STATISTICS IN TRANSITION-new series, Spring 2013 31

STATISTICS IN TRANSITION-new series, Spring 2013

Vol. 14, No. 1, pp. 31–44

SYNTHETIC ESTIMATORS USING AUXILIARY

INFORMATION IN SMALL DOMAINS

P. K. Rai 1, K. K. Pandey 2

ABSTRACT

In the present article we discuss the generalized class of synthetic estimators for
estimating the population mean of small domains under the information of two
auxiliary variables, and describe the special cases under the different values of the
constant beta involved in the proposed generalized class of synthetic estimator. In
addition we have taken a numerical illustration for the two auxiliary variables and
compared the result for the synthetic ratio estimator under single and two
auxiliary variables.

Key words: auxiliary information, small area (domain) estimation, synthetic

estimation, optimum weights.

1. Introduction

An estimator is called a synthetic estimator if a reliable direct estimator for a

larger area, covering several small areas, is used to derive an indirect estimator for
a small area under the assumption that the small areas have the same
characteristics as the large area (Gonzalez, 1973). Such estimators have been
studied by Gonzalez (1973), Gonzalez and Waksberg (1973). It is a fact that if
small domain sample sizes are relatively small the synthetic estimator performs
better than the simple direct estimators, whereas when sample sizes are large the
direct estimators perform better than the synthetic estimators (Schaible, Brock,
Casady and Schnack, 1977). The classes of synthetic estimators proposed by the
above authors give consistent estimators if the corresponding synthetic
assumptions are satisfied. These authors, further, discuss the generalized class of
synthetic estimators under simple random sampling and stratified random

1
Department of Mathematics and Statistics, Banasthali University, Rajasthan India-304022. E-mail:
[email protected].
2
College of management & Economic Studies, University of Petroleum & Energy Studies (UPES),
Energy Acres, P.O. Bidholi, Dehradun-248007. E-mail: [email protected],
[email protected].
32 P. K. Rai, K. K. Pandey: Synthetic estimators …

sampling schemes. In sample surveys usually auxiliary variables are used to

increase the precision of the estimators. A ratio estimator is one of the most
commonly used estimators among others for the population mean or population
total with the help of an auxiliary character. It was shown by Tikkiwal, G.C. and
Ghiya, A. (2004), Tikkiwal, G.C. and Pandey, K.K. (2007), Pandey Krishan K.
and Tikkiwal, G.C. (2010), Pandey, Krishan K. (2010), that when an auxiliary
variable is closely related with the variable under study, the small area estimators
based on auxiliary information perform better than those which do not use
auxiliary information. Further, Tikkiwal, G.C. and Pandey, K.K. (2007) discuss
the generalized class of synthetic and composite estimators under Lahiri-Midzuno
and systematic sampling schemes. The relative performances of these estimators
are empirically assessed for the problem of crop acreage estimation for small
domains.
It is rather difficult to assess the performance of these estimators theoretically.
Here we have discussed the different aspect of the generalized class of synthetic
estimators for small area estimation problems when more than one auxiliary
information is available.

2. Generalized class of synthetic estimators in sample surveys

We define a generalized class of synthetic estimators for estimating the

population mean Y under ‘k’ auxiliary variables x1 , x2 ,.......xk , as follows

βi
k
 x 
ysyn = ∑ Wi y  i  (1)
i =1  Xi 

C0
where β i ' s are equal to the − ρ0i and Wi are the weights to be obtained by
Ci
k
minimizing the variance of (1) subject to the condition that ∑W =1. Here,
i =1
i xi

and X i denote the sample mean and population mean of xi (i = 1, 2,....., p )

respectively, ρi j (i ≠ j =0,1,....., p ) denotes the correlation coefficient between
xi and x j , and Ci (i = 0,1,....., p) denotes the coefficient of variation of xi ; the
suffix 0 stands for the variable y and y is the sample mean of the variable under
study.
 STATISTICS IN TRANSITION-new series, Spring 2013 33

3. Notations and formulation under small domains

Let us represent the important notations which are to be used in this paper.
Suppose that a finite population U= (1,…, i,…, N) is divided into ‘A’ non-
overlapping domains Ua of size Na (a=1,…, A) for which estimates are required.
The domains may be numerous and represent small geographical areas of
a sampled population, which may be a state or a sub-division of the state as the
case may be. Let the characteristic under study be denoted by ‘y’. Further, assume
that the auxiliary information is also available and denoted by ‘x’. A simple
random sample (without replacement) s= (1,…, i, …, n) of size n is selected such
that na (a=1,…, A) units in the sample ‘s’ come from small area ‘a’.
Consequently,
A

∑N
A

a =1
a = N and ∑n
a =1
a =n (2)

Let us consider the case of generalized synthetic estimator for estimating the
population mean Ya for domain ‘a’ under two auxiliary variables x1 and x2 ;
β1 β2
 x   x 
=ysyn , a W1 y  1  + W2 y  2  (3)
 X 1a   X 2a 

Here, W1 and W2 are the weights such that W1 + W2 =1 and β1 , β1 are suitably
chosen constants. To find the expectation and mean square error of the estimator
ysyn , a define
y −Y x1 − X 1 x2 − X 2
=ε0 = , ε1 = , ε2 (4)
Y X1 X2

Then, clearly

(ε 0 ) E=
E= (ε1 ) E=
(ε 2 ) 0 (5)

f 2 f 2 f 2
=E (ε 02 ) = C0 , E (ε12 ) = C1 , E (ε 22 ) C2 (6)
n n n

f f f
and E (ε oε1 ) =
= C01 , E (ε oε 2 ) = C02 , E (ε1ε 2 ) C12 (7)
n n n
 34 P. K. Rai, K. K. Pandey: Synthetic estimators …

where

N −n S y2 S x21 S x22
=f = , 2
C0 = , C12
= , 2
C2 ,
N Y2 X 12 X 12
S y x1 S y x2 S x1x2 (8)
=C01 = , C02 = , C12 ,
Y X1 Y X2 X1 X 2

and

N N
1 1
=S y2 ∑
N − 1 i =1
( yi − Y )=
2
; S x21 ∑
N − 1 i =1
( x1i − X 1 ) 2

1 N
=S ∑
N − 1 i =1
( x2i − X 2 ) 2
2
x2 ;

N
1
=
S yx1 ∑ ( yi − Y )( x1i − X 1 )
N − 1 i =1
N
1
=
S yx2 ∑ ( yi − Y )( x2i − X 2 )
N − 1 i =1
N
1
S=
x1 x2 ∑ ( x1i − X 1 )( x2i − X 2 )
N − 1 i =1
(9)

4. Bias and mean square error

In this section bias and mean square error expressions are considered up to the
terms of order (1/n) only. The ysyn , a can be expressed as
β1 β2
 X (1 + ε1 )   X 2 (1 + ε 2 ) 
ysyn , a =W1Y (1 + ε 0 )  1  + W2Y (1 + ε 0 )  
 X 1a   X 2a 
(10)
 STATISTICS IN TRANSITION-new series, Spring 2013 35

assuming that the contribution of terms involving powers in ε 0 , ε1 and ε 2 higher

than the second order is negligible. The design bias of ysyn , a and MSE ( ysyn , a ) is
given below as

β1
 X1   f  β1 ( β1 − 1) 2 
B ( ysyn , a ) =W1 Y   1 + n  C1 + β1C01  
 X 1a    2! 
β2
 X2   f  β 2 ( β 2 − 1) 2 
+ W2 Y   1 + n  C2 + β 2C0 2   − Ya
 X 2a    2! 

(11)

2
  X  β1  X2  
β2

MSE ( ysyn , a ) Y W1 

= 2 1
 + W2   
  X 1a   X 2 a  
2 β1
 X  f 
+Y W  1 
2
1
2


2
{
 n C0 + β1 (2 β1C1 − C1 + 4C01 ) 
2 2


}
 X 1a 
2 β2
 X  f 
+ Y 2W22  2  {
2 2 2
}
 n C0 + β 2 (2 β 2C2 − C2 + 4C02 ) 
 X 2a 
  2  ( β1 − 1) 2  
β1  C0 + β1  2C01 +
β2 C1   
2  X1   X 2   f   2!  
+ 2W1W2Y      
 X 1a   X 2 a   n + β  2C + ( β 2 − 1) C 2  + β β C  
 2 2  1 2 12 
  
02
2!   
β1
  X1   f  β1 ( β1 − 1) 2  
W1Y   1 +  C1 + β1C01   
  X 1a   n  2!  
− 2Ya  β2  + Ya
2

  X2   f  β 2 ( β 2 − 1) 2  
 +W2 Y  X  1 + n  C2 + β 2C02   
  2a    2!  
(12)

Also, the optimum value for the weights W1opt and W2opt can be obtained by
minimizing mean square error term of (12).
36 P. K. Rai, K. K. Pandey: Synthetic estimators …

5. Special cases: various synthetic estimators

The generalized synthetic estimator ysyn , a reduces to the simple synthetic

estimator if β1 and β 2 equal to zero, i.e. β=
1 β=
2 0
ysyn , a= y= ysyn , s , a (13)

( )
β
and synthetic assumption Ya X a ≅ Y ( X ) β reduces to Ya ≅ Y . Substituting
β=
1 β=
2 0 in the expression (11) we get

B ( ysyn , a ) =W1 Y + W2 Y − Ya =Y − Ya = B ( ysyn , s , a ) (14)

This is the expression for design bias of the simple synthetic estimator. The
design bias of the synthetic estimator vanishes if the synthetic assumption, i.e.
Ya  Y is satisfied. Now β=1 β=2 0 in the expression (12) gives
f 2
MSE ( ysyn , s , a ) = Y 2 (W1 + W2 ) + Y 2 C0 (W1 + W2 ) − 2YaY (W1 + W2 ) + Y a2
2 2

n
2 f N −n 2
= Y= C02 Sy (15)
n Nn
This is the mean square error of simple synthetic estimator under said
synthetic assumption.
The generalized Synthetic estimator ysyn , a reduces to ratio synthetic estimator
under two auxiliary variables, if β1 and β 2 equal to -1, i.e. β1 = β 2 = −1
 y   y 
=
ysyn , r , a W1   X 1a + W2   X 2a (16)
 1
x  2
x
Substituting β1 = β 2 = −1 in the expression (11) and (12) we get the
expressions for the bias and mse for the ratio synthetic estimator.
The generalized synthetic estimator ysyn , a reduces to the product synthetic
estimator under two auxiliary variables, if β1 and β 2 equal to +1, i.e.
β1 = β 2 = +1
 x   x 
=
ysyn , p , a W1 y  1  + W2 y  2  (17)
 X 1a   X 2a 

Substituting β1 = β 2 = +1 in the expression (11) and (12) we get the

expressions for the bias and mse for the product synthetic estimator.
STATISTICS IN TRANSITION-new series, Spring 2013 37

6. Numerical illustration

We consider the study variable as REV84, the real estate values according to
1984 assessment and use the two auxiliary variables as population under
municipalities of 1975 and 1985 of the different geographic region indicator of
Swedish municipalities. Just draw the sample of different sizes using SRSWOR
scheme and analyze the cases for regions 1, 2 and 3 as small domains.
We have considered the cases under single and double auxiliary variables and
computed the biases and mse’s for the different sample sizes. Using the
expressions of optimum weights we have computed the value of weights for the
generalized synthetic estimator under β1 = β 2 = −1 which reduces to synthetic
ratio estimator, thus W1opt = 0.978828466 and W2opt =0.021171534. And Ya =
3011.683, X 1a =28.92308, X 2 a = 255.0192, Y =3133.862676, X 1 = 28.80986,
X 2 = 111.9471831
Under single auxiliary variable the bias and mse for the synthetic ratio
estimator is given by
 N −n 2 
B2= B ( ysyn ,r ,a =
) Y
X
X a 1 +
 Nn
( Cx − Cxy )  − Ya

(18)

2
Y   N −n 
MSE ( =
ysyn ,r ,a )  X a  1 + Nn {3Cx + C y − 4Cxy }
2 2
and
X 
Y  N −n 2 
−2Ya  X a  1 + ( Cx − Cxy )  + Ya2 (19)
X  Nn 
Using the equations above we show the results for bias and mse of synthetic
ratio estimator under two scenarios for the different sample sizes in the given
tables in appendix. The results can be also explored by the following graphical
presentation.

7. Conclusions

At least two auxiliary variables will be the better choice over a single one
when the sample size decreases. In sample surveys it is useful to make use of
information on the auxiliary variable to increase the precision of the estimators.
The above study will provide the motivation towards the use of generalized class
of synthetic estimators in the small area estimation, when the information on two
auxiliary variables is available.
38 P. K. Rai, K. K. Pandey: Synthetic estimators …

Acknowledgements

The authors are grateful to the University Grant Commission (UGC),

F.No.36-341/2008 (SR), New Delhi for providing support and facilitation to this
research and development work.

REFERENCES

AGRAWAL, M. C. and ROY, D. C., (1997). Efficient Estimators for Small

Domains. Jour. Ind. Soc. Ag. Statistics 52(3), 327-337.
GHOSH, M. and RAO, J. N. K., (1994). Small Area Estimation: An Appraisal.
Statistical Science, 9, 55-93.
GONZALEZ, M. E., (1973). Use and Evaluation of Synthetic Estimates.
Proceedings of the Social Statistical Section of American Statistical
Association, 33-36.
GONZALEZ, M. E. and WAKSBERG, J., (1973). Estimation of the Error of
Synthetic Estimates. Paper presented at first meeting of the International
Association of Survey Statisticians, Vienna, Austria, 18-25, August 1973.
HEDAYAT, A. S. and SINHA, B. K., (1991). Design and Inference in Finite
Population Sampling. John Wiley and Sons, New York.
PANDEY, KRISHAN K. and TIKKIWAL, G. C., (2010). ,,Generalized class of
synthetic estimators for small area under systematic sampling design”
Statistics in Transition-new series, Poland, Vol.11 No.1, pp. 75-89.
PANDEY, KRISHAN, K., (2010). ,,Aspects of small area estimation using
auxiliary information”. Book published by VDM Verlag Dr. Muller GmbH &
Co. KG, Germany. (ISBN: 978-3-639-31569-1).
PLATEK, R., RAO, J. N. K., SARNDAL, C. E. and SINGH, M. P., (1987). Small
Area Statistics: An International Symposium. John Wiley and Sons, New
York.
PURCELL, N. J. and KISH, L., (1979). Estimation for Small Domain.
Biometrics, 35, 365-384.
SARNDAL, C. E., SWENSSON, B. and WRETMAN, J., (1992). Model Assisted
Survey Sampling. Springer-Verlag, New York.
SCHAIBLE, W. L., BROCK, D. B., CASADY, R. J. and SCHNACK, G. A.,
(1977). An Empirical Comparison of the Simple Inflation, Synthetic and
Composite Estimators for Small Area Statistics. Proceedings of the American
Statistical Association, Social Statistics Section, 1017-1021.
STATISTICS IN TRANSITION-new series, Spring 2013 39

SINGH, M. P., GAMBINO, J. and MANTEL, H., (1993). Issues and Options in
the Provision of Small Area Data. Proceedings of International Scientific
Conference on Small Area Statistics and survey Design (Held in September,
1992 in Warsaw, Poland), 37-75.
SRIVASTAVA, S. K., (1967). An estimator using auxiliary information in
sample surveys, Calcutta Statistical Association Bulletin, 16, 121-132.
TIKKIWAL, B. D. and TIKKIWAL, G. C., (1991). Sampling Strategies in
Surveys. The Role of the Theory of T-Classes and Computers. Symposium
Ind. Agri. Stat. Research Institute, New Delhi, 287-296.
TIKKIWAL, G. C. and GUPTA, A. K., (1991). Estimation of Population Mean
under Successive Sampling When Various Weights and Regression
Coefficient are Unknown. Biometrical Journal, 33, 529-538.
TIKKIWAL, G. C. and GHIYA, ALKA., (2000). A Generalized Class of
Synthetic Estimators with Application to Crop Acreage Estimation for Small
Domains. Biometrical Journal, 42 (7), 865-876.
TIKKIWAL, G. C. and PANDEY, K. K., (2007). On Some Aspects of Small
Area Estimation Using Auxiliary Information. Ph.D. Thesis under supervision
of Prof. G. C. Tikkiwal, Head of Department of Mathematics and Statistics
J. N. V. University Jodhpur Rajasthan.
40 P. K. Rai, K. K. Pandey: Synthetic estimators …

APPENDICES

FIGURES
Graph 1.
Bias Under Single & Two Auxiliary Variables

255 350

250
300
245

240 250

235
200
Bias

230

150
225

220 100

215
50
210

205 0
160 100 50 10
Sample Sizes

Bias1 Bias2

Graph 2.

MSE's Under Single & Two Auxiliary Variables

200000 350000

180000
300000
160000

140000 250000

120000
200000
MSE

100000

150000
80000

60000 100000

40000
50000
20000

0 0
160 100 50 10
Sample Sizes

MSE1 MSE2
STATISTICS IN TRANSITION-new series, Spring 2013 41

Graph 3.

Bias under Single and Two auxiliary Variables

400

350

300

250
Biases

Bias Under Single AV

200
Bias Under Two AV

150

100

50

0
160 100 50 10
Sample Sizes

Graph 4.
MSE's under Single and Two Auxiliary Variables

400000

350000

300000

250000
MSE's

MSE Under Single AV

200000
MSE Under Two AV

150000

100000

50000

0
160 100 50 10
Sample sizes
42 P. K. Rai, K. K. Pandey: Synthetic estimators …

Graph 5.
Bais under Single and Two Auxiliary Variables

450

400

350

300

250
Bias Under Single AV
Bias

Bias Under Two AV

200

150

100

50

0
160 100 50 10
Sample Sizes

Graph 6.
MSE's under Single and Two Auxiliary Variables

350000

300000

250000

200000
MSE's

MSE Under Single AV

MSE Under Two AV
150000

100000

50000

0
160 100 50 10
Sample Sizes
STATISTICS IN TRANSITION-new series, Spring 2013 43

TABLES

Table No. 6.1

Coefficient of Variation involved in Computation

C02 2.54
C12 3.34
C22 2.81
C01 2.81
C02 1.56
C12 1.87

Table No. 6.2

Bias of Synthetic Ratio Estimator under Single and Two Auxiliary Variables

Sample Sizes Bias Under Single AV Bias Under Two AV

160 221.07 139.14
100 222.30 145.53
50 225.57 162.56
10 251.71 298.79

Table No. 6.3

MSE of Synthetic Ratio Estimator under Single and Two Auxiliary Variables

Sample Sizes MSE Under Single AV MSE Under Two AV

160 51908.99 26914.89
100 54960.31 39043.23
50 64560.66 71385.74
10 176930.48 330125.22

Table No. 6.4

Bias of Synthetic Ratio Estimator under Single and Two Auxiliary Variables

Sample Sizes Bias Under Single AV Bias Under Two AV

160 231.13 152.10
100 242.29 169.76
50 276.56 195.83
10 305.31 350.77
44 P. K. Rai, K. K. Pandey: Synthetic estimators …

Table No. 6.5

MSE of Synthetic Ratio Estimator under Single and Two Auxiliary Variables

Sample Sizes MSE Under Single AV MSE Under Two AV

160 50025.81 36980.87
100 55635.45 42565.40
50 64860.68 82344.91
10 223692.88 362512.46

Table No. 6.6

Bias of Synthetic Ratio Estimator under Single and Two Auxiliary Variables

Sample Sizes Bias Under Single AV Bias Under Two AV

160 331.74 225.37
100 385.35 258.88
50 390.56 325.83
10 403.30 336.73

Table No. 6.7

MSE of Synthetic Ratio Estimator under Single and Two Auxiliary Variables

Sample Sizes MSE Under Single AV MSE Under Two AV

160 69871.32 38565.61
100 66523.33 45021.33
50 72563.24 77452.26
10 213564.13 325613.65