Effectivity of Modified Maximum Likelihood Estimators Using Selected Ranked Set Sampling Data

Abstract: This paper describes the modified maximum likelihood estimator (MMLE) of location and scale parameters based on selected ranked set sampling (SRSS) for normal, uniform and two-parameter exponential distributions. For these distributions, the MMLE of location and scale parameters for SRSS data were compared with the estimators of location and scale parameters for simple random sample (SRS) and ranked set sample (RSS). The MMLE based on SRSS data were found to be advantageous as compared to SRS and RSS estimators for the same number of measurements. The SRSS method with errors in ranking was also described. The minimum correlation between the actual and erroneous ranking was required for MMLE of SRSS to achieve better precision than usual SRS and RSS estimators. When the wrong assumption about the underlying distribution was present, the MMLE of the population mean based on SRSS was better than the RSS estimator of the population mean for all the cases considered.

Abstract: This paper describes the modified maximum likelihood estimator (MMLE) of location and scale parameters based on selected ranked set sampling (SRSS) for normal, uniform and two-parameter exponential distributions.For these distributions, the MMLE of location and scale parameters for SRSS data were compared with the estimators of location and scale parameters for simple random sample (SRS) and ranked set sample (RSS).The MMLE based on SRSS data were found to be advantageous as compared to SRS and RSS estimators for the same number of measurements.The SRSS method with errors in ranking was also described.The minimum correlation between the actual and erroneous ranking was required for MMLE of SRSS to achieve better precision than usual SRS and RSS estimators.When the wrong assumption about the underlying distribution was present, the MMLE of the population mean based on SRSS was better than the RSS estimator of the population mean for all the cases considered.

Introduction
Sampling is the process of selecting some representative part of a population to estimate unknown characteristics by observing the selected part of the population.The assumptions to use the conventional sampling design are that the population of interest is finite and the sample size is also pre-determined.But there are some other situations where a population of interest may be difficult to measure and collection of information are expensive.For these situations, particularly in relation to environmental inquiries, different non-conventional sampling methods can be used.Ranked set sampling (RSS) is one of the sampling methods intended for use in environmental studies.
The RSS is used to increase the precision of estimated yield of a representative samples without the bias of researcher choice (McIntyre, 1952), but this proposal is not being used for a long time.Subsequently some properties (such as unbiasedness, variance and relative precision with respect to SRS) of RSS estimator of population mean (Zheng and Al-Saleh, 2002) and the necessary statistical theory (Takahasi and Wakimoto, 1968) have been established.Of late considerable development has been done in the field of RSS such as RSS has the MMLE for general parameters (Zheng and Al-Saleh, 2002) which has the same expression as the MLE for SRS.Zheng and Al-Saleh (2002) also showed that, the MMLE for the location parameter was always more efficient than the MLE using SRS and for the scale parameter, the MMLE was at least as efficient as the MLE using SRS, when the same sample size was used.They also found that in case of perfect judgement ranking, MMLE was relatively efficient than MLE based on RSS.When the judgement ranking was imperfect using simulation they also found that the MMLE was more robust than the MLE (Zheng and Al-Saleh, 2002).
At present forestry and range researchers are continuing to discover the effectiveness of the RSS.Recently, a number of parametric alternatives of the usual estimation have been suggested (Hossain and Muttlak, 1999;Bhoj, 2000).The SRSS method is an improvement of the RSS method, has the optimal estimators of location and scale parameters (Hossain and Muttlak, 2001).
In this paper, the MMLE of the location and scale parameters based on SRSS data were studied.These estimators were compared with usual estimators based on SRS and RSS data for normal, uniform and two-parameter exponential distributions.The MMLE for SRSS data was also discussed for the situation where errors in ranking might be occurred.And for studying the effect of errors in ranking, we used Dell and Clutter's model (Dell and Clutter, 1972).The minimum value of the correlation coefficient between the actual and erroneous ranking which is required for achieving better precision, with respect to the usual SRS estimator and RSS estimator, were also discussed.
The first step of the RSS procedure was to draw n random sets with n elements in each sample and ranked them (without actual measuring) with respect to the variable of interest.From the first set, the element with the smallest rank was chosen.From the second set, the element with the second smallest rank was chosen.The procedure was continued until the element with the largest rank from the nth sample was chosen.This procedure yielded a total number of n elements chosen to be measured, one from each sample.The cycle might be repeated m times until nm units were measured.These nm units formed the RSS data.
In the SRSS method instead of drawing n random sets of size n, only k < n random sets of size n were drawn, and instead of measuring the ith smallest order statistic of the ith set, the n i th smallest order statistic of the ith set was considered for measurement.The values of n 1 , . . ., n k , 1 ≤ n 1 < • • • < n k ≤ n, is required to determine beforehand.Determination of n 1 , . . ., n k has described extensively earlier (Hossain and Muttlak, 2001).
Present sampling method seems to be quite appealing because SRSS estimators are expected to be more precise than those obtained by SRS and RSS with the same number of measurements when the underlying distribution is known (Hossain and Muttlak, 2001).

MMLEs of Location and Scale Parameters Using SRSS Data
In a SRSS of size k, the measured elements are For simplicity, these are denoted by x (n 1 ) , . . ., x (n k ) .Now, let x (n i ) , the n i th smallest order statistic from the ith random sample of size n, be drawn from a distribution with location and scale parameters µ and σ, respectively.The likelihood function based on the SRSS can be written as (Balakrishnan and Cohen, 1991) The algebraic deduction of the MMLE for SRSS data for normal distribution is described below in short.
Let the sample be drawn from a normal distribution with pdf Equating the first derivatives with respect to µ and σ to zero and let f (z we have the simplified form and These equations do not admit explicit solutions.So, we expand the function F (z (n i ) ) appearing in (2) and (3) in a Taylor series around the point In general we know that, if f (function) can be differentiated n times at a, then we can define the nth Taylor polynomial for f about x = a as The simplified form of the estimator of location parameter is found as follows (Balakrishnan and Cohen, 1991) where where .
The values of C and L in (4) and (5) have to be determined for given population distributions such as normal distribution.
The estimator of the scale parameter can be obtained by finding the positive root of the following equation

Comparison of MMLE Using SRSS Data with Other Estimators
The usual estimators of the population mean µ and variance σ 2 based on SRS of size k is given by  For complete RSS data the estimators of location and scale parameters are respectively.
The relative precisions of μM and σM based on k measurements with respect to μk and σk are defined as φ µ (M |R) and φ σ (M |R), where M and R stand for MMLE for SRSS and RSS.
Table 2 shows that the values of φ µ (M |R) and φ σ (M |R) for various values of n ≤ 7 (k < n) for a normal, a uniform, and a two-parameter exponential distribution.The values of φ µ (M |R) and φ σ (M |R) suggested that the MMLE for SRSS for all the cases considered were (no matter what the values of n and k) better than those based on RSS with same number of measurements.For example, for n = 7 and k = 2, values of φ µ (M |R) are 1.815, 1.752, 1.700, and the values of φ σ (M |R) are 1.697, 7.631, 2.457 for normal, uniform and two-parameter exponential distributions, respectively.

MMLE for SRSS Data with Errors in Ranking
The usual RSS estimator and the MMLE for SRSS data of the population mean and the standard deviation are based on the ranking of the variable of interest.Usually the ranking is done subjectively by rankers and because of that an error may occur in such a ranking.The errors in ranking may have an effect on the estimates (Stokes, 1980).For SRSS, the ranking may not always be perfect, i.e., the n i th smallest observation in the ith set measured by the SRSS method may not be the actual n i th order statistic in the set of size n.Following the method in Dell and Clutter (1972) we call it the n i th 'judgement order statistic'.Let x [n i ] denote the n i th smallest 'judgement order statistic' from the ith random sample of size n.According to the approach of David and Levine (1972), we can suppose that, while ranking, an observer subjectively gives each element a value of a variable y.We assume that the regression of x on y is linear and that the values follow a bivariate normal distribution.Let ρ denote the correlation coefficient between x and y.Now the following can be obtained (Hossain and Muttlak, 2001) The MMLE for SRSS data of the population mean µ using 'judgement order statistics' is given as µ * [k] = C + Lσ , where C and L are defined in (4) and (5).It can be shown that µ * [k] is an unbiased estimator of µ, and its variance is Denoting the relative precision of µ * [k] with respect to the SRS estimator μS by φ µ (E|S) it is found that , where E stands for MMLE for SRSS with errors.
The value of φ µ (E|S) depends on the magnitude of ρ, because φ µ (E|S) = ψ(ρ 2 ), then Austrian Journal of Statistics, Vol. 38 (2009), No. 2, 109-120 It can be shown that φ µ (E|S) ≥ 1 if ρ 2 ≥ ρ 2 0 , where ψ(ρ 2 0 ) = 1.The value of ρ 2 0 can be obtained as where The RSS estimator of population mean in the presence of errors in ranking is given by Dell and Clutter (1972) as with variance The relative precision of µ * k with respect to the RSS estimator μk is Values of φ µ (E|S) for various values of ρ, n, and k showed that values of φ µ (E|S) increased as the magnitudes of ρ increased.The values of |ρ 0 | for various n and k are also given in Table 3. Table 4 gives the values of φ µ (E|R) for various values of ρ, and |ρ 1 | for various n and k.Considering Table 4, it can be seen that for the larger values of ρ, values of φ µ (E|R) were also large.

Sensitivity of MMLE for SRSS of the Population Mean
In this section we studied the effects of any possible misspecification of the assumed distribution.Using the assumed distributions normal, uniform and two-parameter exponential values of φ µ (A|R) are computed for various true distributions (rectangular, normal, two-parameter exponential and exponential) and the values are given in Table 5 and 6 for various n and k.When the number of measurements was small and the cost of measurements was high, instead of SRS and RSS a precise estimator could be obtained by using SRSS method.
For this, we compared MMLE for SRSS with SRS and RSS estimators based on the same numbers of measurements.The results of the present research revealed that, for all the cases considered in this study, the MMLE for SRSS were more efficient than the estimators of SRS obtained from the same number of measurements (Table 1).So for all cases considered, a generous gain in precision was realized by using MMLE for SRSS estimators with k < n measurements.The values of φ µ (M |R) and φ σ (M |R) in Table 2 suggested that for all the cases considered here (no matter what the values of n or k are) the MMLE for SRSS were more efficient than the RSS estimators obtained from the same number of measurements for normal, uniform and two-parameter exponential distributions.
Looking at the values in Table 1 and 2, it can be observed that using the same sample size the values of relative precisions were greater than 1 for all the distributions considered.It also can be seen that, for the same to number of measurements k, the values of relative precisions (for population mean) increased as set size n increased.
In the presence of errors in ranking, MMLE for SRSS performed better than the SRS and RSS estimators when the actual ranking was highly correlated with the 'judgment ranking' (Table 3 and 4).Table 3 also showed that, in favor of MMLE for SRSS to perform better than SRS, the required lower limit of the correlation coefficient between the actual ranking and the 'judgment ranking' can be less than 0.2.Table 4 showed that,

Table 1 :
) respectively.Let φ µ (M |S) and φ σ (M |S) denote the relative precisions of μM with respect to μS and σM with respect to σS , where M and S stand for MMLE for SRSS and SRS estimators.The values of φ µ (M |S) and φ σ (M |S) for various values of n ≤ 7 (k < n) for a normal, a uniform, and a two-parameter exponential distributions are presented in Table1.The tabulated values show that high values of relative precision with respect to the usual SRS estimators of size k can be achieved by MMLE for SRSS data with the same number of measurements.For example, for n = 6 and k = 2, values of φ µ (M |S) are 2.296, 2.503, 2.303 and the values of φ σ (M |S) are 1.646, 20.291, 2.214 for normal, uniform and two-parameter exponential distributions, respectively.Relative precision of MMLE of the population mean and standard deviation for SRSS data compared to the usual SRS estimators μS and σS .

Table 2 :
Relative precision of MMLE of the population mean and standard deviation for SRSS data compared to the usual RSS estimators μR and σR .

Table 3 :
Relative precision φ µ (E|S) of MMLE of the population mean for SRSS data compared to the usual SRS estimator μS when errors occur in ranking.

Table 4 :
Relative precision φ µ (E|R) of MMLE of the population mean for SRSS data compared to the usual RSS estimator μR when errors occur in ranking.

Table 5 :
Values of the relative precision φ µ (A|R) for various true distributions and for the rectangular and normal as assumed distributions.

Table 6 :
Values of the relative precision φ µ (A|R) for various true distributions and for the two-parameter exponential distribution as assumed distribution.