REVSTAT – Statistical Journal

Volume 19, Number 2, April 2021, 275–290

VARIANCE ESTIMATION
IN THE PRESENCE OF MEASUREMENT ERRORS
UNDER STRATIFIED RANDOM SAMPLING

Authors: Neha Singh

– Department of Mathematics & Computing, Indian Institute of Technology (ISM),
Dhanbad-826004, Jharkhand, India
nehaismstats@ gmail.com
Gajendra K. Vishwakarma
– Department of Mathematics & Computing, Indian Institute of Technology (ISM),
Dhanbad-826004, Jharkhand, India
vishwagk@ rediffmail.com
Raj K. Gangele
– Department of Mathematics and Statistics, Dr. Harisingh Gour University
(Central University),
Sagar-470003, M.P., India
rkgangele23@ gmail.com

Received: August 2018 Revised: April 2019 Accepted: April 2019

Abstract:
• This study focuses on the estimation of population variance of study variable in stratified random
sampling using auxiliary information when the observations are contaminated by measurement
errors. Three classes of estimators of variance under measurement error are proposed by using the
approach of Srivastava and Jhajj [18] for the study variable. The properties of the estimator viz.
bias and mean square error of the proposed classes of estimators are provided. The conditions
for which proposed estimators are more efficient compared to usual estimators are discussed. It is
shown that the proposed classes of estimators include a large number of estimators of the population
variance of stratified random sampling and their bias and mean square error can be easily derived.

Keywords:

• variance; auxiliary variable; mean square error; measurement error; stratified random sampling.

AMS Subject Classification:

• 49A05, 78B26.
276 N. Singh, G.K. Vishwakarma and R.K. Gangele

1. INTRODUCTION

In survey sampling, the auxiliary information is mainly used in order to gain efficiency
for the estimation. The literature on estimating the population variance by using auxiliary
variable is substantial and widely discussed. Some authors including, Das and Tripathi [4],
Srivastava and Jhajj [18], Isaki [5], Upadhyay and Singh [19, 20], Singh et al. [13], Prasad and
Singh [8], Biradar and Singh [2], Singh and Biradar [11] have paid their attention towards
the estimation of population variance of study variable using auxiliary information in simple
random sampling. While dealing with planning surveys, in case of heterogeneous population,
stratified random sampling has more importance in precise estimates over the simple random
sampling. Singh and Vishwakarma [14] discussed a general method for the estimation of the
variance of the stratified random sample mean by using auxiliary information.

The theories of survey sampling assume that the observations recorded during data
collection are always free from measurement error. However, this assumption does not meet
in many real-life situations and the data is contaminated with errors. The mean square error
and other properties of the estimator obtained with significant measurement error may lead
to serious fallacious results. Cochran [3], has discussed the source of measurement errors in
survey data. Many authors such as Shalabh [9], Srivastava and Shalabh [16], Maneesha and
Singh [7], Allen et al. [1], Shalabh and Tsai [10], Singh and Vishwakarma [15] have studied
the impacts of measurement errors in the ratio, product and regression methods of estimation
under simple random sampling.

Let us consider a finite heterogeneous population of size N , partitioned into L non-

overlapping strata of sizes Nh , h = 1, 2, ..., L, where L
P
h=1 Nh = N . Let (yhj , xhj ) be the pair
of observed values instead of true pair values (Yhj , Xhj ) of the study character y and the
auxiliary character x respectively of the j-th unit (j = 1, 2, ..., Nh ) in the h-th stratum. Also,
let (yhj , xhj ) be the pair of values on (y, x) drawn from the h-th stratum (j = 1, 2, ..., nh ;
h = 1, 2, ..., L). It is familiar that in stratified random sampling an unbiased estimator of the
population mean (µY = L Nh
P
h=1 Wh µY h ; Wh = N ) of the variable y is given by

L
X
(1.1) ȳst = Wh ȳh ,
h=1

where ȳh = n1h nj=1

P h 1 PNh
yhj is the sample mean of h-th stratum and µY h = Nh j=1 yhj is the
population mean of h-th stratum. Let the observational errors be

(1.2) uhj = yhj − Yhj , vhj = xhj − Xhj ,

which are normally distributed with mean zero and variances σuh 2 and σ 2 respectively. Also
vh
let ρh be the population correlation coefficient between Y and X in h-th stratum. For
simplicity in exposition, it is assumed that the variables uhj and vhj are uncorrelated although
(Yhj , Xhj ) are correlated.

To obtain the bias and mean square error we define

σ̂Y2 h = σY2 h (1 + ε0h ), 2

σ̂Xh 2
= σXh (1 + ε1h ), x̄h = µXh (1 + ε2h ),

such that E(εih ) = 0, ∀i = 0, 1, 2;

Variance Estimation in the Presence of Measurement Errors... 277

2
CXh
AY h AXh 1
E(ε20h ) = , E(ε21h ) = , E(ε22h ) = , E(ε0h ε1h ) = (δ22h − 1),
nh nh nh θXh nh
1 1
E(ε1h ε2h ) = (δ03h CXh ), E(ε0h ε2h ) = (δ21h CXh ),
nh nh
where
σU4 h 2 µ40h σXh
AY h = γ2Y h + γ2U h + , β2h (Y ) = δ40h = , CXh = ,
σY4 h θY2 h µ220h µXh
σV4 h 2 µ04h
AXh = γ2Xh + γ2V h 4 + 2 , β2h (X) = δ04h = ,
σXh θXh µ202h

γ2Y h = β2h (Y ) − 3, γ2Xh = β2h (X) − 3, γ2U h = β2h (U ) − 3,

σY2 h 2
σXh
γ2V h = β2h (V ) − 3, θY h = , θXh = 2 + σ2 ,
σY2 h + σU2 h σXh Vh
Nh
µrsh 1 X
δrsh =
1 , µrsh = (yhj − µY h )r (xhj − µXh )s .
µr20h µs02h 2 Nh
j=1

(r, s) are positive integers, µY h and µXh are the h-th stratum population mean of study and
auxiliary variable respectively. CXh is the coefficient of variation of h-th stratum, θY h and
θXh are the reliability ratio of h-th stratum of study and auxiliary variable respectively and
lying between zero and one.

The variance of the stratified random sample mean is given by

L
X σY2 h
(1.3) V(ȳst ) = Wh2 2
= σst ,
nh
h=1

1 PNh
where σY2 h = Nh j=1 (yij − µ̄Y h )2 is the population variance of h-th stratum.

2 , i.e. V(ȳ ), is given by

The unbiased estimator of σst st

L
2
X s2
2 yh
(1.4) σ̂st = Wh ,
nh
h=1
Pnh
where s2yh = (nh1−1) j=1 (yhj − ȳh )2 is an unbiased estimator of σst 2 . But in the presence of

measurement error s2yh is not an unbiased estimator for σst 2 . In the measurement error case
2
2 σ̂Y h
2 is given by σ̂ 2 =
PL 2 2 2
the unbiased estimator of σst st h=1 Wh nh , where σ̂Y h = (syh − σuh ).

2 in the presence of measurement error is given by

The variance of σ̂st
L
2
X (Wh σY h )4 2
(1.5) V(σ̂st ) = [AY h ] = MSE(σ̂st ).
n3h
h=1

Singh and Karpe [12] have studied the impact of measurement error on separate ratio
and product also combined ratio as well as product estimators for the population mean
under stratified random sampling. We have considered the problem of estimating population
variance using information on the auxiliary variable by adopting Srivastava and Jhajj [18]
278 N. Singh, G.K. Vishwakarma and R.K. Gangele

method in stratified random sampling in the presence of measurement error. Three classes
of estimators for the estimation of population variance are proposed under stratified random
sampling when both the study and auxiliary variables are commingled with measurement
errors as:
2 when the mean µ
i) Estimator of variance σst Xh of the auxiliary variable x in the
h-th stratum of the population is known.
2 when the variance σ 2
ii) Estimation of variance σst Xh of the auxiliary variable x in
the h-th stratum of the population is known.
iii) Estimation of variance σst2 when the mean µ 2
Xh and the variance σXh of the aux-
iliary variable x in the h-th stratum of the population are known.

The crux of this study is to exhibit the effect of measurement errors on the estimates of the
variance of the stratified random sample mean while using auxiliary information.

2. THE PROPOSED CLASSES OF ESTIMATORS

2 of the stratified simple random sample

2.1. Estimation of population variance σst
mean when mean µXh of h-th stratum of the auxiliary variable x in the
population is known

By using information on population mean µXh of the h-th stratum of auxiliary variable,
2 for the study variable is proposed as
a class of estimators of population variance σst

L 
W2
X 
(2.1) σ̂a2 = h
σ̂Y2 h ah (lh ),
nh
h=1

where lh = x̄h /µXh and ah (·) is a function of lh such that ah = 1. It satisfies conditions given
by Srivastava [17] viz. function are continuous and bounded also the first as well as second
order partial derivatives of the function exist. Expanding the function about the point ‘unity’
in a second order Taylor’s series, we have

1
(2.2) ah (lh ) = ah (1) + (lh − 1)a1h (1) + (lh − 1)2 a2h (1),
2

where a1h , a2h are first order and second order derivative with respect to lh about point unity.

L 
W2
  
X 1
σ̂a2 = h
σY2 h (1 2
+ ε0h ) 1 + (lh − 1)a1h (1) + (lh − 1) a2h (1) ,
nh 2
h=1
(2.3) L 
Wh2 2
  
X 1
σ̂a2 = σY h (1 + ε0h ) 1 + ε2h a1h (1) + ε22h a2h (1) ,
nh 2
h=1
Variance Estimation in the Presence of Measurement Errors... 279

L 
W2
X  
(σ̂a2 − 2
σst ) = h
(σY2 h ) ε0h + ε2h a1h (1) + ε0h ε2h a1h (1)
nh
(2.4) h=1

1 2 1 2
+ ε2h a2h (1) + ε0h ε2h a2h (1) .
2 2

Taking expectation on both sides of (2.4) we get

L 
W2 2
  
X 1 CXh
(2.5) Bias(σ̂a2 ) = h
σY2 h δ21h CXh a1h (1) + a2h (1) .
n2h 2 θXh
h=1

For the mean square error we have

L 
W2
X  n o2
(2.6) (σ̂a2 − 2 2
σst ) = h
σY4 h ε0h + ε2h a1h (1) ,
nh
h=1

L 
W2 2
X  n o
(2.7) (σ̂a2 − 2 2
σst ) = h
σY4 h ε20h + ε22h a21h (1) + 2ε0h ε2h a1h (1) .
nh
h=1

Taking expectation up to terms of order n−3 2

h , we get the mean square error of σ̂a as

L 2
(Wh σY h )4
 
X CXh
(2.8) MSE(σ̂a2 ) = AY h + 2
a (1) + 2δ21h CXh a1h (1) .
n3h θXh 1h
h=1

The MSE in (2.8) is minimized for

 
δ21h θXh
(2.9) a1h (1) = − .
CXh

Thus, the resultant minimum MSE of σ̂a2 is given by

L
X (Wh σY h )4 h i
(2.10) min.MSE(σ̂a2 ) = 2
AY h − δ21h θXh .
n3h
h=1

Hence, a theorem can be established as follows.

Theorem 2.1. Up to terms of the order n−3

h ,

L
X (Wh σY h )4 h i
min.MSE(σ̂a2 ) ≥ 2
AY h − δ21h θXh ,
n3h
h=1

 
with equality holding if a1h (1) = − δ21h
CXh
θXh
.
280 N. Singh, G.K. Vishwakarma and R.K. Gangele

The following estimators

L  L 
Wh2 2 α1h Wh2 2 h
X  X  i
2
σ̂a1 = σ̂Y h lh , 2
σ̂a2 = σ̂Y h 2 − lhα1h ,
nh nh
h=1 h=1
L  L 
Wh2 2 Wh2 2 h
   
2
X α1h + lh 2
X i
σ̂a3 = σ̂Y h , σ̂a4 = σ̂Y h α1h + (1 − α1h )lh ,
nh 1 + α1h lh nh
h=1 h=1
L 
Wh2 2 h
X  i
2
σ̂a5 = σ̂Y h α1h + (1 − α1h )lh−1 ,
nh
h=1
L 
Wh2 2 h
 i
σ̂Y h α1h + (1 − α1h )lh−α2h ,
X
2
σ̂a6 =
nh
h=1
L 
Wh2 2 h
X  i−1
2
σ̂a7 = σ̂Y h α1h + (1 − α1h )lh ,
nh
h=1

are some of the members of the proposed class of estimators σ̂a2 . The optimum values of the
scalars α1h and α2h can be derived from the right-hand side (2.9) of and the minimum mean
square error of the listed estimators can be derived from (2.8). The lower bound of the MSE
2 , (i = 1 to 7) is the same as given by (2.10).
of estimators σ̂ai

Following by [17] and Srivastava and Jhajj [18] we have proposed a wider class of
2 as
estimators of σst
L 
Wh2
X 
2
(2.11) σ̂D = Dh (σ̂Y2 h , lh ),
nh
h=1

where function Dh (·, ·) satisfies

∂Dh (·)
Dh (σY2 h , 1) = σY2 h ⇒ D1h (σY2 h , 1) = | 2 = 1.
∂ σ̂Y2 h (σY h ,1)
It can be shown that the minimum MSE of σ̂D 2 and the minimum MSE of σ̂ 2 are equal.
a
We can state that the difference type estimator
L 
Wh2 n 2
X  o
2
(2.12) σ̂std1 = σ̂Y h + d1h (lh − 1) ,
nh
h=1

is a member of class 2
σ̂D where d1h is a suitably chosen constant.

2 of the stratified simple random sample

2.2. Estimation of population variance σst
2
mean when variance σXh of h-th stratum of the auxiliary variable x in the
population is known

2 of the stratified simple random sample mean

A class of estimators of the variance σst
2
when the variance σXh of the auxiliary variable x of the h-th stratum in the population is
known, is defined as
L 
Wh2 2
X 
2
(2.13) σ̂b = σ̂Y h bh (mh ),
nh
h=1
Variance Estimation in the Presence of Measurement Errors... 281

σ̂ 2
where mh = σXh2 , and bh (mh ) is a function of mh such that bh (1) = 1. The function is
Xh
continuous and bounded in R and its first as well as the second order partial derivatives
exist. Now expanding the function at point ‘unity’ in a second order Taylor’s series, we can
write
1
(2.14) bh (mh ) = bh (1) + (mh − 1)b1h (1) + (mh − 1)2 b2h (1),
2
where b1h (1) and b2h (1) are the first order and second order derivative with respect to mh of
the function bh (mh ) about the point ‘unity’.
L 
W2
  
X 1
(2.15) σ̂b2 = h
σ̂Y2 h 2
bh (1) + (mh − 1)b1h (1) + (mh − 1) b2h (1) ,
nh 2
h=1
L 
W2
  
X 1 2
(2.16) σ̂b2 = h
σY2 h (1 + ε0h ) 1 + ε1h b1h (1) + ε1h b2h (1) .
nh 2
h=1

To calculate the bias and the MSE of the estimator we can write
L 
Wh2 2
X  
2 2
(σ̂b − σst ) = σY h ε0h + ε1h b1h (1) + ε0h ε1h b1h (1)
nh
(2.17) h=1

1 1
+ ε21h b2h (1) + ε0h ε21h b2h (1) .
2 2

Taking expectation on both sides of (2.17) we get the bias of σ̂b2 as

L 
W2
  
X 1
(2.18) Bias(σ̂b2 ) = h
σY2 h (δ22h − 1)b1h (1) + AXh b2h (1) .
n2h 2
h=1

For the mean square error we have

L 
W2 2
X  h i2
(2.19) (σ̂b2 − 2 2
σst ) = h
σY4 h ε0h + ε1h b1h (1) ,
nh
h=1
L 
W2 2
X  h i
(2.20) (σ̂b2 − 2 2
σst ) = h
σY4 h ε20h + ε21h b21h (1) + 2ε0h ε1h b1h (1) .
nh
h=1

Taking expectation up to terms of order n−3 2

h , we get the mean square error of σ̂b as

L
X (Wh σY h )4 h i
(2.21) MSE(σ̂b2 ) = AY h + AXh b21h (1) + 2(δ22h − 1)b1h (1) ,
n3h
h=1

which is minimized for

 
δ22h − 1
(2.22) b1h (1) = − .
AXh

Thus, the resultant minimum MSE of σ̂b2 can be written as:

L
(Wh σY h )4 (δ22h − 1)2
X  
(2.23) min.MSE(σ̂b2 ) = AY h − .
n3h AXh
h=1

Hence, a theorem can be established as follows.

282 N. Singh, G.K. Vishwakarma and R.K. Gangele

Theorem 2.2. Up to terms of order n−3

h ,

L
(Wh σY h )4 (δ22h − 1)2
X  
min.MSE(σ̂b2 ) ≥ AY h − ,
n3h AXh
h=1

 
−1
with equality holding if b1h (1) = − δ22h
AXh .

The listed estimators

L  L 
W2 W2
  h i
σ̂Y2 h mηh1h , σ̂Y2 h 2 − mηh1h ,
X X
2 h 2 h
σ̂b1 = σ̂b2 =
nh nh
h=1 h=1
L  L 
Wh2 2 Wh2 2 h
   
2
X η1h + mh 2
X i
σ̂b3 = σ̂Y h , σ̂b4 = σ̂Y h η1h + (1 − η1h )mh ,
nh 1 + η1h mh nh
h=1 h=1
L
X W  2  h i
2
σ̂b5 = h
σ̂Y2 h η1h + (1 − η1h ) m−1
h ,
nh
h=1
L 
Wh2 2 h
 i
σ̂Y h η1h + (1 − η1h )mηh2h ,
X
2
σ̂b6 =
nh
h=1
L 
Wh2 2 h
X  i−1
2
σ̂b7 = σ̂Y h η1h + (1 − η1h )mh ,
nh
h=1

are some of the members of the proposed class of estimators σ̂b2 . The optimum values of the
2 to σ̂ 2 can be derived from (2.22) and the minimum mean square
scalars η1h and η2h from σ̂b1 b7
errors of each of the listed estimators can be derived from (2.21). The lower bound of the
MSE of the estimators σ̂bi2 (i = 1 to 7) is given by (2.23).

2 than σ̂ 2 is
A wider class of estimators of σst b

L 
W2
X 
(2.24) σ̂e2 = h
eh (σ̂Y2 h , mh ),
nh
h=1

where eh (σ̂Y2 h , mh ) is a function of (σ̂Y2 h , mh ) and

∂eh (·)
eh (σY2 h , 1) = σY2 h ⇒ e1h (σY2 h ) = 1 with e1h (σY2 h , 1) = | 2 .
∂ σ̂Y2 h (σY h ,1)

It can be exhibited that up-to third order, the optimum mean square error of σ̂e2 and
σ̂b2 is the same as given by (2.23). It can also be shown that the difference-type estimator

L 
Wh2 n 2
X  o
2
(2.25) σ̂std2
= σ̂Y h + d2h (mh − 1)
nh
h=1

is a specific member of the class of estimator σ̂e2 but not of the σ̂b2 class, where d2h is an
appropriately chosen constant.
Variance Estimation in the Presence of Measurement Errors... 283

2.3. Estimation of population variance σst2 of the stratified simple random sample

mean when the population mean µXh and the variance σXh 2 of h-th stratum
of the auxiliary variable x in the population are known

We define a class of estimators σ̂c2 , for the estimation of variance of the stratified simple
random sample mean when the mean µXh and the variance σXh 2 of the auxiliary variable x
of the h-th stratum in the population are known, as
L 
W2
X 
(2.26) σ̂c2 = h
σ̂Y2 h ch (lh , mh ),
nh
h=1

where ch (lh , mh ) is a function of lh = x̄h /µXh and mh = σ̂Xh 2 /σ 2 , such that c (1, 1) = 1 and
Xh h
it also satisfies similar conditions as mentioned in [18]:

ch (lh , mh ) = ch (1, 1) + (lh − 1)c1h (1, 1) + (mh − 1)c2h (1, 1)
1n o
+ (lh − 1)2 c11h (1, 1) + 2(lh − 1)(mh − 1)c12h (1, 1) + (mh − 1)2 c22h (1, 1)
2
(2.27)
1n
+ (lh − 1)3 c111h (lh∗ , m∗h ) + 3(lh − 1)2 )(mh − 1)c112h (lh∗ , m∗h )
6
o
2 ∗ ∗ 3 ∗ ∗
+ 3(lh − 1)(mh − 1) c122h (lh , mh ) + (mh − 1) c222h (lh , mh ) ,

where

lh∗ = 1 + φ(lh − 1), m∗h = 1 + φ(mh − 1), 0 < φ < 1;

n o
c1h (1, 1), c2h (1, 1)}, {c11h (1, 1), c12h (1, 1), c22h (1, 1) ,
n o
c111h (lh∗ , m∗h ), c112h (lh∗ , m∗h ), c122h (lh∗ , m∗h ), c222h (lh∗ , m∗h ) ,

respectively denote the first, second and third order partial derivatives of the function ch (lh , mh ).
Expressing (2.27) in terms of ε0h , ε1h and ε2h using ch (1, 1) = 1 we have
L 
W2
X  n o
σ̂c2 = h
σY2 h (1 + ε0h ) 1 + ε2h c1h (1, 1) + ε1h c2h (1, 1)
nh
h=1
1n 2 o
+ ε2h c11h (1, 1) + 2ε1h ε2h c12h (1, 1) + ε21h c22h (1, 1)
(2.28) 2
1n 3
+ ε c111h (lh∗ m∗h ) + 3ε1h ε22h c112h lh∗ m∗h
6 2h
o
2 ∗ ∗ 3 ∗ ∗
+ 3ε1h ε2h c122h lh mh + ε1h c222h (lh mh ) .

To calculate the bias and the MSE of the estimator we can write (2.28) as
L 
W2
X  n
σ̂c2 − 2
σst = h
σY2 h ε0h + ε2h c1h (1, 1) + ε1h c2h (1, 1)
nh
h=1
o
(2.29) + ε0h ε2h c1h (1, 1) + ε0h ε1h c2h (1, 1)
1n 2 o
2
+ ε c11h (1, 1) + 2ε1h ε2h c12h (1, 1) + ε1h c22h (1, 1) .
2 2h
284 N. Singh, G.K. Vishwakarma and R.K. Gangele

Taking expectation of both sides of (2.29) we get the bias of the estimator σ̂c2 :
L 
"
X W 2
Bias(σ̂c2 ) = h
σY2 h δ21h CXh c1h (1, 1) + (δ22h − 1)c2h (1, 1)
n2h
h=1
(2.30)  2 #
1 CXh
+ c11h (1, 1) + 2δ03h CXh c12h (1, 1) + AXh c22h(1,1) .
2 θXh

For the mean square error we have

L 
W2 2
X  h i2
(2.31) (σ̂c2 − 2 2
σst ) = h
σY4 h ε0h + ε2h c1h (1, 1) + ε1h c2h (1, 1) ,
nh
h=1

L 
W2 2
X  h
(σ̂c2 − 2 2
σst ) = h
σY4 h ε20h + ε22h c21h (1, 1) + ε21h c22h (1, 1)
(2.32) nh
h=1
i
+ 2ε0h ε2h c1h (1, 1) + 2ε1h ε2h c1h (1, 1)c2h (1, 1) + 2ε0h ε1h c2h (1, 1) .

Taking expectation up-to order n−3 2

h , we get the mean square error of σ̂c as

L 2
(Wh σY h )4

X CXh
MSE(σ̂c2 ) = A Yh + c2 (1, 1) + AXh c22h (1, 1)
n3h θXh 1h
(2.33) h=1

+ 2δ21h CXh c1h (1, 1) + 2δ03h CXh c1h (1, 1)c2h (1, 1) + 2(δ22h − 1)c2h (1, 1) ,

where c1h (1, 1) and c2h (1, 1) denote the first order partial derivatives of ch (lh , mh ) with respect
to lh and mh respectively about the point (1, 1):
 2    
CXh θXh δ03h CXh c1h (1, 1) δ C
(2.34) = − 21h Xh .
δ021h CXh AXh c2h (1, 1) δ22h − 1

By solving (2.34) we can determine the minimum values of c1h (1, 1) and c2h (1, 1) respectively
as
h i
δ03h (δ22h − 1) − δ21h AXh
c1h (1, 1) = h i ,
CXh (AXh /θXh ) − δ03h2

(2.35) h i
δ03h δ21h − (δ22h − 1)/θXh
c2h (1, 1) = h i .
2
(AXh /θXh ) − δ03h

Putting (2.35) in (2.33) we obtain minimum MSE of σ̂c2 as

(2.36)
L
(Wh σY h )4 (δ22h − 1)2 + δ21h
2 θ
 
Xh AXh − 2θXh δ21h δ03h (δ22h − 1)
X
2
min.MSE(σ̂c ) = AY h − ,
n3h 2 θ
(AXh − δ03h Xh )
h=1

L
(Wh σY h )4 {δ21h θXh δ03h − (δ22h − 1)}2
X  
(2.37) min.MSE(σ̂c2 ) = 2
AY h − δ21h θXh − .
n3h 2 θ
(AXh − δ03h Xh )
h=1

Hence, a theorem can be established as follows.

Variance Estimation in the Presence of Measurement Errors... 285

Theorem 2.3. Up to terms of order n−3

h ,

L  2
"  #
(Wh σY h )4 δ21h θXh δ03h − (δ22h − 1)
X 
min.MSE(σ̂c2 ) ≥ AY h − 2
δ21h θXh − ,
n3h 2 θ
(AXh − δ03h Xh )
h=1

with equality holding if

h i
δ03h (δ22h − 1) − δ21h AXh
c1h (1, 1) = h i ,
2
CXh (AXh /θXh ) − δ03h
(2.38) h i
δ03h δ21h − (δ22h − 1)/θXh
c2h (1, 1) = h i .
2
(AXh /θXh ) − δ03h

The present estimators

L 
W2
X 
2
σ̂c1 = h
σ̂Y2 h lh α1h mαh 2h ,
nh
h=1
L 
Wh2 2 h
X  i
2
σ̂c2 = σ̂Y h α1h l1h + (1 − α1h )mα1h2h ,
nh
h=1
L 
Wh2 2 h
X  i
α1h
2
σ̂c3 = σ̂Y h α3h l1h + (1 − α3h )mα2h2h ,
nh
h=1
L 
Wh2 2 h
X  i
2

σ̂c4 = σ̂Y h exp α1h (lh − 1) + α2h (mh − 1) ,
nh
h=1
L 
Wh2 σ̂Y2 h
X 
2
σ̂c5 = h  α2h α3h i,
nh 1 + α l m − 1
h=1 1h 1h 2h
L
Wh2 2 h
X   i
2
σ̂c6 = σ̂Y h 1 − a1h (lh − 1) + α2h (mh − 1) ,
nh
h=1
L 
Wh2 2 h
X  i−1
2
σ̂c7 = σ̂Y h 1 − α1h (lh − 1) + α2h (mh − 1) ,
nh
h=1

are some particular members of the suggested class of estimator σ̂c2 . The mean square error of
these estimators can be obtained from (2.33) by choosing the suitable value for the constants.
2 (i = 1 to 7) is the same as given by (2.37).
The lower bound of the MSE of the estimators σ̂ci

2 wider than σ̂ 2 is proposed as

A class of estimators for σst c

L 
W2
X 
(2.39) σ̂f2 = h
fh (σ̂Y2 h , lh , mh ),
nh
h=1

where fh (σ̂Y2 h , lh , mh ) is a function of (σ̂Y2 h , lh , mh ) such that

∂fh (σY2 h , lh , mh )
fh (σY2 h , 1, 1) = σY2 h ⇒ f1h (σY2 h , 1, 1) = |(σ2 ,1,1) = 1.
∂ σ̂Y2 h Yh
286 N. Singh, G.K. Vishwakarma and R.K. Gangele

It can unveil that up-to order n− 3 2

h the optimum MSE of σ̂f is same as the optimum MSE of
σ̂c2 at (2.36) or (2.37) and is not reduced. The difference-type estimator
L 
Wh2 n 2
X  o
2
(2.40) σ̂std3
= σ̂Y h + d3h (lh − 1) + d4h (mh − 1)
nh
h=1

is a specific member of the class (2.39) but not (2.26), where d3h and d4h are acceptable
constants.

3. EFFICIENCY COMPARISONS

To obtain the conditions for which the proposed classes of estimators σ̂a2 , σ̂b2 and σ̂c2
2 , from (1.5), (2.10), (2.23) and (2.37) we can
perform better than usual unbiased estimator σ̂st
write
L
2
X (Wh σY h )4
(3.1) MSE(σ̂st ) − min.MSE(σ̂a2 ) = 2
δ21h θXh ≥ 0,
n3h
h=1
L
2
X (Wh σY h )4 (δ22h − 1)2
(3.2) MSE(σ̂st ) − min.MSE(σ̂b2 ) = ≥ 0,
n3h AXh
h=1
(3.3)
L 2
"  #
2
X (Wh σY h )4 θXh δ03h δ21h − (δ22h − 1)
MSE(σ̂st ) − min.MSE(σ̂c2 ) = 2
δ21h θXh + ≥ 0.
n3h (AXh − δh032 θ
Xh )
h=1

Remarks: To exhibit the impact of measurement error on MSE of the estimators, let
the observation for both the study variable and auxiliary variable be recorded without error.
Now the MSE of the proposed class of estimators σ̂a2 , to the third degree of approximation is
given as
L
∗
X (Wh σY h )4 h i
(3.4) MSE (σ̂a2 ) = 2
(δ40h − 1) + CXh a21h (1) + 2δ21h CXh a1h (1) ,
n3h
h=1

which is the same as the obtained by Singh and Vishwakarma [14].

From (2.8) and (3.4) we have

L
(Wh σY4 h 1 − θY h 2
    
∗
X 1 − θY h
MSE(σ̂a2 ) − MSE (σ̂a2 ) = (γ2U h + 2) +4
n3h θY h θY h
h=1

2
+ CXh (1 − θXh )a21h (1) .

The difference is always positive in nature, thus we can infer that the presence of measurement
error incorporates larger mean square error than the absence of measurement error.

To obtain the optimum value of the constant differentiating partially (3.4) with respect
to a1h and equate to zero we get
 
δ21h
a1h (1) = − .
CXh
Variance Estimation in the Presence of Measurement Errors... 287

Thus, the resultant minimum mean square error is

L
∗
X (Wh σY h )4 h i
(3.5) min.MSE (σ̂a2 ) = 2
(δ40h − 1) − δ21h .
n3h
h=1

Now the impact of measurement error can be obtained (2.10) and (3.5) from and as
L
"
(Wh σY h )4 1 − θY h 2
X  
2 ∗ 2
min.MSE(σ̂a ) − min.MSE (σ̂a ) = (γ2U h + 2)
n3h θY h
h=1
(3.6)   #
1 − θY h 2
+4 + δ21h (1 − θXh ) .
θY h

The MSE of another proposed class of estimators σ̂b2 for the estimation of σst
2 in the absence

of measurement error is given in (3.7) and is similar to Singh and Vishwakarma [14]:
L
X (Wh σY h )4 h i
(3.7) MSE∗ (σ̂b2 ) = δ40h − 1) + (δ04h − 1)b21h (1) + 2(δ22h − 1)b1h (1) .
n3h
h=1

From equation (2.21) and (3.7) we can write

L
"
(Wh σY h )4 1 − θY h 2
   
∗
X 1 − θY h
MSE(σ̂b2 ) − MSE (σ̂b2 ) = (γ2U h + 2) +4
n3h θY h θY h
h=1
(3.8) #
1 − θXh 2
   
1 − θXh 2
+ (γ2V h + 2) +4 b1h (1) .
θXh θXh

The right-hand side of (3.8) is always positive in nature, thus we can infer that mean square
error of the proposed estimator is always larger when observation is recorded with error.
MSE∗ (σ̂b2 ) is minimized for
 
δ22h − 1
(3.9) b1h (1) = − .
δ04 − 1
Thus, the resultant minimum mean square error is
L
" #
(W σ )4 (δ − 1)2
h Y h 22h
X
(3.10) min.MSE∗ (σ̂b2 ) = (δ40h − 1) − .
n3h (δ04h − 1)
h=1

From (2.23) and (3.10) we can derive the impact of measurement error as
L
"
4 1 − θY h 2
   
2 ∗ 2
X (W h σ Y h ) 1 − θY h
min.MSE(σ̂b ) − min.MSE (σ̂b ) = (γ2U h + 2) +4
n3h θY h θY h
h=1
(3.11) #
(δ22h − 1)2 1 − θXh 2
  
1 − θXh
+ (γ2V h + 2) +4 .
AXh (δ40h − 1) θXh θXh

The mean square error of the third proposed class of estimators σ̂c2 for the estimation of σ̂st
2

in the absence of measurement error is given by Singh and Vishwakarma [14] as

L
X (Wh σY h )4 h
MSE∗ (σ̂c2 ) = 2
(δ40h − 1) + CXh c21h (1, 1) + (δ04h − 1)c22h (1, 1)
n3h
h=1
(3.12)
+ 2δ21h CXh c1h (1, 1) + 2δ03h CXh c1h (1, 1)c2h (1, 1)
i
+ 2(δ22h − 1)c2h (1, 1) .
288 N. Singh, G.K. Vishwakarma and R.K. Gangele

From we can write as

L
"
(Wh σY h )4 1 − θY h 2
   
∗
X 1 − θY h
MSE(σ̂c2 ) − MSE (σ̂c2 ) = (γ2U h + 2) +4
n3h θY h θY h
h=1
1 − θXh 2
   
(3.13) 2 1 − θXh 2
+ CXh c1h (1) + (γ2V h + 2)
θXh θXh
  #
1 − θXh 2
+4 c2h
θXh

and MSE∗ (σ̂c2 ) is minimum for

h i
δ03h (δ22h − 1) − δ21h (δ04h − 1)
c1h (1, 1) = h i ,
2 −1
CXh δ04h − δ03h
(3.14) h i
δ03h δ21h − (δ22h − 1)
c2h (1, 1) = h i .
2 −1
δ04h − δ03h

Thus we can write the resultant minimum MSE as

L 2
"  #
∗ 2
X (Wh σY h )4 2 δ03h δ21h − (δ22h − 1)
(3.15) min.MSE (σ̂c ) = (δ40h − 1) − δ21h − ,
n3h 2 − 1)
(δ04h − δ03h
h=1

which can be easily obtained from (2.37) by putting σU2 h = σV2 h = 0.

Hence, we can derive the impact of measurement error on the mean square error of the
estimator (σ̂c2 ) as

min.MSE(σ̂c2 ) − min.MSE∗ (σ̂c2 ) =

L
"
(Wh σY h )4 1 − θY h 2
   
X 1 − θY h 2
= (γ2U h + 2) +4 + (1 − θXh )δ21h
(3.16) n3h θY h θY h
h=1
 2 − 1) − B (A 2 #
A1 (δ04h − δ03h 1 Xh − δh03 θXh )
− 2 − 1)(A 2 ,
(δ04h − δ03h Xh − δh03 θXh )

where
n o2
A1 = δ03h δ21h θXh − (δ22h − 1) ,
n o2
B1 = δ03h δ21h − (δ22h − 1) .

The right-hand side of (3.16) is the effect of measurement error in the mean square error of
the estimator which is always positive in nature. Thus, the proposed classes of estimators
have larger MSE in the presence of measurement errors in both study and auxiliary variables
than in the absence of measurement errors. When the measurement error is insignificant, the
inference based on these data may remain valid. Nevertheless, when the amount of error is
more significant in observed data, the inference may be invalid and inaccurate and often may
lead to unexpected and undesirable consequences.
Variance Estimation in the Presence of Measurement Errors... 289

4. DISCUSSION AND CONCLUSION

The data available for statistical analysis are always contaminated with measurement
error and may lead to fallacious inference results. When data contains heterogeneity among
units in terms of value, survey users are advised to form several homogeneous groups, and
the sampling design is well known as stratified sampling. To the best of our knowledge, the
study of measurement error for the estimation of variance in stratified random sampling has
not been addressed yet. Singh and Karpe [12] have studied the effect of measurement error
on estimation of population mean in stratified random sampling. Estimation of variance has
vital importance as it has practical uses in real-life. It is discussed by Lee [6], Srivastava
and Jhajj [18], Wu [21] and Singh and Vishwakarma [14] without the measurement error
framework.

The present study deals with the problem of estimation of variance by using auxiliary
information under the stratified sampling framework when observations are contaminated by
measurement errors. Three wider classes of estimators have been proposed. The theoretical
comparisons show that the proposed classes of estimators (σ̂a2 , σ̂b2 and σ̂c2 ) in the presence
of measurement error are more efficient than usual unbiased estimators. Since the proposed
estimators are defined as a class, a large number of estimators become the members of this
class. So the impact of measurement error on the bias and the mean square error of these
estimators can be obtained easily. We can also conclude that the MSE in the presence
of measurement error is larger than in the absence of it. Thus, the present study for the
estimation of variance under measurement error for the stratified random sampling is useful
and may attract others to carry out some work of practical use in this direction.

ACKNOWLEDGMENTS

Authors are deeply indebted to the editor-in-chief Prof. Maria Ivette Gomes and learned
referee for their valuable suggestions leading to improvement the quality of contents and pre-
sentation of the original manuscript. Authors are also thankful to Dr. (Md.) Mojibur Rahman,
Associate Professor of English Language, Department of Humanities and Social Sciences,
Indian Institute of Technology (ISM) Dhanbad, India, for his kind support in improving the
grammar of this manuscript.

REFERENCES

[1] Allen, J.; Singh, H.P. and Smarandache, F. (2003). A family of estimators of popula-
tion mean using multi-auxiliary information in presence of measurement errors, International
Journal of Social Economics, 30(7), 837–849.
[2] Biradar, R.S. and Singh, H.P. (1994). An alternative to ratio estimator for population
variance, Assam Statistical Review, 8(2), 18–33.
290 N. Singh, G.K. Vishwakarma and R.K. Gangele

[3] Cochran, W.G. (1968). Errors of measurement in statistics, Technometrics, 10(4), 637–666.
[4] Das, A.K. and Tripathi, T.P. (1978). Use of auxiliary information in estimating the finite
population variance, Sankhya, 40(C), 139–148.
[5] Isaki, C.T. (1983). Variance estimation using auxiliary information, Journal of the American
Statistical Association, 78(381), 117–123.
[6] Lee, K.H. (1973). Variance estimation in stratified sampling, Journal of the American Sta-
tistical Association, 68(342), 336–342.
[7] Maneesha and Singh, R.K. (2002). Role of regression estimator involving measurement
errors, Brazilian Journal of Probability and Statistics, 16(1), 39–46.
[8] Prasad, B. and Singh, H.P. (1990). Some improved ratio-type estimators of population
variance using auxiliary information in sample surveys, Communication in Statistics – Theory
and Methods, 21(5), 1367–1376.
[9] Shalabh (1997). Ratio method of estimation in the presence of measurement errors, Journal
of the Indian Society of Agricultural Statistics, 50(2), 150–155.
[10] Shalabh and Tsai, J.R. (2017). Ratio and product methods of estimation of population mean
in the presence of correlated measurement errors, Communications in Statistics – Simulation
and Computation, 46(7), 5566–5593.
[11] Singh, H.P. and Biradar, R.S. (1994). Estimation of finite population variance using aux-
iliary information, Journal of the Indian Society of Statistics and Operation Research, 15(1),
47–63.
[12] Singh, H.P. and Karpe, N. (2010). Effect of measurement errors on the separate and com-
bined ratio and product estimators in stratified random sampling, Journal of Modern Applied
Statistical Methods, 9(2), 388–402.
[13] Singh, H.P.; Upadhyaya, L.N. and Namjoshi, U.D. (1988). Estimation of finite population
variance, Current Science, 57(27), 1331–1334.
[14] Singh, H.P. and Vishwakarma, G.K. (2008). Some families of estimators of variance of
stratified random sample mean using auxiliary information, Journal of Statistical Theory and
Practice, 2(1), 21–43.
[15] Singh, N. and Vishwakarma, G.K. (2019). A generalised class of estimator of population
mean with the combined effect of measurement errors and non-response in sample survey,
Revista Investigacion Operacional, 40(2), 275–285.
[16] Srivastava, A.K. and Shalabh (2001). Effect of measurement errors on the regression
method of estimation in survey sampling, Journal of Statistical Research, 35(2), 35–44.
[17] Srivastava, S.K. (1971). A generalized estimator for the mean of a finite population using
multiauxiliary information, Journal of the American Statistical Association, 66, 404–407.
[18] Srivastava, S.K. and Jhajj, H.S. (1980). A class of estimators using auxiliary information
for estimating finite population variance, Sankhya, 42(C), 87–96.
[19] Upadhyaya, L.N. and Singh, H.P. (1983). Use of auxiliary information in the estimation
of population variance, Mathematical Forum, 6(2), 33–36.
[20] Upadhyaya, L.N. and Singh, H.P. (1986). On a dual to ratio estimator for estimating finite
population variance, Nepalese Mathematical Scientific Report, 11(1), 37–42.
[21] Wu, C.F.J. (1985). Variance estimation for the combined ratio and combined regression
estimators, Journal of the Royal Statistical Society, Series B (Methodological), 47(1), 147–154.