL method,
assumes a linear model with random cluster effects, where clusters may be dif-
ferent from the areas. They computed small area estimators by (a) fitting a
linear model with cluster random effects, (b) generating bootstrap random clus-
ter effects, (c) generating bootstrap random errors, (d) constructing a population
from the bootstrap model

Ydj∗ = x′dj β + u∗c + e∗dj , j = 1, . . . , Nd , d = 1, . . . , D, (23)

where u∗c is the bootstrap random effect of cluster c and (e) calculating the
average of the bootstrap elements from area d:
Nd
X
Ȳd∗ = Nd−1 Ydj∗ . (24)
j=1

The ELL estimator of the small area mean Ȳd is then the bootstrap mean

ȲˆdELL = E∗ (Ȳd∗ ),

which, in practice, is obtained from a Monte Carlo approximation by generating

a large number, L, of populations, calculating the area means (24) for each
population and later averaging over the L populations. The bootstrap variance
is used as the mean squared error of the ELL estimator; that is, the ELL method
uses
MSE(ȲˆdELL ) = V ar∗ (Ȳd∗ ),
which again is approximated by Monte Carlo simulations.
In the case of no clusters (as in some establishment surveys), the ELL method
fits the linear model (21) and uses this model to construct bootstrap populations.
Then the bootstrap mean for area d becomes
Nd Nd
1 X
(x′ β + e∗dj ) = ȲˆdSY N + ē∗d ,
X
Ȳd∗ = Nd−1 Ydj∗ =
j=1
Nd j=1 dj

and the ELL estimator is given by

ȲˆdELL = E∗ (Ȳd∗ ) = E∗ (ȲˆdSY N + ē∗d ) = ȲˆdSY N + E∗ (ē∗d ) = ȲˆdSY N ,

due to property (19) of the bootstrap method and the fact that the expectation
E∗ is conditional on the sample. The ELL estimator of the MSE is

ˆ ∗ ∗ ∗ 2 ∗ 2
V ar∗ (e∗dj ) σ2
ELL
MSE(Ȳd ) = V ar∗ (Ȳd ) = E∗ [(Ȳd − E∗ (Ȳd )) ] = E∗ [(ēd ) ] = = e.
Nd Nd
Thus, when fitting a model without cluster effects, the ELL estimator of a small
area mean is essentially the synthetic estimator, which is a good estimator when
the true model is (21). When the true model is (13), the ELL method is not

12
accounting for the area effects and the MSE estimator has a bias equal to σu2 ,
see (22). Thus, the ELL estimator of MSE can lead to serious underestimation
of MSE when the area effects have a substantial variance σu2 .
Now consider the ideal case for ELL, in which clusters are the same as the
small areas. In this case we use the ELL method by fitting the correct model
(13). Then the ELL estimator is again

ȲˆdELL = E∗ (ȲˆdSY N + u∗d + ē∗d ) = ȲˆdSY N ,

and the ELL estimator of the MSE becomes

σe2
MSE(ȲˆdELL ) = V ar∗ (Ȳd∗ ) = E∗ [(Ȳd∗ − E∗ (Ȳd∗ ))2 ] = E∗ [(u∗d + ē∗d )2 ] = σu2 + ,
Nd
which is the MSE of the synthetic estimator under model (13). Thus, when
the clusters are equal to the small areas, the ELL estimator remains essentially
equal to the synthetic estimator, but in this case the ELL variance estimator
is unbiased. Actually, under model (13), the difference between the ELL and
EB methods is that the target quantities are not the same. EB method tries to
estimate (or better predict) the actual area means Ȳd , while the ELL method is
estimating instead the marginal expectations E(Ȳd ).

8 Model-based simulation experiment

A model-based simulation study has been carried out to study the performance
of the proposed EB predictors of small area FGT poverty measures with α = 0
(proportion of people under poverty) and α = 1 (poverty gap). For this, we sim-
ulated populations of size N = 20000, composed of D = 80 areas with Nd = 250
elements in each area d = 1, . . . , D. The response variables for the population
units Ydj were generated from model (13) taking as auxiliary variables two dum-
mies X1 ∈ {0, 1} and X2 ∈ {0, 1} plus an intercept. The values of these two
dummies for the population units were generated from Bernouilli distributions
with success probabilities increasing with the area index for X1 and constant for
X2 ; that is,

p1d = 0.3 + 0.5 d/20; p2d = 0.2, d = 1, . . . , D,

respectively. Here the welfare variables Edj are the exponential of the responses
Ydj ; that is, the transformation T (·) defined in Section 5 is T (x) = log(x). A set of
sample indices s was drawn from the population indices {1, . . . , N }, by a stratified
design with areas as strata and simple random sampling without replacement
within each area. The values of the auxiliary variables for the population units
and the sample indices were kept fixed over all Monte Carlo simulations.
The intercept and the regression coefficients associated with the two auxiliary
variables used to generate populations were β = (3, 0.03, −0.04)′ . In this way,

13
the mean welfare increases when moving from the case (X1 = 0, X2 = 0) to
(X1 = 1, X2 = 0), but decreases when moving from (X1 = 0, X2 = 0) to (X1 =
0, X2 = 1). This implies that the “richest” individuals are those with values
X1 = 1 and X2 = 0. Since the probability p1d of X1 = 1 increases with the
area index but that of X2 = 1 is constant, then the last areas will have more
individuals with larger Ydj and then the FGT poverty measures will decrease
with the area index. The random area effects variance was taken as σu2 = (0.15)2
and the error variance as σe2 = (0.5)2 . The poverty line z was fixed as z = 12,
which is roughly equal to 0.6 times the median of the welfare variables Edj for a
population generated as mentioned above. In this way, the poverty incidence for
the simulated populations is approximately 16%.
Under this setup, I = 1000 populations y(i) were generated from the true
model. For each population i, we carried out the following tasks:

(a) The true area poverty incidences and gaps (FGT measures for α = 0 and
α = 1 respectively) were obtained for each population as
(i) α
Nd
!
(i) 1 X z − E dj (i)
Fαd = I(Edj < z), α = 0, 1, d = 1, . . . , D,
Nd j=1 z

(i) (i)
where Edj = exp(Ydj ) , j = 1, . . . , Nd .

(b) Sample estimators of the same quantities were also calculated as

(i) α
!
(i) 1 X z − E dj (i)
fαd = I(Edj < z), α = 0, 1, d = 1, . . . , D.
nd j∈s z
d

In this simulation study sample estimators are equal to direct estimators.

(c) Model (13) was fitted to the sample data. Then, substituting the estimated
(iℓ)
model parameters in (15) and (16), L = 50 out-of-sample vectors yr ,
ℓ = 1, . . . , L were generated from the conditional distribution (14) using
(i)
(17). Then using the sample data ys and the generated out-of-sample data
(iℓ)
yr , EB predictors of the area poverty incidences and gaps were calculated.
For this, first, for each sample unit j we considered the sample values
(i) α
!
(i) z − Edj (i)
Fαdj = I(Edj < z), j ∈ sd , d = 1, . . . , D,
z

(i)
while for each out-of-sample unit j, we computed the EB predictors of Fαdj
using the L simulated vectors,
(iℓ) α
L
!
EB(i) 1 X z − E dj (iℓ)
F̂αdj = I(Edj < z), j ∈ rd , d = 1, . . . , D,
L ℓ=1 z

14
 for α = 0, 1. Then the EB predictor of the FGT measure of order α = 0, 1
was obtained as
" #
EB(i) 1 X (i) X EB(i)
F̂αd = F + F̂ , d = 1, . . . , D.
Nd j∈s αdj j∈r αdj
d d

(d) The ELL estimators of the poverty measures were calculated. For
this, first the model (13) was fitted to the sample data and then L = 50
populations were generated using the parametric bootstrap described in
Section 6. For each population, the poverty measures were obtained and
finally, the results were averaged over the L = 50 populations.

Means over Monte Carlo populations i = 1, . . . , I of the true values of the

FGT measures of order α = 0, 1 were computed as
I
1 X (i)
E(Fαd ) = F , d = 1, . . . , D.
I i=1 αd

These means are the reference values for comparison. Finally, for all estimators,
namely, EBP, sample and ELL, the mean, variance and mean squared error over
Monte Carlo populations i = 1, . . . , I were obtained for each area d = 1, . . . , D.
(i)
Figure 1 a) plots the means of the true poverty incidences F0d , EB esti-
EB(i) (i) ELL(i)
mates F̂0d , direct estimates f0d and ELL estimates F̂0d for each area
d = 1, . . . , D. Figure 1 b) plots the MSEs of EB, direct and ELL estimators.
Similarly, Figures 2 a) and 2 b) plot the analogous quantities for the poverty
gap F1d . Observe that the means of EB estimates are for most areas close to the
mean true values. On the other hand, the direct estimates seem to be too unsta-
ble (due to small sample sizes) and the ELL estimators seem too stable, that is,
they do not seem to reproduce the between area variation. In terms of MSE, EB
estimators are clearly much better than the other two estimators. Surprisingly,
the MSEs of ELL estimators are larger than those of direct estimators.
Turning to MSE estimation, the parametric bootstrap procedure described in
Section 6 was implemented with B = 500 replicates and the results are plotted
in Figure 3. The number of Monte Carlo simulations was I = 500 and the
true values of MSE were independently computed with I = 50000 Monte Carlo
simulations. Observe that the bootstrap MSE estimator seems to track the
pattern of the true MSE values.

9 Design-based simulation experiment

A design-based simulation experiment was also carried out to study the perfor-
mance of estimators over repeated samples drawn from a fixed population. Only
one population was generated as described in Section 8, with the same popula-
tion and sample sizes, and using the same values of model parameters. Then,

15
 a) b)

70
16.5
EB Sample ELL
True
EB

60
Sample

MSE poverty incidence (x10000)

Poverty incidence (x100) ELL

16.0

50
40
15.5

30
20
15.0

10
0 20 40 60 80 0 20 40 60 80

Area Area

Figure 1: a) Means (×100) and b) MSEs (×104 ) over simulated populations of

true values, EB, sample and ELL estimators of the poverty incidence F0d for each
area d.

in each replication out of I = 1000, a new sample was drawn from this fixed
population according to SRS without replacement within each area. From each
sample, the three types of estimators of poverty measures, namely EBP, direct
and ELL were obtained.
Results for the poverty incidence are displayed in Figures 4 and 5. Figure 4
shows the true values and the means over the Monte Carlo samples of the EB,
direct and ELL estimators. Observe that the ELL estimators remain practically
constant across the areas. On the other hand, EB estimators track the true
values well, even though these estimators are supposed to have good theoretical
properties with respect to the model. Of course, direct estimators perform better
because they are design-unbiased. In terms of MSE, Figure 5 shows that ELL
estimators have small MSEs for some of the areas and large for the other areas,
while the MSE of EB and direct estimators remain small for all areas. For most
areas, the MSE of EB estimators is smaller than that of the direct estimators.

10 Application
The EB method was applied to compute poverty incidences and poverty gaps
in Spanish provinces crossed with gender. For this, data from the European
Survey on Income and Living Conditions (EUSILC) from the year 2006 has been
used. The welfare variable for the individuals is the normalized annual income
calculated following the standard procedure of the Spanish Statistical Institute
(INE). This variable has been transformed by adding a fixed quantity to make
it always positive and then taking logarithm. This transformed variable acts as
the response in the nested-error regression model. As auxiliary variables we have
considered the indicators of the 5 quinquennial groupings of the variable age,

16
 a) b)

3.8

6
True EB Sample ELL
EB

3.7
Sample

5
ELL

MSE poverty gap (x10000)

3.6
Poverty gap (x100)

4
3.5

3
3.4

2
3.3

1
3.2

0 20 40 60 80 0 20 40 60 80

Area Area

Figure 2: a) Means (×100) and b) MSEs (×104 ) over simulated populations of

E
true values, EB, sample and ELL estimators of the poverty gaps F̂1d for each
area d.

the indicator of having Spanish nationality, the indicators of the 3 levels of the
variable education level, and the indicators of the 3 categories of the variable
employment, with categories “unemployed”, “employed” and “inactive”. From
each variable, one of the categories was considered as base reference, omitting
the corresponding indicator and then including an intercept in the model.
The values of the dummy indicators are not known for the out-of-sample
units, but the EB method requires only the knowledge of the total number of
people with the same x-values. These totals were estimated using the sampling
weights attached to the sample units in the EUSILC.
The MSEs of the poverty measures were estimated by using the bootstrap
procedure described in Section 6. Values of EB estimates and coefficients of
variation (CVs) for the poverty incidence and the poverty gap are listed in Tables
1 and 2 respectively. In these tables we show also the direct estimates (2) and
their estimated variances following standard formulas in sampling theory, but
taking as observations the quantities Fαdj , j ∈ sd and using the EUSILC sampling
weights. Observe that the CVs of EB estimates are much smaller than those of
direct estimators for all except few domains, in which the CVs are similar for
both estimators. This improvement in efficiency is even more dramatic for the
poverty gap than for the poverty incidence. Moreover, the reduction in CV tends
to be greater for domains with small sample sizes. National statistical offices
usually establish a maximum publishable CV. For these data, the estimated CVs
of direct estimators of poverty incidences exceeded the level of 10% for 78 (out of
the 104) domains while those of the EB estimators exceeded this level for only 28
domains. If we increase the level to 20%, then the direct estimators have greater
CV for 17 domains but the CV of EB estimators exceeded 20% only for the first
domain.
Cartograms of the estimated poverty incidences and the poverty gaps in Span-

17
 a) b)
True MSE True MSE

0.86
10.8
Bootstrap MSE Bootstrap MSE

MSE Poverty incidence (x10000)

MSE Poverty gap (x10000)

0.84
10.6

0.82
10.4

0.80
10.2

0.78
10.0

0 20 40 60 80 0 20 40 60 80

Area Area

Figure 3: True MSEs (×104 ) of EB predictors and bootstrap MSE estimates

with B = 500 for each area d: a) Poverty incidence, b) Poverty gap.

ish provinces for males and females have been constructed using the EB esti-
mates, see Figures 6 and 7. In these maps we can see that the poorer provinces
concentrate mainly in the south and west parts of Spain. Provinces with critical
poverty incidences (over 30%) for men are, in the south: Almerı́a and Córdoba;
west: Badajoz, Ávila, Salamanza and Zamora and then Cuenca, situated east of
Madrid. For women the poverty incidences increase in most provinces, becoming
critical also, in the south: Granada, Jaén, Albacete and Ciudad Real, and in
the north: Palencia and Soria. The poverty level for Lérida (north-east) seems
unexpected considering that this province belongs to the region of Catalonia,
which is commonly considered as a rich region.
The poverty gap measures the degree of poverty instead of the quantity of
people under poverty. For a region with many people whose income is under the
poverty line but very close to it, the poverty gap will be close to zero. Observe
that the provinces with an income of over 12.5% under the poverty line are
also among those provinces with critical values of poverty incidence, except for
the northern provinces such as Lérida, which do not have significant gaps in
comparison with the rest of the provinces.

11 Conclusions
In this paper Empirical Best (EB) methodology to estimate poverty measures
is proposed. Parametric bootstrap is used for mean squared error (MSE) es-
timation. Simulation results show the good performance of EB estimators in
comparison with the direct and the ELL estimators. Simulation results confirm
the discussion that the ELL estimator is basically a synthetic-type estimator
derived from a linear regression model.
Model (17) illustrates a parallelism between ELL and EB methods. When

18
 35
True
EB
Sample

30
ELL
Poverty incidence (x100)

25
20
15
10
5

0 20 40 60 80

Area

Figure 4: Means (×100) over Monte Carlo samples of true values, EB, sample
and ELL estimators of the poverty incidence F0d for each area d.

EB
Sample
300

ELL
MSE poverty incidence (x10000)

250
200
150
100
50
0

0 20 40 60 80

Area

Figure 5: MSEs (×104 ) over Monte Carlo samples of EB, sample and ELL esti-
mators of the poverty incidence F0d for each area d.

19
 Poverty incidence − Men Poverty incidence − Women

45

45
40

40
35

35
under 15 under 15
15 − 20 15 − 20
20 − 25 20 − 25
30

30
25 − 30 25 − 30
over 30 over 30

−15 −10 −5 0 5 −15 −10 −5 0 5

Figure 6: Cartograms of estimated percent poverty incidences in Spanish

provinces for Men and Women.

Poverty gap − Men Poverty gap − Women

45

45
40

40
35

35

under 5 under 5
5 − 7.5 5 − 7.5
7.5 − 10 7.5 − 10
30

30

10 − 12.5 10 − 12.5
over 12.5 over 12.5

−15 −10 −5 0 5 −15 −10 −5 0 5

Figure 7: Cartograms of estimated percent poverty gaps in Spanish provinces for

Men and Women.

20
the clusters in the ELL method are taken to be equal to the small areas, the
ELL method generates a full population or census file of responses Ydj from the
bootstrap model (23) with vc∗ = vd∗ . Then the poverty measure is calculated from
this census file. The procedure is repeated a large number of times and finally the
computed poverty measures are averaged over bootstrap replications. The EB
method also creates a new census file, but first plugging in the observed sample
elements Ydj in their corresponding place, and then generating only the non-
sample values from the conditional model (17). The main difference between
model (17) and bootstrap model (23) used for the ELL method is the term
−1
σu2 1Nd −nd 1′nd Vds (ys −Xs β) appearing in the conditional mean given in (16). The
rest of the procedure is the same as in the ELL method. Thus, this term makes
an improvement for areas that are not fully explained by auxiliary variables and
therefore reduces the MSE of estimators significantly.
We remark that EB is a model-based method that relies on the validity of
the model. Thus, model selection procedures and model diagnostics are essential
in the practical application of this methodology.

References
Ballini, F., Betti, G. and Neri, L. (2006). Poverty and inequality mapping in
the Commonwealth of Dominica. Preprint.

Battese, G. E., Harter, R. M. and Fuller, W. A. (1988). An Error-Components

Model for Prediction of County Crop Areas Using Survey and Satellite
Data. Journal of the American Statistical Association, 83, 28–36.

Bell, W. (1997). Models for county and state poverty estimates. Preprint,
Census Statistical Research Division.

Elbers, C., Lanjouw, J. O. and Lanjouw, P. (2003). Micro-level estimation of

poverty and inequality. Econometrica, 71, 355–364.

Fay, R. E. and Herriot, R. A. (1979). Estimation of income from small places:

An application of James-Stein procedures to census data. Journal of the
American Statistical Association, 74, 269–277.

Foster, J., Greer, J. and Thorbecke, E. (1984). A class of decomposable poverty

measures, Econometrica, 52, 761–766.

González-Manteiga, W., Lombardı́a, M. J., Molina, I., Morales, D. and Santa-

marı́a, L. (2008). Journal of Statistical Computation and Simulation, 75,
443–462.

Hall, P. and Maiti, T. (2006). On Parametric Bootstrap Methods for Small

Area Prediction. Journal Royal Statistical Society, Series B, 68, 221–238.

21
Haslett, S. and Jones, G. (2005). Small area estimation using surveys and some
practical and statistical issues. Statistics in Transition, 7, 541–555.

Neri, L., Ballini, F. and Betti, G. (2005). Poverty and inequality in transition
countries. Statistics in Transition, 7, 135–157.

Pfeffermann, D., Skinner, C. J., Holmes, D. J., Goldstein, H. and Rasbash, J.

(1998). Weighting in Unequal Probabilities in Multilevel Models. Journal
of the Royal Statistical Society B, 60, 23–40.

Rao, J. N. K. (2003). Small Area Estimation. London: Wiley.

Royall, R. M. (1976). The Linear Least Squares Prediction Approach to Two-

Stage Sampling, Journal of the American Statistical Association, 71, 657–
664.

Tarozzi, A. and Deaton, A. (2007). Using census and survey data to estimate
poverty and inequality for small areas. Preprint.

22
Application results
Table 1. Results on poverty incidence: Spanish SILC data.
w EB w EB w EB
Province Sex Nd nd f0d F̂0d var(f0d ) mse(F̂0d ) CV(f0d ) CV(F̂0d ) Ratio
Álava M 99354 95 7.10 12.84 0.6751 0.6805 36.60 20.32 1.80
Álava F 108422 96 14.60 12.50 1.5637 0.6000 27.08 19.60 1.38
Albacete M 184058 163 30.11 29.22 1.0801 0.4617 10.92 7.35 1.48
Albacete F 186503 183 30.58 33.74 1.1558 0.4618 11.12 6.37 1.75
Alicante M 929288 526 17.96 19.45 0.2793 0.1466 9.31 6.23 1.49
Alicante F 931405 552 17.95 22.59 0.2482 0.1601 8.78 5.60 1.57
Almerı́a M 341228 204 35.34 32.88 1.3701 0.3642 10.47 5.80 1.80
Almerı́a F 318857 193 33.54 35.72 1.1329 0.5020 10.04 6.27 1.60
Ávila M 56601 56 20.85 31.48 2.5112 1.2061 24.03 11.03 2.18
Ávila F 61708 60 20.42 38.51 2.4398 1.3285 24.19 9.46 2.56
Badajoz M 351985 472 31.97 36.56 0.5177 0.1703 7.12 3.57 1.99
Badajoz F 346810 515 34.90 39.13 0.4958 0.1947 6.38 3.57 1.79
Baleares M 477561 609 12.76 11.55 0.2297 0.1042 11.88 8.84 1.34
Baleares F 472843 660 15.57 14.05 0.2228 0.1130 9.59 7.57 1.27
Barcelona M 2617681 1358 9.82 10.49 0.0569 0.0524 7.68 6.90 1.11
Barcelona F 2752431 1483 11.80 13.10 0.0619 0.0494 6.67 5.37 1.24
Burgos M 215155 168 15.35 16.72 0.7776 0.3736 18.16 11.56 1.57
Burgos F 211240 167 17.82 18.33 0.8419 0.4097 16.28 11.04 1.47
Cáceres M 169833 261 33.20 24.69 0.8999 0.2099 9.04 5.87 1.54
Cáceres F 184785 302 41.91 28.24 0.9402 0.2689 7.32 5.81 1.26
Cádiz M 642053 373 34.75 26.88 0.5629 0.2013 6.83 5.28 1.29
Cádiz F 681522 422 35.92 31.63 0.5660 0.2316 6.62 4.81 1.38
Castellón M 201428 113 18.03 14.79 1.1401 0.6489 18.73 17.22 1.09
Castellón F 197726 123 17.71 17.35 1.2807 0.6008 20.21 14.13 1.43
Ciudad Real M 265393 260 27.97 28.39 0.8795 0.3060 10.60 6.16 1.72
Ciudad Real F 254508 239 33.82 30.18 0.9979 0.3598 9.34 6.29 1.49
Córdoba M 356218 217 37.15 30.16 1.0097 0.3636 8.55 6.32 1.35
Córdoba F 364583 230 41.55 33.32 0.9769 0.5025 7.52 6.73 1.12
La Coruña M 509141 457 22.89 24.66 0.3935 0.1549 8.67 5.05 1.72
La Coruña F 563190 533 19.36 25.36 0.2888 0.1789 8.78 5.27 1.66
Cuenca M 92275 96 35.91 35.26 3.2294 0.6676 15.82 7.33 2.16
Cuenca F 86760 87 43.95 35.35 3.0522 0.9183 12.57 8.57 1.47
Gerona M 307975 145 11.23 13.29 0.6006 0.4512 21.83 15.98 1.37
Gerona F 245519 138 11.38 15.38 0.5847 0.5672 21.25 15.48 1.37
Granada M 371735 188 31.53 29.16 0.8038 0.3423 8.99 6.35 1.42
Granada F 424598 229 35.74 36.34 0.8358 0.3340 8.09 5.03 1.61
Guadalajara M 87591 92 13.53 12.74 0.9309 0.6339 22.55 19.76 1.14
Guadalajara F 79560 86 16.77 15.83 1.1738 0.9055 20.43 19.01 1.08
Guipúzcoa M 323719 279 10.10 11.30 0.3226 0.2488 17.79 13.96 1.27
Guipúzcoa F 348524 291 11.70 14.56 0.3297 0.2250 15.52 10.30 1.51
Huelva M 223158 121 34.24 29.06 1.6956 0.4976 12.03 7.68 1.57
Huelva F 214587 123 30.95 29.13 1.5825 0.5606 12.85 8.13 1.58
Huesca M 96617 125 11.67 17.11 0.8957 0.6126 25.64 14.47 1.77
Huesca F 91147 105 14.44 18.99 1.1769 0.8048 23.76 14.94 1.59
Jaén M 380752 233 24.11 28.60 0.8679 0.2878 12.22 5.93 2.06
Jaén F 356344 230 24.76 32.31 0.8523 0.4010 11.79 6.20 1.90
León M 204462 209 18.96 22.60 0.7566 0.3772 14.51 8.59 1.69
León F 225753 228 20.69 24.17 0.7933 0.3639 13.62 7.89 1.72

23
Lérida M 214123 127 16.82 25.74 1.2317 0.6116 20.87 9.61 2.17
Lérida F 218051 133 15.57 27.36 1.0677 0.5777 20.98 8.79 2.39
La Rioja M 149238 519 16.75 18.57 0.3618 0.1383 11.36 6.33 1.79
La Rioja F 147554 500 18.50 21.45 0.3613 0.1557 10.27 5.82 1.77
Lugo M 175462 169 32.68 24.51 1.4726 0.3914 11.74 8.07 1.45
Lugo F 167892 177 30.17 26.87 1.3821 0.4034 12.32 7.47 1.65
Madrid M 2816184 893 7.95 12.06 0.0513 0.0619 9.01 6.52 1.38
Madrid F 3011923 996 9.36 13.91 0.0527 0.0704 7.76 6.04 1.29
Málaga M 693871 361 18.96 27.95 0.5823 0.2031 12.73 5.10 2.50
Málaga F 702667 397 22.88 32.45 0.5375 0.2190 10.13 4.56 2.22
Murcia M 668714 868 23.83 25.35 0.2471 0.1027 6.60 4.00 1.65
Murcia F 660107 902 23.67 28.70 0.2371 0.1087 6.50 3.63 1.79
Navarra M 286947 525 11.03 9.13 0.1812 0.1405 12.21 12.98 0.94
Navarra F 289947 603 12.98 11.40 0.2185 0.1211 11.39 9.66 1.18
Orense M 120257 118 24.94 25.07 1.3845 0.5993 14.92 9.77 1.53
Orense F 137587 140 21.27 22.12 1.0968 0.4809 15.57 9.92 1.57
Oviedo M 511169 742 12.26 16.01 0.1823 0.0824 11.01 5.67 1.94
Oviedo F 546817 864 12.56 16.59 0.1464 0.0893 9.63 5.70 1.69
Palencia M 75638 71 27.29 26.16 2.3375 1.0455 17.72 12.36 1.43
Palencia F 72558 72 30.63 30.13 2.5426 1.0907 16.46 10.96 1.50
Las Palmas M 592262 458 24.45 24.65 0.5337 0.1615 9.45 5.16 1.83
Las Palmas F 580265 485 29.76 25.40 0.5325 0.1520 7.75 4.85 1.60
Pontevedra M 494161 434 13.03 19.15 0.2430 0.1620 11.97 6.64 1.80
Pontevedra F 525627 462 15.69 22.66 0.2803 0.1865 10.67 6.03 1.77
Salamanca M 151335 166 26.92 31.46 1.2025 0.3862 12.88 6.25 2.06
Salamanca F 152234 162 31.58 33.56 1.3392 0.4030 11.59 5.98 1.94
Tenerife M 366253 370 17.63 24.14 0.3997 0.1590 11.34 5.22 2.17
Tenerife F 376690 392 17.07 26.36 0.3078 0.2006 10.28 5.37 1.91
Santander M 267290 424 8.79 16.00 0.1586 0.1398 14.33 7.39 1.94
Santander F 279191 443 12.65 16.93 0.2339 0.1678 12.09 7.65 1.58
Segovia M 62518 57 39.16 19.24 3.7441 1.0910 15.63 17.17 0.91
Segovia F 63217 58 47.03 26.74 3.7211 1.2032 12.97 12.97 1.00
Sevilla M 816795 472 22.06 19.61 0.2737 0.1575 7.50 6.40 1.17
Sevilla F 853057 491 28.07 24.04 0.3173 0.1493 6.35 5.08 1.25
Soria M 26431 24 23.68 26.33 8.0523 2.0666 37.89 17.26 2.19
Soria F 17211 17 26.73 31.48 11.6416 2.7052 40.37 16.52 2.44
Tarragona M 264627 129 12.96 14.86 0.6612 0.5761 19.85 16.15 1.23
Tarragona F 255490 139 16.19 19.28 0.6499 0.5197 15.75 11.82 1.33
Teruel M 53380 66 12.43 17.13 1.0073 0.8420 25.53 16.94 1.51
Teruel F 65002 78 16.22 22.26 1.3288 1.0112 22.48 14.29 1.57
Toledo M 288335 278 23.86 26.22 0.6157 0.1871 10.40 5.22 1.99
Toledo F 305241 272 20.56 22.50 0.5188 0.2784 11.08 7.42 1.49
Valencia M 1169258 686 17.53 17.89 0.1995 0.0940 8.06 5.42 1.49
Valencia F 1197478 742 19.64 20.78 0.1978 0.1162 7.16 5.19 1.38
Valladolid M 305496 292 14.98 15.34 0.4731 0.2216 14.52 9.70 1.50
Valladolid F 322530 306 18.59 18.29 0.4771 0.2352 11.75 8.38 1.40
Vizcaya M 576042 515 9.08 10.01 0.1458 0.1267 13.30 11.24 1.18
Vizcaya F 590094 532 10.26 11.57 0.1753 0.1175 12.91 9.37 1.38
Zamora M 101433 109 36.14 34.67 2.6728 0.7388 14.30 7.84 1.82
Zamora F 98337 100 36.53 32.84 2.3562 0.7964 13.29 8.59 1.55
Zaragoza M 466651 555 10.64 15.42 0.2081 0.1232 13.55 7.20 1.88
Zaragoza F 462937 574 10.22 15.34 0.1435 0.0989 11.72 6.48 1.81
Ceuta M 35705 223 40.87 30.26 1.5506 0.3482 9.63 6.17 1.56

24
Ceuta F 40426 247 42.80 33.15 1.4583 0.3804 8.92 5.88 1.52
Melilla M 30595 179 32.60 19.27 1.1714 0.3783 10.50 10.09 1.04
Melilla F 27498 180 29.56 25.45 1.0739 0.5579 11.09 9.28 1.19
Columns respectively denote province, gender, population size, sample size, direct estimate of
poverty incidence, EB estimate, estimated variance of direct estimator, estimated MSE of EB
estimator, CV of direct estimator, CV of EB estimator and ratio of CVs of direct estimators
over EB estimators. Estimated poverty incidences and CVs in percentage.

Table 2. Results on poverty gap: Spanish SILC data.

w EB w EB w EB
Province Sex Nd nd f1d F̂1d var(f1d ) mse(F̂1d ) CV(f1d ) CV(F̂1d ) Ratio
Álava M 99354 95 2.49 3.75 1.0904 1.1907 41.94 29.09 1.44
Álava F 108422 96 1.53 3.65 0.4942 1.1315 45.97 29.12 1.58
Albacete M 184058 163 9.63 10.53 2.9626 0.9286 17.87 9.15 1.95
Albacete F 186503 183 11.72 12.68 3.5333 0.9250 16.03 7.59 2.11
Alicante M 929288 526 5.00 6.29 0.5269 0.2937 14.53 8.62 1.68
Alicante F 931405 552 5.89 7.55 0.6127 0.3407 13.28 7.73 1.72
Almerı́a M 341228 204 10.81 12.30 2.3507 0.6839 14.19 6.73 2.11
Almerı́a F 318857 193 11.18 13.64 2.8714 1.0880 15.16 7.65 1.98
Ávila M 56601 56 10.82 11.64 7.0443 2.3237 24.54 13.09 1.87
Ávila F 61708 60 12.30 15.40 6.1424 2.8775 20.15 11.01 1.83
Badajoz M 351985 472 12.59 14.11 1.2979 0.3086 9.05 3.94 2.30
Badajoz F 346810 515 12.15 15.46 1.0543 0.4007 8.45 4.09 2.06
Baleares M 477561 609 2.88 3.34 0.4130 0.1955 22.28 13.24 1.68
Baleares F 472843 660 2.94 4.23 0.2716 0.2227 17.72 11.17 1.59
Barcelona M 2617681 1358 3.07 3.00 0.1224 0.0997 11.38 10.53 1.08
Barcelona F 2752431 1483 3.60 3.92 0.1297 0.1027 10.00 8.17 1.22
Burgos M 215155 168 4.22 5.21 2.2735 0.6704 35.72 15.70 2.27
Burgos F 211240 167 3.50 5.81 1.4983 0.8377 34.93 15.75 2.22
Cáceres M 169833 261 7.54 8.52 1.2188 0.3704 14.65 7.14 2.05
Cáceres F 184785 302 9.33 10.13 1.2620 0.5620 12.03 7.40 1.63
Cádiz M 642053 373 7.24 9.38 0.9284 0.4024 13.31 6.76 1.97
Cádiz F 681522 422 10.95 11.65 1.4154 0.5101 10.87 6.13 1.77
Castellón M 201428 113 3.97 4.48 2.8120 1.2242 42.19 24.68 1.71
Castellón F 197726 123 3.86 5.51 2.0386 1.3159 36.97 20.83 1.77
Ciudad Real M 265393 260 7.30 10.07 0.9995 0.6338 13.70 7.91 1.73
Ciudad Real F 254508 239 7.15 10.86 0.9134 0.7758 13.36 8.11 1.65
Córdoba M 356218 217 8.22 10.82 1.2822 0.6983 13.77 7.72 1.78
Córdoba F 364583 230 8.01 12.26 1.1694 1.1819 13.50 8.87 1.52
La Coruña M 509141 457 7.34 8.47 0.7480 0.2867 11.78 6.32 1.86
La Coruña F 563190 533 8.33 8.72 0.8716 0.3791 11.20 7.06 1.59
Cuenca M 92275 96 8.83 13.41 2.4195 1.4071 17.62 8.84 1.99
Cuenca F 86760 87 10.73 13.36 3.0724 2.0791 16.33 10.80 1.51
Gerona M 307975 145 1.87 3.95 0.5700 0.7954 40.35 22.56 1.79
Gerona F 245519 138 2.15 4.67 0.7537 1.0857 40.30 22.29 1.81
Granada M 371735 188 13.55 10.56 4.0423 0.6923 14.84 7.88 1.88
Granada F 424598 229 16.81 14.02 4.8343 0.7568 13.08 6.20 2.11
Guadalajara M 87591 92 1.52 3.80 0.2823 1.2206 34.88 29.10 1.20
Guadalajara F 79560 86 2.55 4.90 0.4615 1.7868 26.63 27.28 0.98
Guipúzcoa M 323719 279 2.60 3.25 0.9591 0.4277 37.69 20.09 1.88
Guipúzcoa F 348524 291 4.42 4.38 1.3093 0.4294 25.90 14.95 1.73
Huelva M 223158 121 10.46 10.37 7.2743 0.9412 25.78 9.36 2.75

25
Huelva F 214587 123 9.13 10.40 4.2187 1.0980 22.49 10.07 2.23
Huesca M 96617 125 2.56 5.39 1.2775 1.2615 44.18 20.86 2.12
Huesca F 91147 105 3.04 6.06 1.7064 1.5781 42.92 20.72 2.07
Jaén M 380752 233 9.63 10.28 1.8186 0.5968 14.01 7.51 1.86
Jaén F 356344 230 11.41 11.94 2.1644 0.8385 12.89 7.67 1.68
León M 204462 209 7.14 7.58 2.2850 0.7474 21.16 11.41 1.85
León F 225753 228 7.56 8.31 2.2879 0.8014 20.00 10.77 1.86
Lérida M 214123 127 9.22 9.08 4.8531 1.2797 23.88 12.46 1.92
Lérida F 218051 133 9.34 9.77 4.5156 1.2979 22.75 11.66 1.95
La Rioja M 149238 519 4.05 5.97 0.3139 0.2546 13.83 8.46 1.64
La Rioja F 147554 500 4.34 7.14 0.2958 0.3245 12.52 7.98 1.57
Lugo M 175462 169 8.64 8.40 6.9390 0.7199 30.50 10.10 3.02
Lugo F 167892 177 5.26 9.40 1.3626 0.8026 22.20 9.54 2.33
Madrid M 2816184 893 3.37 3.58 0.3812 0.1145 18.33 9.45 1.94
Madrid F 3011923 996 3.59 4.26 0.3350 0.1442 16.14 8.92 1.81
Málaga M 693871 361 8.95 9.90 1.9024 0.4162 15.41 6.52 2.37
Málaga F 702667 397 10.80 12.04 1.9554 0.4561 12.95 5.61 2.31
Murcia M 668714 868 7.54 8.74 0.4296 0.2175 8.69 5.34 1.63
Murcia F 660107 902 8.30 10.31 0.4393 0.2373 7.99 4.73 1.69
Navarra M 286947 525 2.99 2.53 0.3732 0.2389 20.45 19.28 1.06
Navarra F 289947 603 2.73 3.31 0.2752 0.2450 19.23 14.96 1.29
Orense M 120257 118 7.28 8.66 3.6924 1.1440 26.41 12.36 2.14
Orense F 137587 140 4.77 7.44 2.0973 0.9954 30.34 13.41 2.26
Oviedo M 511169 742 2.54 4.95 0.2335 0.1618 19.02 8.12 2.34
Oviedo F 546817 864 3.11 5.14 0.2300 0.1857 15.41 8.38 1.84
Palencia M 75638 71 5.65 9.10 2.9335 2.1179 30.32 15.99 1.90
Palencia F 72558 72 6.08 10.92 3.1612 2.5262 29.27 14.56 2.01
Las Palmas M 592262 458 7.63 8.40 1.5179 0.3192 16.15 6.73 2.40
Las Palmas F 580265 485 8.46 8.78 1.6326 0.3139 15.11 6.38 2.37
Pontevedra M 494161 434 3.00 6.09 0.1949 0.3100 14.73 9.14 1.61
Pontevedra F 525627 462 4.40 7.57 0.3114 0.3888 12.68 8.24 1.54
Salamanca M 151335 166 9.87 11.50 2.3273 0.7451 15.46 7.51 2.06
Salamanca F 152234 162 8.85 12.74 2.0322 0.8510 16.11 7.24 2.22
Tenerife M 366253 370 8.07 8.20 1.0073 0.2909 12.44 6.58 1.89
Tenerife F 376690 392 9.35 9.18 1.2484 0.4136 11.95 7.00 1.71
Santander M 267290 424 2.59 4.94 0.3045 0.2634 21.30 10.38 2.05
Santander F 279191 443 2.95 5.29 0.3210 0.3526 19.20 11.22 1.71
Segovia M 62518 57 7.01 6.30 4.5293 2.1717 30.36 23.41 1.30
Segovia F 63217 58 10.90 9.64 5.5058 2.9114 21.52 17.71 1.22
Sevilla M 816795 472 3.42 6.34 0.1609 0.2819 11.72 8.38 1.40
Sevilla F 853057 491 4.53 8.14 0.3045 0.3112 12.19 6.85 1.78
Soria M 26431 24 15.28 9.13 76.9805 3.9189 57.42 21.68 2.65
Soria F 17211 17 23.46 11.84 122.9756 5.5980 47.27 19.99 2.37
Tarragona M 264627 129 1.95 4.53 0.2800 1.0997 27.15 23.14 1.17
Tarragona F 255490 139 2.79 6.16 0.4304 1.1317 23.52 17.26 1.36
Teruel M 53380 66 4.48 5.49 5.9649 1.6368 54.54 23.29 2.34
Teruel F 65002 78 5.16 7.38 3.8559 1.9972 38.09 19.15 1.99
Toledo M 288335 278 7.69 9.18 1.3151 0.3438 14.92 6.39 2.33
Toledo F 305241 272 5.85 7.58 0.8997 0.6082 16.21 10.28 1.58
Valencia M 1169258 686 5.08 5.70 0.9538 0.1722 19.24 7.28 2.64
Valencia F 1197478 742 4.26 6.78 0.3187 0.2456 13.25 7.31 1.81
Valladolid M 305496 292 6.38 4.71 1.1430 0.4269 16.75 13.87 1.21
Valladolid F 322530 306 7.45 5.82 1.3767 0.4782 15.76 11.89 1.33

26
Vizcaya M 576042 515 2.57 2.80 0.2783 0.2338 20.49 17.27 1.19
Vizcaya F 590094 532 2.26 3.35 0.1756 0.2177 18.56 13.92 1.33
Zamora M 101433 109 12.58 13.10 5.5333 1.5147 18.71 9.40 1.99
Zamora F 98337 100 9.86 12.18 4.7252 1.7185 22.04 10.76 2.05
Zaragoza M 466651 555 4.29 4.77 0.7891 0.2377 20.69 10.23 2.02
Zaragoza F 462937 574 5.08 4.72 0.9837 0.1956 19.53 9.38 2.08
Ceuta M 35705 223 14.79 11.09 3.3694 0.7296 12.41 7.70 1.61
Ceuta F 40426 247 20.68 12.52 5.5107 0.8832 11.35 7.50 1.51
Melilla M 30595 179 11.87 6.22 7.3207 0.7442 22.80 13.86 1.64
Melilla F 27498 180 12.47 8.82 3.5770 1.1392 15.16 12.10 1.25
Columns respectively denote province, gender, population size, sample size, direct estimate
of poverty gap, EB estimate, estimated variance of direct estimator, estimated MSE of EB
estimator, CV of direct estimator, CV of EB estimator and ratio of CVs of direct estimators
over EB estimators. Estimated poverty gaps and CVs in percentage.

27