Methods for Estimating the Parameters of the

Weibull Distribution

Mohammad A. Al-Fawzan

King Abdulaziz City for Science and Technology

P.O. Box 6086, Riyadh 11442, Saudi Arabia.
Fax: +966 1 4813991
Email: mfawzan@[Link]

May 2000
Revised October 2000
 Abstract

In this paper, we present some methods for estimating Weibull parameters,

namely, shape parameter ( ) and scale parameter ( ). The Weibul distribution is an
important distribution especially for reliability and maintainability analysis. The
presented methods can be classified into two categories: graphical and analytical
methods. Computational experiments on the presented methods are reported.

1. Graphical Methods

Usually, the graphical methods are used because of their simplicity and
speed. However, they involve a great probability of error. Next we discuss two main
graphical methods.

1.1 Weibull Probability Plotting

The Weibull distribution density function (Mann et al. (1974)) is given by:
x
x

( )
f (x) = ( ) 1
e
, > 0 , > 0 , x 0 (1)

The cumulative Weibull distribution function is given by:
x
( )

F (x) = 1 e (2)

where; is the shape parameter, is the scale parameter, and is the location
parameter.

To come up with the relation between the CDF and the two parameters
( , ), we take the double logarithmic transformation of the CDF.

From (2) and letting = 0, we have

2
 x
( )

1 F (x) = e

x
( )
1
= e
1 F (x)

1 x
ln[ ] = ( )
1 F (x)

1
ln ln[ ] = ln ln x
1 F (x)
The last equation is an equation of a straight line. To plot F (x) versus x , we
apply the following procedure:

1. Rank failure times in ascending order.

2. Estimate F (x i ) of the i th failure.
3. Plot F (x i ) vs. x in the Weibull probability paper.
To estimate F (x i ) in (2 and 3) above, we may use one of the following
methods presented in Table 1 where n is number of data points.

Method F ( xi )
Mean Rank i
n +1
Median Rank i 0. 3
n + 0. 4
Symmetrical CDF i 0. 5
n
Table 1. Methods for estimating F ( x i ) .

1.2 Hazard Plotting Technique

The hazard plotting technique is an estimation procedure for the Weibull

parameters. This is done by plotting cumulative hazard function H (x ) against failure
times on a hazard paper or a simple log-log paper.
The hazard function is given below:
x
( ) 1
h(x) =

3
The cumulative hazard function is given below:
x
H ( x) = h ( x) = ( ) (3)

We can transform (3) by taking the logarithm as follows

ln H ( x ) = { ln x ln }

1
ln x = ln H ( x ) + ln (4)

From (4), we can then plot ln H ( x) versus ln x using the following procedure:

1. Rank the failure times in ascending order.

2. For each failure, calculate i = 1
( n +1 )1
3. For each failure, calculate H = 1 + 2 + + i
4. Plot ln H vs. ln x .
5. Fit a straight line.

Upon completing the plotting, the estimated parameters will be as follows:

x 1
= =
y slope
at H = 1, = x

2. Analytical Methods
Due to the high probability of error in using graphical methods, we prefer to
use the analytical methods. This is motivated by the availability of high-speed
computers. In the following, we discuss some of the analytical methods used in
estimating Weibull parameters.

2.1 Maximum Likelihood Estimator (MLE)

The method of maximum likelihood (Harter and Moore (1965a), Harter and
Moore (1965b), and Cohen (1965)) is a commonly used procedure because it has
very desirable properties. Let x1 , x 2 , ..., x n be a random sample of size n drawn
from a probability density function f X ( x; ) where is an unknown parameter. The
likelihood function of this random sample is the joint density of the n random
variables and is a function of the unknown parameter. Thus
n
L = f X ( x i , )
i (5)
i= 1

4
is the likelihood function. The maximum likelihood estimator (MLE) of , say $ ,
is the value of that maximizes L or, equivalently, the logarithm of L . Often, but
not always, the MLE of is a solution of
d log L
=0
d
where solutions that are not functions of the sample values x1 , x 2 , ..., x n are
not admissible, nor are solutions which are not in the parameter space. Now, we are
going to apply the MLE to estimate the Weibull parameters, namely the shape and
the scale parameters. Consider the Weibull pdf given in (1), then likelihood function
will be
xi
)
x n (
L ( x1 , , x n ; , ) = ( )( i ) 1 e (6)
i =1
On taking the logarithms of (6), differentiating with respect to and in turn and
equating to zero, we obtain the estimating equations
ln L n n 1 n
= + ln xi xi ln xi = 0 (7)
i =1 i =1
ln L n 1 n
= + x i = 0 (8)
2
i=1

On eliminating between these two equations and simplifying, we have

n
x i ln xi
1 1 n
i=1
ln xi = 0
(9)
n
n
x i i=1

i =1

which may be solved to get the estimate of $ = . This can be accomplished by the
use of standard iterative procedures (i.e., Newton-Raphson method). Once is
determined, can be estimated using equation (8) as
n
x i
= i=1
(10)
n

2.2 Method of Moments (MOM)

The method of moments is another technique commonly used in the field of

parameter estimation. If the numbers x1 , x 2 , ..., x n represent a set of data, then an
unbiased estimator for the kth origin moment is
1 n
m k = xik (11)
n i =1

where; mk stands for the estimate of mk . In Weibull distribution, the kth moment
readily follows from (1) as

5
 1 k
k
k = (
) (1 + )

(12)

where signifies the gamma function

(s) = 0 xs 1 e x dx,( s > 0)
Then from (12), we can find the first and the second moment as follows
1
1 1
m 1 = k = ( )
(1 + ) (13)

1
2 2
1 2
m2 = + =
2 2

1 + 1 + (14)

k k

When we divide m 2 by the square of m1 , we get an expression which is a function of

only
2 1
(1 + ) 2
(1 + )
k2
=
k2 2
(1 +
1
)

(15)
On taking the square roots of (15), we have the coefficient of variation
2 1
(1 + ) 2 (1 + )

CV = (16)
1
(1 + )

Now, we can form a table for various CV by using (16) for different
values. In order to estimate and , we need to calculate the coefficient of
variation ( CV )d of the data on hand. Having done this, we compare ( CV )d with
CV using the table. The corresponding is the estimated one ($ ). The scale
parameter ( ) can then be estimated using the following
= {x [(1 ) + 1 ]}

where x is the mean of the data.

2.3 Least Squares Method (LSM)

The third estimation technique we shall discuss is known as the Least Squares
Method. It is so commonly applied in engineering and mathematics problems that is
often not thought of as an estimation problem. We assume that a linear relation
between the two variables (see section 1). For the estimation of Weibull parameters,
we use the method of least squares and we apply it to the results of section 1. Recall
that
1
ln ln[ ] = ln x ln (17)
1 F (x)

6
Equation (17) is a linear equation. Now, we can write

1 n 1
x = ln ln
(18)
n i =1 (1 i )
n + 1

1 n
Y = ln x i (19)
n i=1

n 1 n 1 n
n (ln x i ) (ln{ln[ ]}) ln(ln[ ]) ln x i
i=1 (1 n + 1 )
i
i =1 (1 n + 1 )
i

=
i =1

2 (20)
2 n
n
n (ln x i ) (ln x i )
i=1 i=1
(Y x )
= e (21)
From equations (18)-(21), we can calculate the estimate of and .

3. Numerical Example
In order to illustrate and compare the methods described earlier, we have
coded the three analytical methods MLE, MOM and LSM in BASIC Language and we
have used a Pentium PC with 133MHz . To compare the three methods, we used the
MSE (Mean Squared Error) test. MSE can be calculated as below
n
MSE = { F ( x i ) F ( x i )} 2 (22)
i =1

( x )

i 0 .3
where F ( x i ) = 1 e i and F (xi ) =
n + 0 .4

Example 1:
Consider the following example where xi represents the ith failure time.
Table 2. Data for example 1.
i xi
1 0.438
2 2.413
3 3.073
4 3.079
5 3.137
6 3.198
7 3.918
8 4.287
9 4.508
10 4.981

7
 11 5.115
12 5.592
13 5.848
14 5.958
15 6.013

Table 3 below shows the complete results of the estimates of the shape and scale
parameters and the mean squared error (MSE) for each method.

Table 3. Results of example 1.

Method MSE
MLE 2.923 4.552 3. 57 10 3
MOM 2.941 4.587 3. 17 10 3
LSM 1.8515 4.756 8. 4 10 3

Since MOM has the minimum MSE, then =2.941 and =4.587.

In comparing our results with the well-known software called

STATGRAPHICS, we found that STATGRAPHICS is using the MLE and our MLE
estimates are exactly the same as the estimates of STATGRAPHICSs MLE.
However, we have the advantage of having different estimates for different methods
and the ability of choosing one of them.

4. Computational Results
The objective of our experiments is to compare the three methods namely,
LSM, MLE and MOM. We have generated random samples with known parameters.
For each sample, we have varied the size from 20 to 100. To be able to compare, we
calculated the total deviation (TD) for each method as follows:

TD = + (23)

where; and are the known parameters, and and are the estimated parameters
by any method.

Table 4 shows the complete results. The last column of the table shows the
best method which yields the minimum total deviation. Notice that the maximum TD
is 0.55 for all methods. This means that at the worst case the estimated parameters
are within 50% of their actual values.

Table 5 shows a summary of the results of Table 4. As it is obviously seen,

MOM is the best method. MOM achieves the best estimate 12 times out of 25 which

8
is approximately 50% of the time. Its average deviation from the actual values is
17% with a standard deviation of 14% which is quite good. MOM achieves good
results because it involves more calculations and require more computation time
than LSM or MLE. However, for a sample of size 100, MOM takes only few
seconds.

Table 4. Comparison between LSM, MLE and MOM

n S LSM MLE MOM Best
i TD TD TD
z
e
1 1 10 20 1.34 7.08 0.63 1.22 7.3 0.5 1.1 7.09 0.4 MOM
40 0.85 10.96 0.24 0.94 10.74 0.13 0.98 10.8 0.1 MOM
60 0.84 12.17 0.37 0.926 11.87 0.26 0.94 11.9 0.25 MOM
80 1.07 10.83 0.16 1.11 10.8 0.21 1.15 10.9 0.24 LSM
100 0.87 11.18 0.24 0.93 10.97 0.16 0.93 10.8 0.15 MOM
2 2.3 145 20 2.36 180.3 0.27 2.7 178.4 0.41 2.7 178.1 0.4 LSM
40 3.4 155.1 0.55 3.3 155.8 0.51 3.33 155.2 0.52 MLE
60 2.19 132.5 0.13 2.38 131.8 0.13 2.35 131.6 0.11 MOM
80 2.43 139.06 0.09 2.25 140.5 0.05 2.25 140.1 0.05 MLE
100 2.44 156.9 0.14 2.62 155.6 0.21 2.6 155.6 0.2 MOM
3 2.9 357 20 2.91 358.7 0.01 3.35 354.9 0.16 3.33 354.1 0.15 LSM
40 3.32 365.7 0.14 3.6 364.1 0.25 3.4 363.6 0.19 LSM
60 2.63 373.8 0.13 2.95 370.5 0.05 2.9 370.4 0.04 MOM
80 2.62 373.04 0.14 3.03 367.8 0.07 3.0 367.8 0.06 MOM
100 2.7 367.6 0.1 3.02 363.7 0.06 3.0 367.4 0.05 MOM
4 3.5 1270 20 3.77 1276.4 0.08 4.4 1270.3 0.27 4.1 1264.6 0.19 LSM
40 3.12 1200 0.16 3.57 1192.4 0.08 3.4 1189.6 0.09 MLE
60 3.07 1250.7 0.13 3.3 1248.5 0.06 3.2 1245.3 0.1 MLE
80 3.09 1304.5 0.14 3.8 1282.8 0.1 3.6 1284.2 0.04 MOM
100 3.46 1230.9 0.04 3.52 1228.2 0.03 3.57 1223.5 0.05 MLE
5 1.9 872 20 1.5 888.8 0.22 1.83 867.7 0.03 1.9 870.6 0.01 MOM
40 2.1 949.6 0.19 2.6 920.9 0.47 2.8 926.7 0.53 LSM
60 2.3 957.3 0.32 2.25 962.3 0.3 2.25 959.9 0.28 MOM
80 1.64 907.8 0.17 1.95 884.7 0.04 2.0 890.2 0.08 MLE
100 1.91 827.3 0.06 1.96 824.8 0.08 1.98 824.8 0.1 LSM

9
 Table 5. Summary of Results
Method No. of times the Average Standard deviation
method gives the percentage of of the devi ations
best estimate deviation from
actual values ()
()
LSM 7 21 % 15 %
MLE 6 18 % 15 %
MOM 12 17 % 14 %

5. Method Selection
We have described two graphical methods and three analytical methods for
estimating Weibull parameters and . Now, the question is which method do we
use? The answer depends on whether one needs a quick or an accurate estimation. In
what follows, the methods are ranked according to their accuracy or speed. The
order of the methods based on their speed (computing time) are
1. Any graphical method.
2. Least Squares Method.
3. Maximum Likelihood Estimator.
4. Method of Moments.

The order of the methods based on their accuracy are

1. Applying the three analytical methods (MLE,MOM and LSM)
and selecting the best one which gives the minimum mean squared
error.
2. Method of Moments.
3. Maximum Likelihood Estimator.
4. Least Squares Method.
5. Any graphical method.

6. Conclusion
In this paper, we have presented both graphical and analytical methods for
estimating the Weibull distribution parameters. It has been shown from the
computational results that the method which gives the best estimates is the method
of moments.

10
References
1. Mann, N. R., Schafer, R. E., and Singpurwalla, N. D., Methods for statistical
analysis of reliability and life data, 1974, John Wiley and Sons, New York.

2. Engelhardt, M., "On simple estimation of the parameters of the Weibull or

extreme-value distribution", Technometrics, Vol. 17, No. 3, August 1975.

3. Mann, N. R. and K. W. Fertig , "Simplified efficient point and interval estimators

of the Weibull parameters", Technometrics, Vol. 17, No. 3, August 1975.

4. Cohen, A. C., "Maximum likelihood estimation in the Weibull distribution based

on complete and on censored samples", Technometrics, Vol. 7, No. 4, November
1965.

5. Harter, H. L. and A. H. Moore, "Point and interval estimators based on order

statistics, for the scale parameter of a Weibull population with known shape
parameter", Technometrics, Vol. 7, No. 3, August 1965a.

6. Harter, H. L. and A. H. Moore, "Maximum likelihood estimation of the parameters

of Gamma and Weibull populations from complete and from censored samples",
Technometrics, Vol. 7, No. 4, November 1965b.

7. Stone, G. C. and G. Van Heeswijk, "Parameter estimation for the Weibull

distribution", IEEE Trans. on Elect. Insul., Vol. EI-12, No. 4, August 1977.

11