356 IEEE TRANSACTIONS ON RELIABILITY, VOL. R-26, NO.

5, DECEMBER 1977

Monte Carlo Comparisons of Parameter Estimators

Alan J. Gross, Associate Member IEEE and does provide a nearly unbiased estimator of a. However,
Dan Lurie the estimator by Thoman et al, is not an ML estimator (it is
an ML estimator multiplied by a constant). Furthermore,
Key Words-Bain-Antle estimators, Maximum likelihood estimators, even if this nearly unbiased estimator were used in small
Monte Carlo simulation, Weibull distbution. samples, it requires the calculation of the ML estimator and
thus the convergence problems that are often present in small
Reader:Aids-
Purpose: Widen state of the art
sample ML estimates of a are not remedied.
Special math needed: Basic statistical theory
Results useful to: Reliability engineers and theoreticians 2. ESTIMATES USED IN THE COMPARATIVE
PROCEDURES
Abstract-A Monte Carlo Simulation was carried out in order to
compare three different estimators of the 2-parameter Weibull dis- Let R (x) R (x; 0, a) be given by (1). If y = R (x), it is
tribution. The estimators were the ML (maximum likelihood)
estimators and two other estimator pairs suggested by Bain & Antle. well known that y is uniformly distributed on (0,1). Thus, to
The Bain-Antle estimators are better than the ML estimator for small generate a random sample of n joint Weibull variables xl, ...,
samples (in that their bias, standard deviation, and rms error are
smaller), whereas the ML estimator is superior in large samples.
xn it suffices to generate a random sample of n joint uniform
variables y,. Yn on (0,1) and make the transformation

I. INTRODUCTION
xi/0 = (- 1nyi)t (2)
The 2-parameter Weibull Sf is
for fixed ax and 0. Since 0 is a true scale parameter, it can be
R(x; 0, a) = exp[- (x/0fa], x > 0, a > 0, 0 > 0. (1) set equal to one. Thus, only n and a need be varied in generat-
ing x1, , Xn
Since this distribution is used extensively in failure analysis In each generated sample of size n three pairs of estimates
studies, much attention has been paid to methods of estimating are computed for 0 and a. The estimates are: 01 and al, the
0 and a. ML estimates; 02 and 3. The latter two sets
Methods of estimating 0 and a have been investigated by are Bain-Antle estimates. The notation 1i and 7ri imply the
many authors. Cohen [2,3] and Harter & Moore [5] derive sum and product, respectively, over i = 1 to n.
the maximum likelihood (ML) estimators. Menon [7], The ML estimators, analogous to those derived by Cohen
Gumbel [4], Miller & Freund [8], and Bain & Antle [1] all [2,3], are
consider simple alternatives to the ML method to obtain
estimators of 0 and a. Their purpose in obtaining these =
A A
.x.&l)/nI£l1
simplified estimates is to avoid the rather involved computa-
tions required by ML estimation. The solution for a' (the ML A A

estimate of a) usually uses iteration such as the Newton- (1ixial ln xi)/lixi'l - (Xi ln xi)/n - all - 0. (4)
Raphson technique; so convergence problems can and often
do occur in small samples especially when a is less than one. The first set of Bain-Antle estimators [1] is
The purpose of this article is to compare the two estimators
of Bain & Antle [1] to the ML estimators. Monte Carlo simu- - = n-1 (nca2)-1 (5)
lation is used for several sample sizes and values of 0 and a. 02 7i X(i) /7ii(
Only complete samples are considered here. Such com-
parisons are useful to the engineer and scientist who a2= S /52 (6)
ordinarily use ML estimates for Weibull parameters. As is
demonstrated by these comparisons, the simple estimators x(l) 6 x 2S. . X(n) is the ordered sample of Weibull
proposed in [1] are useful alternatives to the ML estimators observations;
when-
1. sample sizes are small t in-j+1
2. a<l.
Thoman et al. [91 provide a table of unbiasing factors for & for Y i-ln x(s), Zi~ lnt, y and Syare the usual sum of cross
48 sample sizes between 5 and 120. This table is quite useful products for y and z and sum of squares for y, respectively.
IEEE TRANSACTIONS ON RELIABILITY, VOL. R-26, NO. 5, DECEMBER 1977 357

The second set of Bain-Antle [1] is 1. With 0 = 1 specific values of a are chosen for this simu-
lation so that the reader can tell at a glance the bias associated
i="r-x ) I.W.(na3)1 (7) with the estimators al, a2, 03 in estimating a and of the
estimators 01, 02, 03, in 'recapturing' 0 = 1. The rms error
03= 2
S/S; (8) (RMSE) is also given for each estimator. The simulation with
0 = 1 is equivalent to the use of a pivotal function whose dis-
tribution is independent of the unknown parameter(s) (0
where wi - ln [1 - i/(n + 1)], and vi ln wi; i = 1, 2, ... n. in this case). See Lemon [6] or Thoman et al. [9] for a
discussion of pivotal functions.
3. MONTE CARLO SIMULATIONS 2. The new simulation compares both of the Bain-Antle
estimators with the ML estimator and with each other,
The three steps in the simulation are as follows: although it does not compare the other estimators mentioned
in [1].
1. Values of a and n used in the simulation are: a = 0.5, 1.0, 3. Comparisons are made for a range of Weibull distribu-
5.0; n = 10, 25, 50, 100. This gives a total of 12 comparisons. tions including those with increasing, decreasing, and con-
While this is not an all inclusive list, it does include a decreas- stant hazard rates.
ing hazard rate (a = 0.5), a constant hazard rare (a = 1.0) and an From the simulation results the following conclusions can be
increasing hazard rate (a = 5.0). made:
2. For each fixed value of a and n, 1000 random Weibull - In samples of n . 10 both Bain-Antle estimators are
samples were drawn by means of (2) with 0 1, and a psuedo- superior to the ML estimators. In fact the second Bain-Antle
random number generator from the uniform distribution estimator is the best estimator even though its bias in
U(0,1).2.W U(O, I). ~ ~ ~ ~ ~ ~ ~ estimating
2. When
at iS larger than the first Bain-Antle estimator.
eth sapeszn=25teM
the sample esiaobenso
size n = 25, the ML estimator
3. For each of the 1000 random samples, the estimators begins to
& Oi; i = 1, 2, 3, were computed each pair with its cor- act more favorably. However, on balance, the second Bain-Antle
estimator is still the best among the three on the basis of
responding bias, standard
' deviation,' and rms error. ~~~~~~~~~~overall
smallest bias and RMSE for the two parameters.
3. For sample sizes of 50 or larger, the desirable large
4. DISCUSSION OF THE SIMULATION sample properties of the ML estimators prevail. This is, the
ML estimator is the best among the three on the basis of
Prior to discussing the simulation, it is important to point overall smallest bias and RMSE for the two parameters.
out its features and the difference between it and the simula- The main idea that this article endeavors to convey
tion of Bain & Antle [1]. is that when a reliability engineer or statistician ought to

TABLE 1
Monte Carlo Simulation of the Weibull Parameters Based on 1000 Random Samples
Bias Standard Deviation RMS Error
n a1 a2 a3 61 2 63 a1 a2 a3 61 02 83 a1 a2 a3 e1 e2 63
10 .081 -.052 -.068 .117 .113 .002 .176 .155 .148 .571 .395 .339 .192 .161 .161 .583 .410 .339
25 .031 -.038 -.046 .024 .029 -.020 .089 .095 .089 .232 .214 .200 .094 .100 .100 .232 .217 .200
50 .015 -.030 -.035 .009 .007 -.019 .055 .071 .063 .152 .141 .138 .063 .071 .077 .152 .141 .138
100 .009 -.021 -.024 .003 -.002 -.016 .045 .045 .045 .105 .100 .100 .045 .055 .055 .105 .100 .100
(a-0.5, 9-1.0)
10 .172 -.094 -.127 .103 .103 -.006 .359 .313 .300 .510 .402 .343 .398 .325 .325 .520 .415 .343
25 .050 -.097 -.113 .039 .039 -.010 .170 .179 .176 .245 .219 .205 .176 .205 .207 .247 .224 .205
50 .026 -.064 -.073 .018 .014 -.013 .114 .138 .134 .158 .148 .141 .118 .152 .152 .158 .148 .141
100 .011 - .046 -2 . . -.015 .084 .100 .100 .110 . 0 .105 .084 .110 .114 .110 .105 .105
(a-1.0, 9-1.0)
10 .828 -.478 -.643 .115 .111 .000 1.722 1.440 1.380 .501 .375 .324 1.911 1.518 1.522 .514 .391 .324
25 .323 -.382 - .466 . 035 .036 -.014 .876 . 952 .929 .242 .219 .205 .934 1.026 1.040 . 249 .221 . 205
50 .158 -.283 - .332 .018 .015 -.012 .5S78 .714 .704 .158 .148 .145 .607 .768 .778 .161 . 152 . 145
100 .076 - .195 - .223 . 007 .004 -.011 .408 .5S12 . 509 .110 .110 .105 .416 .5S49 .55S6 .110 .110 .105
(a-5.0, 0-1.0)
358 GROSS/LURIE: MONTE CARLO COMPARISONS OF PARAMETER ESTIMATORS OF THE WEIBULL DISTRIBUTION

be careful in determining which parameter estimation [8] I. Miller, J.E. Freund, Probability and Statistics for Engineers,
scheme should be used when fitting a Weibull distribution to Prentice Hall, Inc., Englewood Cliffs, New Jersey, 1958.
failure data. When dealing with sample sizes of 25 or [9] D.R. Thoman, L.J. Bain, C.E. Antle, "Inferences on the
parameters of the Weibull distribution", Technometrics, vol 11,
less, it is not the best procedure to use ML estimators. 1969 Aug, pp 445-4601
ACKNOWLEDGMENTS
Gross A. J.; Department of Biometry; Medical University of South
The following people have been instrumental in the pre- Carolina; 171 Ashley Avenue; Charleston, SC 29403 USA.
paration of this article: Rocco L. Brunelle of Eli Lilly Com-
pany. Min-An Huang of the Ohio State University, an Dr. Alan J. Gross is a Professor of Biometry at the Medical University
anonymous referee whose comments on an earlier draft of South Carolina. From 1971 to 1976 he was Associate Professor
of the article were very valuable, and the Editor for of Public Health at the University of Massachusetts. Previous assign-
ments were at UCLA, RAND Corporation, Booz-Allen Applied
suggesting many improvements. Research, and the National Cancer Institute. Dr. Gross has
published more than two dozen technical articles appearing in
REFERENCES various statistical, biomedical, and reliability related journals in-
cluding these transactions. He is co-author of the book Survival Dis-
[1] L.J. Bain, C.E. Antle, "Estimation of Parameters in the Weibull tributions: Reliability Applications in the Biomedical Sciences
distribution", Technometrics, vol 9, 1967 Nov, pp 621-627. published by John Wiley & Sons, 1975.
[2] A.C. Cohen, "Maximum likelihood estimation in the Weibull
distribution based on complete and censored samples", Lurie Dan; Department of Biometry; Medical University of South
Technometrics, vol 7, 1965 Nov, pp 579-588. Carolina; 171 Ashley Avenue; Charleston, SC 29403 USA.
[3] A.C. Cohen, "life testing and early failure" (Query), Tech-
nometrics, vol 8, 1966 Aug, pp 539-545. Dr. Dan Lurie is a statistical consultant, Applied Statistics Group,
[4] E.J. Gumbel, Statistics of Extremes, Columbia University Press, Nuclear Regulatory Commission, Bethesda, Maryland. From 1971
New York, 1958. until 1976, Dr. Lurie was an Assistant and subsequently Associate
[5] H.L. Harter, A.H. Moore, "Maximum likelihood estimation Professor of Biometry at the Medical University of South Carolina.
of the parameters of gamma and Weibull populations from From 1964 until 1967 he was a consultant at the Biometrics Section
complete and censored samples", Technometrics, vol 7, 1965 of the School of Aerospace Medicine, Brooks AFB, Texas. He
Nov, pp 639-643. received his BS (Mathematics, SMU) in 1961, his MS (Statistics,
[6] G.H. Lemon, "Maximum likelihood estimation for the three SMU) in 1964, and his PhD (Statistics, Texas A&M)
parameter Weibull distribution based on censored samples",
X ' May,' pp 247-254.
Technometrics, vol 17, 1975
~~~~~~~~~~Manuscriptreceived 1977 January 27; revised 1977 August 3.
[7] M.V. Menon, "Estimation of the shape and scale parameters
of the Weibull distribution, Technometrics, vol 5, 1963 May,
pp 175-182. nun1

Correspondence Correspondence items are not refereed

Hence REFERENCE
3
PiPi = P1P2G1t1 + + P2P3(J.2 + [5] U. Isphording, "Methods of calculating the reliability parameters
i=1 of redundant complex systems", Electronics and Communica-
+ P3P1 (43 + 1ul) 2P1P2P3(§p1 + 92 + 93).
- tion (A.E. U) vol 22, 1968, pp 337-342, (Full text in German).

Generally, one can take the polynomial form of Ps X multiply P5 Dr. W. Schneeweiss; D 75 Karlsruhe; Bertha-v.-Suttner-Str. 3a; WEST
by .Uand multiply each term PiP1Pk ..............
by (jui + aj+ Uk+*-.......)..GERMANY
Also one can start from a "functional tree", find a polynomial
form of the availability A~of the system in terms of the avail-
abilities of the subsystems, multiply As by X5 (the failure rate) Letrrcie'7Jl5 eie 7Ag 3
and multiply each term A iA1A k *@by (Xi + Xi+ Xk+)
I thank Dr. Singh for having initiated this discussion.