5.2 ANÁLISIS Y DESCRIPCIÓN DE LOS PUESTOS DE TRABAJO
5.2.1 Gerente
The multivariate Fay-Herriot Model (MFH) is an extension of the univariate Fay-
Herriot (UFH) model defined in (2.18) and (2.19). Let us consider an (r×1) vector of survey estimators θˆd such that
ˆ
θd=θd+ed, d = 1,2, . . . , D (2.26)
where θd = (θd1, θd2, . . . , θdr)0 is a (r ×1) vector with θd = g( ¯Ydt). Let θˆd =
(ˆθd1,θˆd2, . . . ,θˆdr)0 be a (r×1) vector of survey estimators of θd. The independent
sampling errors, ed = (ed1, ed2, . . . , edr)0, are normally distributed with E(ed) = 0
and Var(ed) = Ψd. Here 0is a (r×1) vector andΨd is a (r×r) matrix. ¯Ydr is the
dth small area mean forrth characteristic. The linking model is :
θd=Zd0β+vd, d= 1,2, . . . , D (2.27)
where β is a (p×1) vector of regression coefficients, vd = (vd1, vd2, . . . , vdr) are
identically, independently and multivariate normally distributed area specific ran-
dom effects with E(vd) = 0,Var(vd) =Σv, and Zd is a (p×r) matrix of auxiliary
variables.
Combining the sampling (2.26) and the linking model (2.27), the MFH model
becomes
ˆ
θd=Zd0β+vd+ed, d= 1,2, . . . , D (2.28)
The BLUP estimator of θˆd is given as
ˆ θ∗d=Σv(Ψd+Σv) −1 ˆ θd+Ψd(Ψd+Σv) −1 Zd0βˆ (2.29)
(Rao and Molina, 2015, p.236) where βˆis given by ˆ β = " D X d=1 Zd0(Ψd+Σv)−1Zd0 #−1" D X d=1 Zd0(Ψd+Σv)−1θˆd # (2.30)
The MSE of the BLUP estimator θˆd∗ is obtained as
MSEθˆd∗= Ψ−d1 + Σv−1 −1 + Ψ−d1+ Σv−1 −1 Σv−1Zd Zd0(Ψd+Σv)−1Zd −1 Zd0Σv−1(Ψd+Σv)−1 (2.31)
(Rao and Molina, 2015, p.236). The multivariate model (2.28) is a natural extension
of the UFH model (2.20). Since the multivariate area level Fay-Herriot model in-
corporates the correlations between the components of the direct estimators, it may
provide more efficient estimators of small area statistics compared to those based
on separate univariate models for each variable (Rao, 2003b, pp.236-237).
Gonz´alez-Manteiga et al. (2008) developed a multivariate extension of the Fay-
Herriot model in order to estimate the MSE of multidimensional response variables
in small areas. The method of moments is used to estimate the unknown parameters
(σv2) in their study. They estimated MSE using the analytical MSE approximation
(2.25) of Prasad and Rao (1990). Several parametric bootstrap MSE estimators are
also studied by Gonz´alez-Manteiga et al.(2008).
Gonz´alez-Manteiga et al. (2008) conducted a simulation study to examine the
efficiency of the bivariate Fay-Herriot model (BFH) model over the UFH model
as follows: Consider the response vectors θd = (θd1, θd2)
0
, d = 1,2, . . . , D with
the mean vectors θˆd =
ˆ θd1,θˆd2 0 . The covariates (Zd1, Zd2) 0
are generated from
a bivariate normal distribution with mean µZ1 = µZ2 = 10, variances σ2Z1 = 1
and σ2
Z2 = 2 with covariance σZ12 = 1/ √
vd1 = vd2 are generated from a normal distribution with mean 0 and variance 1.
Vectors of sampling errorsed= (ed1, ed2)
0
are generated from the multivariate normal
distribution with mean vector 0and variance-covariance matrix Ψd = (Ψdjk), j, k =
1,2 where Ψdjk = rjk
√
wd and wd is heteroscedasticity weights. For generating ed,
Gonz´alez-Manteiga et al. (2008) used r11 = 1, r22 = 2 and r12 = ρe
p
(r11r22) and
the heteroscedasticity weights wd =
p
Z2
d1+Zd22. The regression coefficients are
β1 = β2 = 1. The two response variables are generated using the model (2.28).
Both the BFH and UFH models are fitted and the empirical MSE (2.25) of the
corresponding EBLUP (ˆθEd1 and ˆθEd2) are computed. For comparison, the median
percentage reduction from the UFH to the BFH is obtained for each of the 1000
datasets.
The correlation between sampling errors (ρe) was varied from -0.5 to 0.5. Their
study showed that the reduction in MSE from the UFH to the BFH model is greater
for the second variable since the auxiliary variables and the sampling errors are
generated with larger variance. The reduction in MSE is as large as 60% when
ρe = −0.5 for the first variable and up to about 30% for ρe = −0.5 for the sec-
ond variable. In this thesis, a simulation experiment has been conducted following
Gonz´alez-Manteiga et al. (2008) in Chapter 3. Based on the simulation study in
Chapter 3 (Section 3.5), it appears that Gonz´alez-Manteiga et al. (2008) may per-
haps have actually used heteroscedasticity weights wd = 1.
Benavent and Morales (2016) introduced a general class of multivariate Fay-
Herriot models from Gonz´alez-Manteiga et al. (2008) allowing for different covari-
small area parameters. Several simulation studies are performed by Benavent and
Morales (2016) in order to assess the behavior of multivariate EBLUP and to com-
pare the multivariate Fay-Herriot model with univariate Fay-Herriot by calculating
MSE estimators. Their developed methods are applied to data from the 2005 and
2006 Spanish living condition surveys and estimated the poverty proportions and
gaps at province level. The empirical research of Benavent and Morales (2016) based
on Monte Carlo simulations showed that the multivariate EBLUPs have lower MSE
than the corresponding univariate EBLUPs.
Porter et al. (2015) improved multivariate estimates by considering both multi-
variate outcomes and latent spatial dependence. They consider three models, one in
which the outcome-by-space dependence structure is separable, one that accounts
for the cross dependence between spatial location and outcome variables through
the use of a generalized multivariate conditional autoregressive (GMCAR) structure
and the another one is the state-of-the-art multivariate model with unstructured
dependence between outcomes and no spatial dependence. They demonstrated that
GMCAR-FH model shows excellent results. A simulation study of Porter et al.
(2015) based on the Missouri county lattice structure finds gains in MSE of around
10% when cross-correlations are weaker, and around 30-70% when they are stronger,
compared to a non-spatial multivariate model (2.28).
The multivariate extension of the univariate Fay-Herriot model (2.20) for small
area estimation has been suggested by Arima et al. (2017) for two reasons: one is
to produce improved small area predictions by capturing dependence between the
different population characteristics, for example, an estimate of year-to-year change
and an associated measure of uncertainty. In small area estimation, it is also im-
portant to account random error of the covariates in the modelling (Arima et al.,
2017). Ybarra and Lohr (2008) observed that when covariates for a small area are
measured with sufficient error, then model-based small area estimators yield worse
predictions (higher mean-squared error) than the direct estimators.
Arima et al.(2015) developed a Bayesian approach of the univariate Fay-Herriot
model (2.20) when some of the covariates are measured with error and their approach
is extended to multivariate generalization by Arimaet al.(2017). Arimaet al.(2017)
provided a Bayesian analysis of a multivariate Fay-Herriot model (2.28) with func-
tional measurement errors accounting for random errors in some of the covariates.
The functional measurement error model of Arima et al. (2017) assumes that the
underlying values of the covariates (Zd in (2.28)) are fixed but unknown quanti-
ties. They applied their approach to data on 2010 and 2011 US country poverty
rates of school-aged children, for predicting 2011 poverty rates and the 2010-2011
changes. The measurement error model showed results in great improvements in pre-
diction of country poverty rates compared to the direct estimates. The authors also
showed that for estimating yearly changes of poverty rates, the bivariate model that
accounts the dependence of the model errors of each year leads to lower standard
errors compared to univariate model that assumes independence of the measurement