The use of auxiliary information has increased widely in survey sampling to improve the precision of estimates. A lot of works have been done by many authors using auxiliary information. The usage of the transformed auxiliary variable in estimating finite population mean is presented in [1]. The use of two auxiliary variables in estimation has been introduced by [2]. A modified ratio estimator for the estimation of finite population mean when median of the auxiliary variable was available is shown in [3]. The modified estimator shows the significant role using of auxiliary information. The cum-dual ratio estimator was more efficient than other existing estimator proved by [4]. A new modified product estimator is better than other existing estimator when median of the auxiliary variable is known [5].

Regression estimator is important in estimation theory for its improved precision. It has been widely used in estimating population parameter. Regression estimator is more useful in many practical situations including business, economics, time series forecasting, agriculture etc. [6]. Before applying regression estimator, the regression parameter needs to be estimated. In this study, we proposed new regression type estimator where correlation coefficient is used instead of regression parameter. In many practical situations, population correlation coefficients are known [7].

The structure of the paper is as follows: Section 2 contains the concept of proposed estimator and its bias and mean square error. Section 3 shows the efficiency comparison of proposed estimator compared to that of ratio, product and simple random sampling mean estimator. Section 4 illustrates simulation study of the proposed estimator. Section 5 presents the real-life example of proposed estimator. Finally, section 6 provides the conclusion of our study.

In this section we proposed our new estimator. For this purpose, we derived bias and mean square error of proposed estimator.

Consider a finite population U = ( u 1 , u 2 , ⋯ , u N ) of size N units. Let y and x denote the study variable and auxiliary variable respectively. For estimating population mean Y ¯ , a sample of size n ( n < N ) is drawn using simple random sampling without replacement (SRSWOR) from the population U.

Where, the population mean is Y ¯ = 1 N ∑ y i .

Hansen, Hurwitz and Madow (1953) consider the linear regression estimator of the population mean Y ¯ as

y ¯ L R = y ¯ + β ^ ( X ¯ − x ¯ ) .

where, β ^ = s x y s x 2 an estimate of regression coefficient is β = S x y S x 2 .

The new regression type estimator for estimating population mean Y ¯ as

y ¯ z = y ¯ + r ( X ¯ − x ¯ ) (1)

where, r = s x y s x s y is an estimate of correlation coefficient ρ = S x y S x S y .

To calculate bias and mean square error of new regression type estimator, we consider first degree approximation.

Let us define [6]

e 0 = ( y ¯ Y ¯ − 1 ) , e 2 = ( x ¯ X ¯ − 1 ) , e 2 = ( s y 2 S y 2 − 1 ) , e 3 = ( s x 2 S x 2 − 1 ) , e 4 = ( s x y S x y − 1 )

This implies that,

y ¯ = Y ¯ ( 1 + e 0 ) , x ¯ = X ¯ ( 1 + e 1 ) , s y = S y ( 1 + e 2 ) 1 2 , s x = S x ( 1 + e 3 ) 1 2 , s x y = ( 1 + e 4 ) .

E ( e 0 ) = E ( e 1 ) = E ( e 2 ) = E ( e 4 ) = 0.

E ( e 0 2 ) = ( 1 − f n ) C y 2 , E ( e 1 2 ) = ( 1 − f n ) C x 2 , E ( e 0 e 1 ) = ( 1 − f n ) C x C y ρ x y .

E ( e 1 e 2 ) = ( 1 − f n ) C x λ 21 , E ( e 1 e 3 ) = ( 1 − f n ) C x λ 03 , E ( e 1 e 4 ) = ( 1 − f n ) C x λ 12 ρ x y .

f = n N , C y 2 = S y 2 Y ¯ 2 , C x 2 = S x 2 X ¯ 2 , ρ = S x y S x S y .

S x 2 = 1 N − 1 ∑ ( x i − X ¯ ) 2 , S y 2 = 1 N − 1 ∑ ( y i − Y ¯ ) 2 , S x y = 1 N − 1 ∑ ( y i − Y ¯ ) ( x i − X ¯ ) .

Theorem 1. The bias of the proposed estimator y ¯ z of population mean Y ¯ is given by

B i a s ( y ¯ z ) = 1 2 ( 1 − f n ) S x ( ρ λ 21 + ρ λ 03 − 2 λ 12 ) .

Proof:

From Equation (1) we can write,

y ¯ z = y ¯ + r ( X ¯ − x ¯ )

⇒ y ¯ z = Y ¯ ( 1 + e 0 ) + r { X ¯ − X ¯ ( 1 + e 1 ) }

⇒ y ¯ z = Y ¯ + Y ¯ e 0 + r ( X ¯ − X ¯ − X ¯ e 1 )

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − r X ¯ e 1

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − s x y s x s y X ¯ e 1

The proposed estimator y ¯ z , in terms of e 0 , e 1 , e 2 , e 3 and e 4 can be easily written as

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − S x y ( 1 + e 4 ) S x ( 1 + e 3 ) 1 2 S y ( 1 + e 2 ) 1 2 X ¯ e 1

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − S x y S x S y X ¯ e 1 ( 1 + e 4 ) ( 1 + e 3 ) − 1 2 ( 1 + e 2 ) − 1 2 (2)

Using binomial expansion, we consider first order approximation and ignoring the higher order approximation. Hence, Equation (2) can be written as

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − ρ X ¯ e 1 ( 1 + e 4 ) ( 1 − e 2 2 + 3 8 e 2 2 − ⋯ ) ( 1 − e 3 2 + 3 8 e 3 2 − ⋯ )

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − ρ X ¯ ( e 1 + e 1 e 4 ) ( 1 − e 2 2 + 3 8 e 2 2 − ⋯ ) ( 1 − e 3 2 + 3 8 e 3 2 − ⋯ )

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − ρ X ¯ ( e 1 − e 1 e 2 2 − e 1 e 3 2 + e 1 e 4 ) (3)

After simplification, Equation (3) can be expressed as

⇒ y ¯ z − Y ¯ = Y ¯ e 0 − ρ X ¯ e 1 + ρ X ¯ e 1 e 2 2 + ρ X ¯ e 1 e 3 2 − ρ X ¯ e 1 e 4 (4)

Now, taking expectation of both sides of Equation (4) we get,

⇒ E ( y ¯ z − Y ¯ ) = Y ¯ E ( e 0 ) − ρ X ¯ E ( e 1 ) + 1 2 ρ X ¯ E ( e 1 e 2 )                                           + 1 2 ρ X ¯ E ( e 1 e 3 ) − ρ X ¯ E ( e 1 e 4 )

⇒ E ( y ¯ z − Y ¯ ) = 0 − 0 + 1 2 ρ X ¯ ( 1 − f n ) C x λ 21 + 1 2 ρ X ¯ ( 1 − f n ) C x λ 03                                           − ρ X ¯ ( 1 − f n ) C x λ 12 ρ

⇒ E ( y ¯ z − Y ¯ ) = ( 1 − f n ) ( 1 2 ρ X ¯ C x λ 21 + 1 2 ρ X ¯ C x λ 03 − X ¯ C x λ 12 )

⇒ B i a s ( y ¯ z ) = ( 1 − f n ) X ¯ C x ( 1 2 ρ λ 21 + 1 2 ρ λ 03 − λ 12 ) .

Simplifying this, we get the bias of the estimator y ¯ z . Hence this is the theorem.

Theorem 2. The mean squared error of the proposed estimator y ¯ z , to the first order approximation is

M S E ( y ¯ z ) = ( 1 − f n ) ( Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y ) .

Proof:

Squaring both sides of Equation (3) and then taking expectation

⇒ ( y ¯ z − Y ¯ ) 2 = { Y ¯ e 0 − ρ X ¯ ( e 1 − e 1 e 2 2 − e 1 e 3 2 + e 1 e 4 ) } 2

⇒ ( y ¯ z − Y ¯ ) 2 = ( Y ¯ e 0 ) 2 + ρ 2 X ¯ 2 ( e 1 − e 1 e 2 2 − e 1 e 3 2 + e 1 e 4 ) 2                                         − 2 Y ¯ e 0 ρ X ¯ ( e 1 − e 1 e 2 2 − e 1 e 3 2 + e 1 e 4 )

⇒ ( y ¯ z − Y ¯ ) 2 = Y ¯ 2 e 0 2 + ρ 2 X ¯ 2 e 1 2 − 2 Y ¯ e 0 ρ X ¯ e 1

⇒ E ( y ¯ z − Y ¯ ) 2 = Y ¯ 2 E ( e 0 2 ) + ρ 2 X ¯ 2 E ( e 1 2 ) − 2 ρ Y ¯ X ¯ E ( e 0 e 1 )

By the definition of mean squared error (MSE), we have

⇒ M S E ( y ¯ z ) = E ( y ¯ z − Y ¯ ) 2 = Y ¯ 2 E ( e 0 2 ) + ρ 2 X ¯ 2 E ( e 1 2 ) − 2 ρ Y ¯ X ¯ E ( e 0 e 1 )

⇒ M S E ( y ¯ z ) = Y ¯ 2 ( 1 − f n ) C y 2 + ρ 2 X ¯ 2 ( 1 − f n ) C x 2 − 2 ρ X ¯ Y ¯ ( 1 − f n ) ρ C x C y

⇒ M S E ( y ¯ z ) = ( 1 − f n ) ( Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y ) .

Hence this is the theorem.

This section illustrates the conditions, for which the mean square error of newly proposed regression type estimator will be minimum compared to that of simple random sampling mean estimator, product estimator and ratio estimator.

Theorem 3. The proposed estimator y ¯ z is more efficient than the product estimator y ¯ p if

ρ < 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ − 2 Y ¯ 2 C x C y ) > 0. or ρ > 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ − 2 Y ¯ 2 C x C y ) < 0.

Proof:

The proposed estimator y ¯ z is more efficient than product estimator y ¯ p if

M S E ( y ¯ z ) < M S E ( y ¯ p )

⇒ ( 1 − f n ) ( Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y )         < ( 1 − f n ) Y ¯ 2 ( C y 2 + C x 2 + 2 ρ C x C y )

⇒ Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − Y ¯ 2 C y 2 − Y ¯ 2 C x 2 − 2 Y ¯ 2 ρ C x C y < 0

⇒ ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 − 2 Y ¯ 2 ρ C x C y < 0

⇒ ρ ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ − 2 Y ¯ 2 C x C y ) < 0

Now there are two cases:

Case 1. The inequality will be satisfied if

ρ < 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ − 2 Y ¯ 2 C x C y ) > 0.

This condition holds in practice.

Case 2. The inequality will be satisfied if

ρ > 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ − 2 Y ¯ 2 C x C y ) < 0.

This condition holds in practice.

Theorem 4. The estimator y ¯ z is more efficient than the ratio estimator y ¯ R if

ρ < 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ + 2 Y ¯ 2 C x C y ) > 0. or ρ > 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ + 2 Y ¯ 2 C x C y ) < 0.

Proof:

The proposed estimator y ¯ z is more efficient than ratio estimator y ¯ R if

M S E ( y ¯ z ) < M S E ( y ¯ R )

⇒ ( 1 − f n ) ( Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y )         < ( 1 − f n ) Y ¯ 2 ( C y 2 + C x 2 − 2 ρ C x C y )

⇒ Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − Y ¯ 2 C y 2 − Y ¯ 2 C x 2 + 2 ρ Y ¯ 2 C x C y < 0

⇒ ρ ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ + 2 Y ¯ 2 C x C y ) < 0

⇒ ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 + 2 ρ Y ¯ 2 C x C y < 0

Now there are two cases:

Case 1. The inequality will be satisfied if

ρ < 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ + 2 Y ¯ 2 C x C y ) > 0.

This condition holds in practice.

Case 2. The inequality will be satisfied if

ρ > 0 and ( ρ X ¯ 2 C x 2 − 2 ρ X ¯ Y ¯ C x C y − Y ¯ 2 C x 2 ρ + 2 Y ¯ 2 C x C y ) < 0.

This condition holds in practice.

Theorem 5. The estimator y ¯ z is more efficient than simple random sample mean estimator if

ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y < 0.

Proof:

The proposed estimator y ¯ z is more efficient than simple random variance estimator y ¯ if

M S E ( y ¯ z ) − M S E ( y ¯ ) < 0

⇒ ( 1 − f n ) ( Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y ) − ( 1 − f n ) Y ¯ 2 C y 2 < 0

⇒ Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − Y ¯ 2 C y 2 < 0

⇒ ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y < 0.

This condition holds in practice.

Theorem 6. The estimator y ¯ z is more efficient than simple linear regression estimator y ¯ L R if

X ¯ 2 C x 2 − 2 X ¯ Y ¯ C x C y + S y 2 < 0.

Proof:

The proposed estimator y ¯ z is more efficient than simple linear regression estimator y ¯ L R if

M S E ( y ¯ z ) − M S E ( y ¯ L R ) < 0

⇒ ( 1 − f n ) ( Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y ) − ( 1 − f n ) S y 2 ( 1 − ρ 2 ) < 0

⇒ Y ¯ 2 C y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − S y 2 ( 1 − ρ 2 ) < 0

⇒ Y ¯ 2 S y 2 Y ¯ 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − S y 2 + ρ 2 S y 2 < 0

⇒ S y 2 + ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y − S y 2 + ρ 2 S y 2 < 0

⇒ ρ 2 X ¯ 2 C x 2 − 2 ρ 2 X ¯ Y ¯ C x C y + ρ 2 S y 2 < 0

⇒ X ¯ 2 C x 2 − 2 X ¯ Y ¯ C x C y + S y 2 < 0.

This condition does not hold in practice.

In this section, we will discuss the simulation study for the proposed estimator. For this purpose, we have to generate data from standard normal distribution. Steps for generating data from normal distribution and estimating relative efficiency of different estimator with simple random sampling estimator are given below:

Step 1: Choose a random sample of size n.

Step 2: Generate normal random variable with specified mean and variance.

Step 3: Calculate mean square error and relative efficiency of considered estimator.

Step 4: Repeat step1 to step3 until desired result is obtained.

Table 1 provides the estimated values of Mean Square Error and Relative Efficiency of proposed estimator and different estimators for simulated data.

The above information delineates that proposed estimator is more efficient than other estimator for simulating data.

To illustrate the applicability of the proposed estimators over other estimators through real data sets, we used data in [6]. The data used in this study to illustrate the application of proposed estimator is a database for the amount of real and non-real estate farm loans in different states during 1997. The result, description and necessary data statistics of the populations are given as follows:

y = Amount (in $000) of real estate farm loans in different states during 1997

x = Amount (in $000) of non-real estate farm loans in different states during 1997

N = 50 , x ¯ = 964.70 , y ¯ = 582.58 , n = 30 , s x 2 = 1145801 , s y 2 = 432636.7 , c x 2 = 1.2311 , c y 2 = 1.2747 , ρ = − 0.1222 , f = 0.6.

Table 2 shows the estimated values of the bias, mean square error and relative efficiency of proposed estimator and different estimator for real data.

Figure 1 shows the bias and mean square error of the proposed estimator is less than the bias and mean square error of the existing random sample estimator, product and ratio estimators. To examine the efficiency of the proposed estimator y ¯ z over the estimator y ¯ , y ¯ R and y ¯ p , we calculated the percentage of relative efficiency of different estimator of Y ¯ with respect to usual estimator y ¯ provided in Table 1 and Table 2 and Figure 2. The proposed estimator y ¯ z performed better than other estimators.

In this demonstration, we developed a new regression type estimator using the correlation coefficient Equation (1) and theoretically showed that the proposed estimator is more efficient than other usual ratio estimators, product estimator and simple random sampling mean estimator under some certain conditions. These theoretical conditions are also satisfied by the results of a numerical example.

The authors gratefully acknowledge the excellent research facility provided by Bangabandhu Sheikh Mujibur Rahman Science and Technology University. M. Zahid Hasan would like to specifically thank Ministry of Science and Technology Government of the People’s Republic of Bangladesh for research support through R & D.

The authors declare no conflicts of interest regarding the publication of this paper.

Hasan, M.Z., Sultana, M., Fatema, K., Hossain, Md.A. and Hossain, M.M. (2020) A New Regression Type Estimator and Its Application in Survey Sampling. Open Journal of Statistics, 10, 1010-1019. https://doi.org/10.4236/ojs.2020.106057