sae under non-normal random effects
TRANSCRIPT
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
SAE under Non-normal Random Effects
Gauri Sankar Datta
Department of StatisticsUniversity of Georgia
Athens, GA 30602, [email protected]
Joint work with Drs. K. Irimata, J. Maples and E. SludUS Census Bureau
In Likelihood-free Methods of InferenceFebruary 18-19, 2019
Padua, Italy
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Outline
1 Small area models as hierarchical models, random effectsmodels and mixed linear models
2 Mixed linear models in SAE
3 Empirical Best Linear Unbiased Prediction
4 Approximation of MSE of EBLUP
5 MSE of EBLUP under non-normality
6 MSE estimation without normality
7 Simulation study
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Small Area Models
1 Explicit linking models based on random effects that accountfor between area variation beyond that is explained byauxiliary variables are known as “small area models”.
2 Resulting indirect estimators are “model-based estimators”.
3 Model-based indirect estimators are now the norm in SAE
4 Small area models are classified into two broad types:
Area level models relate aggregate direct SA estimators toarea-specific covariates. Such models are used if unit (orelement) level data are not available due to privacy or otherreasons. Has advantage to use design weights.Unit level models relate the unit values of a response variableto unit-specific covariates and/or area-specific covariates.
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Frequentist approach to small area estimation
Frequentist approach: empirical best linear unbiasedprediction (EBLUP) or empirical Bayes (EB)
In frequentist approach, the model parameters (such as β, σ2v ,
defined later) are estimated from the marginal distribution ofdata
While frequentist approach is popular, clever approximationsare usually needed for accurate estimates of mean squarederror by accounting for estimation of variance parameters
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Bayesian approach to small area estimation
Bayesian approach: Hierarchical Bayes (HB)
In HB approach, a prior distribution is assigned to the modelparameters
While the HB approach is computationally challenging,sensible inference is obtained from the posterior distribution
Since HB approach automatically accounts for varianceparameter estimation, it has an inherent conceptual advantageover the EB/EBLUP approach
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Longitudinal Model: A Mixed Linear Model
1 Fay-Herriot model (for area-level data) and nested errorregression model (for unit-level data) are special cases ofmixed linear model
Yi = Xiβ + Zivi + ei , i = 1, . . . ,m,
Xi and Zi are known matrices;
2 v = (vT1 , . . . , vTm )T ∼ MVN(0,G (ψ)), independent of
e = (eT1 , . . . , eTm )T ∼ MVN(0,R(ψ)) .
3 R(ψ) = Block Diag(R1(ψ), · · · ,Rm(ψ)).
4 Here ψ denotes the vector of variance components parameters
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Matrix Representation of a Mixed Linear Model
1 Y = (Y T1 , · · · ,Y T
m )T , X = (XT1 , · · · ,XT
m )T (stack the Xi ’s)
2 Z = Block Diag(Z1, · · · ,Zm)⇒ Y = Xβ + Zv + e
3 Y ∼ N(Xβ,Σ(ψ)), Σ(ψ) = R(ψ) + ZG (ψ)ZT
4 R(ψ),G (ψ) are known matrices of ψ
5 If G (ψ) is block diagonal, then Σ(ψ) is block diagonal. Truefor the Fay-Herriot and the Nested Error Regression models.
6 Model with block diagonal variance matrix is termedlongitudinal model by Rao and Molina (2015)
7 Goal: Predict θi = hTi β + λTi v for suitable known hi and λi
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Fay-Herriot Model: A Popular Model for Area-Level Data
1 Frequently only area-level summary statistics are available toestimate θi , ith area population mean, i = 1, · · · ,m
2 Area-level summary θi (or Yi ) directly estimates θi3 Fay and Herriot (1979) proposed their model for Yi
4 Sampling model: Yi = θi + ei , eiind∼ N(0,Di ), i = 1, · · · ,m
5 Known sampling variances Di : D = Diag(D1, · · · ,Dm)
6 Fay-Herriot model connects θi to covariate xi by a
7 Linking model : θi = xTi β + vi , viind∼ N(0, σ2
v ), i = 1, · · · ,m8 A matched model
9 Z = Im, G = σ2v Im, Σ(ψ) = D + σ2
v Im.
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Best Linear Prediction (BLP) for Mixed Linear Models
For the mixed linear model
Y = Xβ + Zv + e,
we first obtained the best linear predictor (BLP) forθi = hTi β + λTi v . To simplify the problem, first assume that β, ψare known. We predict v by its conditional expectation.
1 v |Y = y ∼ N(G (ψ)Σ−1(ψ)(y−Xβ),G (ψ){I−Σ−1(ψ)G (ψ)))
2 BLP of θi : θi ,BP(β, ψ,Y ) = hTi β + sTi (Y − Xβ), withsTi = λTi G (ψ)Σ−1(ψ).
3 Next we estimate unknown β. Assume ψ is known.
4 Marginally, Y ∼ N(Xβ,Σ(ψ))
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Best Linear Unbiased Prediction (BLUP) for θi
1 Estimate β by generalized least squares (GLS) estimator:
2 β(ψ) = (XTΣ−1(ψ)X )−1XTΣ−1(ψ)Y
3 Using β(ψ) in θi ,BP(β, ψ,Y ), leads to BLUP of θi4 θi (ψ) = hTi β(ψ) + sTi (ψ)(Y − X β(ψ))
5 Finally, unknown ψ is estimated from the data by ψ
6 ANOVA, ML, REML and MOM for estimating ψ
7 EBLUP of θi = θi ,EBL = θi (ψ)
8 These estimates of ψ are even functions of OLS residuals
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Details for the Fay-Herriot Model
1 Sampling model: Yi = θi + ei , eiind∼ N(0,Di ), i = 1, · · · ,m
2 Sampling variances Di ’s are assumed known
3 Linking model : θi = xTi β + vi , viind∼ N(0, σ2
v ), i = 1, · · · ,m4 vi |Y = y ∼ N( σ2
vσ2v+Di
(yi − xTi β), σ2v{1− (σ2
v + Di )−1σ2
v})
5 vi |Y = y ∼ N( σ2v
σ2v+Di
(yi − xTi β), σ2vDi
σ2v+Di
= g1i (ψ))
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
BLUP for the Fay-Herriot model
The BP of θi is
θi ,BP(β, ψ,Y ) = xTi β +σ2v
σ2v + Di
(Yi − xTi β)
= Yi − Bi (Yi − xTi β) (1)
BLUP : θi (ψ,Y ) = θi (σ2v ,Y ) = Yi − Bi (Yi − xTi β)
1 BLUP of θi is a WTD Avg. of Yi and synthetic regressionpredictor xTi β, where β = (XTΣ−1(σ2
v )X )−1XTΣ−1(σ2v )Y
2 Bi = Di/(σ2v + Di ) is the shrinkage coefficient, shrinks more
to synthetic regression for large Di
3 Under uniform prior for β, BLUP is the Bayes predictor of θi
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
MSE of BLUP
θi (σ2v )− θi = {θi ,BP(β, ψ,Y )− θi}+ {θi (σ2
v )− θi ,BP(β, ψ,Y )}= t1i + t2i
1 BLUP needs only first two moments
2 t1i and t2i are uncorrelated
3 t2i = {θi (σ2v )− θi ,BP(β, ψ,Y )} = Bix
Ti {β − β}
4 E{t2i}2 = B2i x
Ti (XTΣ−1Xi )
−1xi = g2i (σ2v )
5 MSE(θi (σ2v )) = g1i (σ
2v ) + g2i (σ
2v ) [⊥ decomposition]
6 g1i (σ2v ) measures uncertainty due to predicting vi
7 g2i (σ2v ) measures uncertainty in estimating β
8 In general, g1i (σ2v ) = O(1), g2i (σ
2v ) = O(m−1)
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
MSE of EBLUP and Approximation1 Finally, EBLUP of θi is θi ,EBL = Yi − Bi (σ
2v )(Yi − xTi β)
2 θi ,EBL − θi =
{θi ,BP(β, ψ,Y )− θi}+ {θi (σ2v )− θi ,BP}+ {θi ,EBL − θi (σ2
v )}.3 θi ,EBL − θi = t1i + t2i + t3i
4 If σ2v is an even function of OLS residuals, red, blue and green
terms are uncorrelated under normality.
5 MSE(θi ,EBL) = g1i (σ2v ) + g2i (σ
2v ) + E{θi ,EBL − θi (σ2
v )}2
6 Usually, no simple expression of the green term exists
7 So MSE of EBLUP θi usually has no exact expression
8 PR (1990), LR (1995), DL (2000), DRS (2005) and othersapproximated E{θi ,EBL − θi (σ2
v )}2 accurately up to o(m−1).
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
MSE Approximation and Estimation
1 By Taylor expansion,{θi ,EBL − θi (σ2
v )} ≈ B ′i (σ2v )(σ2
v − σ2v )(Yi − xTi β), and
2 E{θi ,EBL − θi (σ2v )}2 ≈ {B ′i (σ2
v )}2var(σ2v )(Di + σ2
v ) = g3i (σ2v )
3 Note that var(σ2v ) is of the order O(m−1). Thus, the green
term in MSE expression, i.e., g3i (σ2v ) is also of order O(m−1).
4 For any estimator of ψ with bias of the order o(m−1), Prasadand Rao (1990) provided analytical approximation to theestimator of MSE (denoted mse)
5 For other estimators of ψ, the mse estimates are given byDatta and Lahiri (2000), Datta, Rao and Smith (2005)
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Second Order Approximations of MSE1 With σ2
v is an estimator of the variance component
MSE [θi ,EBL] = g1i (σ2v ) + g2i (σ
2v ) + g3i (σ
2v ) + o(m−1),
g3i (σ2v ) =
D2i
(σ2v + Di )4
E (Yi − xTi β)2h(σ2v ) =
D2i
(σ2v + Di )3
h(σ2v ),
2 where h(σ2v ) is the asymptotic variance of σ2
v .
3 An estimator mse, is second order unbiased if
E (mse) = MSE + o(1/m).
4 mse is called an accurate estimator of MSE(θi ,EBL)
5 Accurate estimation of MSE(θi ,EBL) is a major issue in SAE
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Second Order Unbiased Estimation of MSE
If σ2v is an unbiased estimator of σ2
v , such as the Prasad-Raoestimator, then
E [g1i (σ2v )] = g1i (σ
2v )−g3i (σ
2v ) + o(1/m).
Also, if σ2v is asymptotically unbiased,
E [g2i (σ2v )] = g2i (σ
2v ) + o(1/m), E [g3i (σ
2v )] = g3i (σ
2v ) + o(1/m).
With an unbiased estimator σ2v , a second order unbiased estimator
of MSE is
mse(σ2v ) = g1i (σ
2v ) + g2i (σ
2v ) + 2g3i (σ
2v ).
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Various methods of variance estimation
1 ANOVA or Henderson’s method (Prasad and Rao, 1990):
2 σ2vPR = (m − p)−1[
∑mi=1 u
2i −
∑mi=1 Di (1− hii )],
σ2vPR = max{σ2
vPR , 0}3 ui OLS residual and hii = xTi (XTX )−1xi .
4 σ2vPR is an unbiased estimator (Normality not needed). Then
msePR(σ2vPR) = g1i (σ
2vPR) + g2i (σ
2vPR) + 2g3i (σ
2vPR)
5 Note: Instead of truncating σ2vPR at a lower bound zero, we
suggested using a larger bound
LF =
√2DH√m
(due to Fuller).
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Other estimators of σ2v : Fay-Herriot method
1 PR (1990) estimator of σ2v is subject to large variability
2 Note: E [Y T{Σ−1 − Σ−1X (XTΣ−1X )−1XTΣ−1}Y ] = m − p
3 FH (1979) suggested the estimating equation
4 Q(σ2v )
def=
∑mi=1
(Yi−xTi β)2
σ2v+Di
= m − p [Estimating equation]
5 Q(σ2v ) is a decreasing function of σ2
v with Q(σ2v ) ↓ 0
6 FH estimator of σ2v is
σ2v ,FH = I (Q(0) > m − p)Q−1(m − p) + I (Q(0) ≤ m − p)LF .
7 The estimator is biased.
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
ML and REML methods
1 DL (2000) discussed ML and REML in SAE
2 Likelihood function: L(β, σ2v ) = |Σ|−
12 exp[−1
2
∑mi=1
(Yi−xTi β)2
σ2v+Di
]
3 Profile likelihood function: LP(σ2v ) = L(β, σ2
v )
4 MLE of σ2v : σ2
v ,ML maximizes LP(σ2v )
5 Likelihood equation for σ2v :
m∑i=1
[(Yi − xTi β)2
(σ2v + Di )2
− 1
σ2v + Di
] = 0
6 DL(2000) showed that σ2v ,ML is negatively biased
7 DL(2000) showed that bias of σ2v ,RE is o(m−1);
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Fay-Herriot model under non-normality
Yi = θi + ei , θi = xTi β + vi , i = 1, · · · ,mE (ei ) = 0, V (ei ) = Di , E (vi ) = 0, V (vi ) = σ2
v
Error terms (sampling and model) ei ’s and vi ’s are allassumed uncorrelated
Under moments assumptions, EBLUPs of θi can be derived
The MSE depends on the fourth moments of ei ’s, usually notavailable from the area summary statistics
Assume sampling errors approximately normal, may be okay
Lahiri & Rao (1995) derived MSE with non-normal vi ’s
[Bayesian version of the problem was considered by Datta &Lahiri (1995) with normal mixture]
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
EBLUP and MSE
Lahiri & Rao estimated σ2v by PR estimator and β by GLS
Prediction error: θi ,EBL − θi =
{θi ,BP(β, ψ,Y )− θi}+ {θi (σ2v )− θi ,BP}+ {θi ,EBL − θi (σ2
v )}.θi ,EBL − θi = t1i + t2i + t3i
t1i & t2i are uncorrelated; corr(t2i , t3i ) = O(m−1)
MSE (θi ,EBL) = g1i + g2i + E [t23i ] + 2E [t1i t3i ] + o(m−1)
The last two terms,O(m−1), depend on σ2v and µ4,v = E (v4
i )
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Estimation of MSE
LR (1995) showed that the analytic estimator
msePR(σ2v ,PR) = g1i (σ
2v ,PR) + g2i (σ
2v ,PR) + 2g3i (σ
2v ,PR)
is an accurate estimator of MSE even under non-normality
This implies robustness of msePR under non-normality
Hall and Maiti (2006) considered non-normal SAE prediction
They considered for unit-level data the NER model undermoments assumption for the error terms eij ’s and vi ’s
The MSE approximation depends on E (e4ij) = µ4,e and µ4,v
They proposed bootstrap estimation of the MSE
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Estimation equations for non-normal FH model
Estimating equation for β: GLS
m∑i=1
yi − xTi β
Di + σ2v
xi = 0
Estimating equation for σ2v :
m∑i=1
{(yi − xTi β)2
(Di + σ2v )a− (Di + σ2
v )1−a} = 0
(i) a = 0 corresponds to usual MOM (PR version)(ii) a = 1 corresponds to FH version(iii) a = 2 corresponds to ML (for normality) version
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
MSE approximation for EBLUP
Let φ = (σ2v , µ4,v )
MSE (θi ,EBL) = g1i + g2i + g3i + g4i + o(m−1)
E (t21i ) = g1i (σ
2v )
E (t22i ) = g2i (σ
2v ) = O(m−1)
E [t23i ] = g3i (φ) + o(m−1)
2E [t1i t3i ] = g4i (φ) + o(m−1)
g3i = O(m−1), g4i = O(m−1)
MSE (θi ,EBL) = g1i (σ2v ) + gi (φ) + o(m−1)
gi (φ) = O(m−1)
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Bootstrap bias correction
Pfeffermann and Corea (2012) suggested parametricbootstrap to estimate MSE
Hall and Maiti (2006) suggested nonparametric bootstrap
From the property of σ2v we know
E [g1i (σ2v )] = g1i (σ
2v ) + bi (φ) + o(m−1)
We estimate bi (φ) by bootstrapping
MSE approximation depends on 2nd and 4th moments of vi
Estimate µ4,v from the residuals
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Bootstrap estimation of MSE
1 Bootstrap from a population matching estimated 2nd and 4thmoments of vi
2 As in Hall and Maiti, we generate v∗i from a 3−pointdistribution
3 Generate e∗i from N(0,Di )
4 Based on the estimator φ and FH model, we generate F sets
of {θ(f )i : i = 1, · · · ,m}, f = 1, · · · ,F and datasets
{Y (f ) = (Y(f )1 , · · · ,Y (f )
m )T : f = 1, · · · ,F}
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Bootstrap estimation of MSE (continued)
1 Corresponding to the f th dataset we obtain φ(f ) and EBLUP
θ(f )i ,EBL of θ
(f )i . Based on the F generated sets of data and
small area means, we define an empirical MSE of θi ,EBL by
m(F )i =
1
F
F∑f =1
(θ(f )i ,EBL − θ
(f )i )2. (2)
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Bootstrap estimation of MSE (continued)
1 Eφ[m(F )i ] = g1i (φ) + bi (φ) + gi (φ) + o(m−1)
2 Eφ[g1i (φ(f ))] = g1i (φ) + 2bi (φ) + o(m−1)
3 m(F )i − { 1
F
∑Ff =1 g1i (φ
(f ))− g1i (φ)} is a second-order
unbiased estimator of the MSE(θi ,EBL(Yi ; φ)).
4 This estimator uses the known form of g1i (φ)
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Simulation results
We compare our bootstrap method of estimating MSE withLahiri-Rao method
Following Lahiri-Rao we use three distributions (normal,exponential, double exponential), two sample sizes (m=20,m=30) and two sampling variance patterns (Pattern (a) withσ2v = .2 and four groups of sampling variances, and Pattern
(b) with σ2v = 1 and five groups of sampling variances)
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Empirical MSE x 102: m=20, Pattern (a), σ2v = 0.2
Di 0.5 0.33 0.25 0.2γi 0.29 0.375 0.44 0.50
NormalLR(R) 16.9 14.8 13.1 11.7a=0 16.9 14.8 13.1 11.8a=1 16.8 14.7 13.0 11.7
Double ExponentialLR(R) 16.4 14.4 12.7 11.3a=0 16.3 14.3 12.7 11.4a=1 16.3 14.3 12.6 11.3
Shifted ExponentialLR(R) 16.3 14.0 12.4 11.0a=0 16.1 14.0 12.4 11.1a=1 16.1 13.9 12.3 11.0Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Empirical MSE x 102: m=20, Pattern (b), σ2v = 1
Di 0.5 0.33 0.25 0.2γi 0.29 0.375 0.44 0.50
NormalLR(R) 55.1 44.4 36.4 30.9a=0 55.2 44.8 37.0 31.6a=1 55.2 44.6 36.7 31.2
Double ExponentialLR(R) 53.49 43.0 35.4 30.3a=0 53.3 43.2 35.9 30.9a=1 53.62 43.2 35.6 30.5
Shifted ExponentialLR(R) 52.7 41.9 34.6 29.6a=0 52.4 42.2 35.1 30.2a=1 52.9 42.1 34.8 29.7Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Empirical MSE x 102: m=30, Pattern (a), σ2v = 0.2
Di 0.5 0.33 0.25 0.2 0.14Normal
LR 16.3 14.4 12.7 11.6 10.5LR(R) 15.9 13.9 12.5 11.2 10.3a=0 16.0 14.0 12.6 11.3 10.5a=1 15.9 13.9 12.5 11.2 10.3
Double ExponentialLR 15.8 14.0 12.5 11.2 10.4
LR(R) 15.7 13.7 12.2 11.0 10.0a=0 15.7 13.7 12.4 11.2 10.2a=1 15.7 13.7 12.2 11.0 10.0
Shifted ExponentialLR 15.8 13.7 12.2 11.1 10.0
LR(R) 15.7 13.5 12.0 10.8 9.8a=0 15.5 13.5 12.1 11.0 10.0a=1 15.6 13.5 12.0 10.8 9.8
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Empirical MSE x 102: m=30, Pattern (b), σ2v = 1
Di 1 0.67 0.5 0.4 0.33γi 0.5 0.6 0.67 0.71 0.75
NormalLR 53.6 42.8 35.4 30.3 26.3
LR(R) 53.0 42.3 35.3 30.0 26.0a=0 53.1 42.5 35.5 30.2 26.3a=1 53.0 42.4 35.4 30.0 26.1
Double ExponentialLR 51.9 41.9 34.9 29.6 26.2
LR(R) 52.2 41.6 34.7 29.7 25.6a=0 52.0 41.7 35.0 30.0 26.9a=1 52.3 41.7 34.8 29.8 25.7
Shifted ExponentialLR 51.6 41.0 34.1 29.4 25.5
LR(R) 51.7 41.0 34.1 29.4 25.3a=0 51.3 41.0 34.4 29.8 25.7a=1 51.8 41.1 34.2 29.4 25.3
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Relative Bias: m=20, Pattern (a), σ2v = 0.2
Di 0.5 0.33 0.25 0.2γi 0.29 0.375 0.44 0.50
NormalLR 5.0 9.0 12.0 17.0
LR(R) 9.6 12.8 17.1 20.5B(a=0) -9.7 -11.4 -11.3 -11.5
Double ExponentialLR 6.0 11.0 14.0 21.0
LR(R) 12.4 15.8 20.4 24.6B(a=0) -7.9 -10.7 -10.9 -11.1B(a=1) -8.3 -10.8 -11.2 -11.4
Shifted ExponentialLR 7.0 13.0 16.0 23.0
LR(R) 13.5 18.4 23.3 27.8B(a=0) -7.8 -9.6 -9.9 -10.2B(a=1) -8.1 -9.7 -10.3 -10.7
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Relative Bias: m=20, Pattern (b), σ2v = 1
Di 1 0.67 0.5 0.4γi 0.5 0.6 0.67 0.71
NormalLR -0.2 0.9 0.4 2.3
LR(R) -1.1 -0.9 0.6 1.9B(a=0) -2.9 -3.9 -3.1 -2.5B(a=1) -2.7 -3.6 -2.8 -2.2
Double ExponentialLR -0 1.6 1.2 3.0
LR(R) -0.6 0.3 2.3 3.4B(a=0) -3.5 -4.7 -4.2 -4.4B(a=1) -3.8 -4.6 -4.1 -4.0
Shifted ExponentialLR 0.2 2.0 2 4.5
LR(R) -0.8 1.6 3.7 5.2B(a=0) -4.9 -5.5 -5.4 -5.3B(a=1) -5.1 -5.3 -5.1 -5.1
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Relative Bias: m=30, Pattern (a), σ2v = 0.2
Di 0.5 0.33 0.25 0.2 0.14γi 0.29 0.375 0.44 0.50 0.54
NormalLR 1.6 2.7 3.5 5.0 5.0
LR(R) 2.7 4.5 5.3 7.4 8.3B(a=0) -6.5 -7.4 -8.6 -8.3 -8.9B(a=1) -5.6 -6.7 -7.9 -7.8 -8.3
Double ExponentialLR 2.6 3.0 4.0 6.0 7.0
LR(R) 3.5 5.3 7.0 8.8 11.1B(a=0) -6.2 -7.6 -8.7 -9.2 -8.8B(a=1) -5.6 -7.1 -8.0 -8.5 -8.2
Shifted ExponentialLR 2.1 4.0 4.5 9.0 9.5
LR(R) 3.5 6.5 8.8 10.0 13.1B(a=0) -6.9 -7.7 -8.4 -10.0 -9.0B(a=1) -6.5 -7.5 -8.2 -9.6 -8.8
Gauri Sankar Datta SAE under Non-normal Random Effects
Small area models as hierarchical models, random effects models and mixed linear modelsMixed linear models in SAE
Empirical Best Linear Unbiased PredictionApproximation of MSE of EBLUP
MSE of EBLUP under non-normalityMSE estimation without normality
Simulation study
Relative Bias: m=30, Pattern (b), σ2v = 1
Di 1 0.67 0.5 0.4 0.33γi 0.5 0.6 0.67 0.71 0.75
NormalLR 0.1 0.6 0.2 0.8 0
LR(R) -0.8 -0.2 -0.6 0.2 0B(a=0) 0.1 0.2 -0.3 0.6 0.2B(a=1) 0.0 0.4 -0.01 0.5 0.2
Double ExponentialLR 0.9 0 0.4 0.6 0.1
LR(R) -1.3 -0.4 -0.4 0 0.9B(a=0) -0.6 -0.6 -1.2 -1.0 -0.4B(a=1) -0.6 -0.4 -1.0 -1.0 -0.3
Shifted ExponentialLR -0.4 0.4 0.8 0.4 1.7
LR(R) 1.8 0 0.3 0.2 1.5B(a=0) -1.6 -1.1 -1.6 -2.4 -1.5B(a=1) -1.5 -0.9 -1.4 -2.1 -1.4
Gauri Sankar Datta SAE under Non-normal Random Effects