MODIFIED ML ESTIMATION: SIZE-BIASED LOMAX DISTRIBUTION Dr. R. Subba Rao1, A. Naga Durgamamba2, Dr. R.R.L.Kantam3 1
Department of Basic Science, Shri Vishnu Engg. College for Women, Bhimavaram ,Andhra Pradesh, (India)
2
Department of Humanities and Science, Raghu Institute of Technology, Dakamarri,Vizag, Andhra Pradesh, (India) 3
Department of Statistics, Acharya Nagarjuna University, Guntur, Andhra Pradesh, (India)
ABSTRACT Size biased Lomax distribution is considered with a known shape parameter. Some of its distributional characteristics are presented. Estimation of its scale parameter by the well known maximum likelihood method is modified by two different approaches in order to yield linear estimators. The proposed methods are compared with respect to simulated sampling characteristics.
Keywords: Asymptotic Variance, Log Likelihood Function, MLE, Order Statistics, Size Biased Lomax Distribution.
I INTRODUCTION Estimation of parameters of the size biased Lomax model is considered. Let α and σ be the two parameters of the size biased Lomax distribution (SBLD), where α is shape parameter and σ is scale parameter. First we establish the well known maximum likelihood method of estimation of the parameters from complete sample when one of the parameters is known while the other is unknown. Hence we start with general maximum likelihood estimation of α and σ. As the estimating equations are to be solved by numerical iterative techniques we suggest some modifications to maximum likelihood method from complete sample. Such studies based on modified maximum likelihood estimation which is attempted by many researchers with respect to other standard reliability models that include K. Rosaiah et al. derived (i) ML and Modified ML estimation in gamma distribution with known prior relation among the parameters (1993a) and (ii) On modified maximum likelihood estimation of gamma parameters (1993b), R.R.L. Kantam and G. Srinivasa Rao developed Reliability estimation in Rayleigh distribution with censoring some
287 | P a g e
approximations to ML Method (1993), R.R.L. Kantam and B. Sriram determined Maximum likelihood estimation from censored samples- Some modifications in length biased version of exponential model (2003), Pareto Distribution – Some methods of Estimation constructed by R. Subba Rao et al. (2010) and Modified Maximum Likelihood Estimation in Pareto-Rayleigh distribution studied by R. Subba Rao et al. (2015). In this paper discussion of complete sample is given in Section 2. Since the process of estimation involves lot of numerical computations all such results are presented in the form of numerical tables towards the end of the Section with appropriate identification and labels.
II ESTIMATION FROM COMPLETE SAMPLE Let f (x) be the probability density function of the two parameter size biased Lomax distribution and is given by: f x
1 1 x x ; x 0, 1, 0 1
(2.1)
where α and σ are shape and scale parameters respectively. Let x1 x2 x3 ........ xn be a complete ordered sample of size n drawn from the above distribution. The log likelihood equations to estimate α and σ from the given complete sample are given by x x log L n n x log 1 1 . 1 2 ......... 1 n 1 n x log L n n log 1 i 1 i 1
(2.2)
xi
n 2 log L 2n 1 x i i 1 1
(2.3)
For M.L.Es of α and σ
log L 0,
log L 0
After simplification, these equations become n
2 1 2
xi
log 1 i 1
n
288 | P a g e
n
2 1 i 1 2
log 1 Zi (2.4)
n
xi 2 1 x i 1 1 i n
2n
2n
1
n
zi
1 z i 1
(2.5) i
where It can be seen that equation (2.5) can be solved only by iterative method for ‘σ’. The M.L.E of α is an analytical expression involving ‘σ’. In order to overcome the iterative techniques that may sometimes lead to convergence problems we approximate the expression h zi
zi 1 zi
(2.6)
of the log likelihood equation (2.5) for estimating ‘σ’ by a linear expression say h zi i i zi
(2.7)
in certain admissible ranges of
. Such approximations are not feasible for the log likelihood equation of ‘α’. Hence
we develop our approximate M.L. method for estimation of ‘σ’ with known ‘α’. As per the parametric specifications we take α = 3. After using the linear approximation given by equation (2.7) in the equation (2.5) and solving it for ‘σ’ we get n
ˆ
1 i xi
(2.8)
i 1
n
2n 1 i i 1
as an approximate M.L.E. of ‘σ’, which is a linear estimator. Similar approximations are in Tiku (1967) and Balakrishnan (1990). We suggest two methods finding
of equation (2.7).
Method I:
289 | P a g e
Let P i
i , i 1, 2, 3,...........n n 1
Let Zi , Zi be the solutions of the following equations
F Zi pi and F Zi pi
where pi pi
pi qi , n
pi pi
pi qi n
The solutions of Zi , Zi in our size biased Lomax distribution are
F Zi Pi Zi F 1 Pi F Zi Pi Zi F 1 Pi The slope
of the linear approximation in the equation (2.7) are respectively given by
h Z i h Z i
i And
and intercept
(2.9)
Z i Z i
i h Zi i Zi
The values of
and
(2.10)
in this method for n = 5, 10, 15, 20 and 25 for α = 3 are given in Table (2.1).
Method II: Consider the Taylor’s expansion of h zi
zi in the neighbourhood of ith quantile of size biased Lomax 1 zi
population. We get another linear approximation for h Z i is given by
h Zi i i Zi where i h1 Zi with
(2.11) as the ith quantile for the population given by
F Zi Pi Zi F 1 Pi ,
Pi
i n 1
i h Zi i Zi Substituting these approximations in the equation (2.8) we get another linear estimator of ‘σ' with different values of and
. The values of
and
in this method for n = 5, 10, 15, 20 and 25 for α = 3 are given in Table (2.2).
These two methods are asymptotically as efficient as exact maximum likelihood estimators as can be seen from the following narration (Tiku et al. 1986; Balakrishnan, 1990).
290 | P a g e
In the above two modified methods, the basic principle is that certain expressions in the log likelihood equation are linearized in a neighbourhood of the population quantile which depends on the size of the sample also. The larger the size, the narrower is the neighbourhood and hence the closer is the approximation. That is, the exactness of the approximation becomes finer and finer for large values of ‘n’. Hence the approximate log likelihood equation and the exact log likelihood equation tend to each other as n
. Hence the exact and modified MLEs are
asymptotically identical (Tiku et al. 1986). The same may not be true in small samples and these are to be assessed with the help of small sample characteristics of the MMLEs. Because of difficulties in finding the analytical sampling variances, we compared the modified MLEs of two methods (Method I and Method II) of estimates through Monte Carlo simulation. 1,000 random samples of size n = 5 (5) 25 each are generated from size biased Lomax distribution with α = 3. For each sample with α = 3 the
and
of Method I and Method II are as given in Tables 2.1 and 2.2 and are used in
equation 2.8 to get the modified MLE of ‘σ’ by Method I and Method II respectively. The empirical variances of MMLEs by Method-I and Method-II are compared and given in Table 2.3. A comparison of the sampling characteristics namely the bias, variance and M.S.E of Methods I and II, reveal that MMLE of Method I is preferable to that of Method II.
Table – 2.1 Intercept ( ) and Slope ( ) in MMLE by Method I (when α = 3)
n
i
5
1
0.100281
0.454169
5
2
0.115767
0.367651
5
3
0.226304
0.226304
5
4
0.367651
0.115767
5
5
0.454169
0.100281
10
1
0.050617
0.592522
10
2
0.054466
0.539664
10
3
0.100664
0.430874
291 | P a g e
10
4
0.151329
0.344394
10
5
0.20775
0.271357
n
i
10
6
0.271357
0.20775
10
7
0.344394
0.151329
10
8
0.430874
0.100664
10
9
0.539664
0.054466
10
10
0.592522
0.050617
15
1
0.033662
0.660326
15
2
0.0354
0.660326
15
3
0.064332
0.528229
15
4
0.09502
0.454645
15
5
0.127878
0.391957
15
6
0.163176
0.336824
15
7
0.201227
0.287377
15
8
0.242441
0.242441
15
9
0.287377
0.201227
15
10
0.336824
0.163176
15
11
0.391957
0.127878
15
12
0.454645
0.09502
15
13
0.528229
0.064332
15
14
0.660326
0.0354
15
15
0.660326
0.033662
292 | P a g e
20
1
0.025159
0.702527
20
2
0.026152
0.669137
20
3
0.047119
0.588105
n
i
20
4
0.069009
0.522942
20
5
0.092057
0.467197
20
6
0.116366
0.417954
20
7
0.142028
0.373576
20
8
0.169153
0.333035
20
9
0.19788
0.295632
20
10
0.228379
0.260869
20
11
0.260869
0.228379
20
12
0.295632
0.19788
20
13
0.333035
0.169153
20
14
0.373576
0.142028
20
15
0.417954
0.116366
20
16
0.467197
0.092057
20
17
0.522942
0.069009
20
18
0.588105
0.047119
20
19
0.669137
0.026152
20
20
0.702527
0.025159
25
1
0.02006
0.732041
293 | P a g e
25
2
0.020706
0.703045
25
3
0.037107
0.629696
25
4
0.054068
0.570574
25
5
0.071757
0.519872
25
6
0.090227
0.474969
n
i
25
7
0.109517
0.434396
25
8
0.129668
0.397225
25
9
0.150729
0.362829
25
10
0.172764
0.330756
25
11
0.195841
0.300672
25
12
0.22052
0.272322
25
13
0.245502
0.245502
25
14
0.272322
0.22052
25
15
0.300672
0.195841
25
16
0.330756
0.172764
25
17
0.362829
0.150729
25
18
0.397225
0.129668
25
19
0.434396
0.109517
25
20
0.474969
0.090227
25
21
0.519872
0.071757
25
22
0.570574
0.054068
25
23
0.629696
0.037107
294 | P a g e
25
24
0.703045
0.020706
25
25
0.732041
0.02006
Table – 2.2 Intercept ( ) and Slope ( ) in MMLE by Method II (when α = 3) n
i
5
1
0.067158
0.54886
5
2
0.047807
0.610511
5
3
0.119761
0.427631
5
4
0.221399
0.280337
5
5
0.375814
0.14974
10
1
0.034592
0.662612
10
2
0.022148
0.724506
10
3
0.05183
0.596505
10
4
0.08787
0.495012
10
5
0.130433
0.408122
10
6
0.180423
0.330898
10
7
0.239605
0.260616
10
8
0.311234
0.195467
10
9
0.40198
0.133942
10
10
0.529845
0.074036
15
1
0.023187
0.71864
15
2
0.014349
0.774776
15
3
0.032882
0.670212
15
4
0.054562
0.587391
15
5
0.079107
0.516588
15
6
0.406506
0.453802
15
7
0.13691
0.396883
15
8
0.170616
0.344502
15
9
0.208085
0.295759
15
10
0.25
0.25
295 | P a g e
15
11
0.297394
0.206717
15
12
0.351899
0.165477
15
13
0.416325
0.12586
15
14
0.496279
0.087338
15
15
0.606743
0.048869
20
1
0.017404
0.753558
20
2
0.010593
0.804744
n
i
20
3
0.024019
0.714059
20
4
0.039447
0.642223
20
5
0.056595
0.580801
20
6
0.075357
0.526331
20
7
0.095717
0.476955
20
8
0.117712
0.431528
20
9
0.14143
0.389286
20
10
0.167003
0.349683
20
11
0.194614
0.312313
20
12
0.224509
0.276861
20
13
0.257022
0.243075
20
14
0.292605
0.210746
20
15
0.331893
0.17969
20
16
0.375814
0.14974
20
17
0.425816
0.120725
20
18
0.484366
0.092437
20
19
0.556381
0.064562
20
20
0.654951
0.03637
25
1
0.013915
0.777993
25
2
0.008389
0.825209
25
3
0.018894
0.743984
25
4
0.030839
0.679618
25
5
0.043976
0.624565
25
6
0.058194
0.575727
25
7
0.073441
0.53144
25
8
0.089707
0.490684
296 | P a g e
25
9
0.107003
0.452776
25
10
0.125363
0.417231
25
11
0.144836
0.383689
25
12
0.165495
0.351873
25
13
0.187429
0.321567
25
14
0.210755
0.292594
25
15
0.235618
0.264808
25
16
0.262204
0.238087
25
17
0.290749
0.212325
n
i
25
18
0.321567
0.187429
25
19
0.355073
0.163313
25
20
0.391846
0.139894
25
21
0.432728
0.117086
25
22
0.479026
0.094792
25
23
0.532965
0.072876
25
24
0.598965
0.051108
25
25
0.68804
0.028919
Table – 2.3 Sample Characteristics of MMLE of 'σ' from Complete Sample Method I and Method II
BIAS α
VARIANCE
MSE
N MMLE-I
MMLE-II
MMLE-I
MMLE-II
MMLE-I
MMLE-II
3
5
0.25713
0.59366
0.76178
1.135183
0.82789
1.48761
3
10
0.08878
0.29662
0.29284
0.39633
0.30072
0.48431
3
15
0.04531
0.19875
0.13584
0.18001
0.13789
0.21951
3
20
0.03345
0.153892
0.08263
0.1058
0.08374
0.12949
297 | P a g e
3
25
0.01304
0.11383
0.05929
0.07359
0.05947
0.08655
REFERENCES [1] K. Rosaiah, R.R.L. Kantam and V.L. Narasimham, ML and Modified ML estimation in gamma distribution with known prior relation among the parameters, Pakistan J. of Statist., Vol.9(3)B, 1993a, 37-48. [2] K. Rosaiah, R.R.L. Kantam and V.L. Narasimham, On modified maximum likelihood estimation of gamma parameters, J. of Statistical Research, Bangladesh, 27, Nos.1 & 2, 1993b, 15-28. [3] R.R.L. Kantam and G. Srinivasa Rao, Reliability estimation in Rayleigh distribution with censoring some approximations to ML Method, Proceedings of II Annual Conference of Society for Development of Statistics, Acharya Nagarjuna University, 1993, 56-63. [4] R.R.L. Kantam and G. Srinivasa Rao, Log-logistic distribution: Modified Maximum likelihood estimation, Gujarat Statistical Review, 29, Nos.1-2, 2002, 25-36. [5] R.R.L. Kantam and B. Sriram, Maximum likelihood estimation from censored samples- Some modifications in length biased version of exponential model, Statistical methods, 5(1), 2003, 63-78. [6] R. Subba Rao, R.R.L. Kantam and G. Prasad, Modified Maximum Likelihood Estimation in Pareto-Rayleigh Distribution, Conference Proceedings of National Seminar on Recent Developments in Applied Statistics – Golden Research Thoughts, 2015, 140-152. [7] R. Subba Rao, R.R.L. Kantam and G. Srinivasa Rao, Pareto Distribution – Some methods of Estimation, International Journal of Computational Mathematical Ideas, Vol.2, No.2, 2010, 82-92. [8] M.L. Tiku, Estimating the Mean and Standard Deviations from a Censored Normal Sample, Biometrika, Vol.4, 1967, 155-165. [9] N. Balakrishnan, Approximate Maximum Likelihood Estimation for a Generalized Logistic Distribution, Journal of the Statistics Planning & inference, Vol.26, 1990, 221-236. [10] M.L. Tiku, W.Y. Tan and N. Balakrishnan, Robust Inference, Marcel Dekker, INC, New York, 1986.
298 | P a g e