Journal of the European Ceramic Society 26 (2006) 1099–1105

Unbiased estimation of Weibull parameters

with the linear regression method
Dongfang Wu a, ∗ , Jiancheng Zhou a , Yongdan Li b
a Department of Chemistry and Chemical Engineering, Southeast University, Nanjing 210096, China
b Department of Catalysis Science and Technology, School of Chemical Engineering, Tianjin University, Tianjin 300072, China

Received 13 October 2004; received in revised form 13 January 2005; accepted 23 January 2005
Available online 26 February 2005

Abstract

Monte Carlo simulations were used to search for the probability estimator for the unbiased estimate of the Weibull parameters in the linear
regression method. Compared with commonly-used probability estimators, the estimator proposed gives a more accurate estimation of the
Weibull modulus and the same estimation precision of the scale parameter. It is found that the estimator proposed is more conservative than
the estimator Pi = (i − 0.5)/n recommended by previous authors, and hence results in a higher safety in reliability predictions. The unbiased
properties of the estimated Weibull parameters were validated with actual experimental data. It is also concluded that the estimated Weibull
modulus from actual experimental data is more dispersive than that from Monte Carlo simulation, which arises from the fact that the strength
data from actual experiments does not perfectly follow the Weibull statistics.
© 2005 Elsevier Ltd. All rights reserved.

Keywords: Fracture; Strength; Mechanical properties; Failure analysis; Weibull statistics

1. Introduction where m and σ 0 are the Weibull modulus and the scale pa-
rameter, respectively. The Weibull modulus m, also called
Weibull statistics has been commonly used to character- the shape parameter, represents the scatter in the fracture
ize the statistical variation in the fracture strength of brittle strength. A higher m leads to a steeper distribution func-
materials such as ceramics, glasses and solid catalysts.1–4 It tion and thus a lower dispersion of the fracture strength. The
is based on a “weakest link theory”, which means that the scale parameter σ 0 corresponding to the fracture stress with
most serious flaw in the material will control the strength, a failure probability of 63.2% is closely related to the mean
like a chain breaking if the weakest link fails. The most seri- strength of the distribution, σ̄.1,5
ous flaw is not necessarily the largest one because its severity  
also relies on its location and orientation. In other words, the 1
σ̄ = σ0 Γ 1+ (2)
flaw subjected to the highest stress intensity factor will be m
strength controlling.
Using Weibull’s two-parameter distribution, the cumula- where Γ is the gamma function. For the Weibull modulus
tive probability of failure P at or below a stress σ is repre- of 5–20, a typical range for technical ceramics,5 Γ (1 + 1/m)
sented by1,2,5 takes values between 0.9 and 1.

 m  There are several methods available in the literature,5–14
σ for the determination of the Weibull distribution parameters
P = 1 − exp − (1)
σ0 from a set of experimentally measured fracture stresses. It
has been shown that the maximum likelihood (ML) method
∗ Corresponding author. leads to the highest estimation precision of the Weibull mod-
E-mail address: dfwu@seu.edu.cn (D. Wu). ulus, which has been recommended by previous authors.1,5,7

0955-2219/$ – see front matter © 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jeurceramsoc.2005.01.044
1100 D. Wu et al. / Journal of the European Ceramic Society 26 (2006) 1099–1105

However, the most widely used may be the linear regression the unbiased estimation of the Weibull parameters. A Monte
(LR) method due to its simplicity. Moreover, the ML method Carlo simulation was used for this purpose.
results more often in an overestimation of the Weibull modu-
lus than underestimation, and hence results in a lower safety
than the LR method in reliability prediction.10,14 From an 2. Monte Carlo simulation
engineering point of view, the LR method is, therefore, to be
preferred. From Eqs. (3a)–(3d), it is shown that the probability esti-
In the LR method, the measured fracture stresses are mator should have the following general expression.
ranked in ascending order and then a probability of failure Pi ,
is assigned to each stress σ i . Since the true value of Pi is un- i−α
Pi = (5)
known, a prescribed estimator has to be used. The following n+β
four expressions are often applied to define the probability
estimator.8–12 By varying the values of α and β, large numbers of functions
can be obtained. Clearly, the α-value is required to be less than
i − 0.5 unity; otherwise a negative value of the probability of failure
Pi = (3a)
n in case of i = 1 arises. Due to the requirement of statistical
i inference, the β-value is also necessary to be very small,
Pi = (3b) usually not larger than unity, especially for n ≤ 50.14 The
n+1
disadvantage of a large β-value is that a poor precision of
i − 0.3 the estimated scale parameter takes place.13 Therefore, the
Pi = (3c)
n + 0.4 numeric areas of the α and β-values are generally 0 ≤ α < 1
i − 3/8 and 0 ≤ β ≤ 1, respectively.
Pi = (3d) If α and β are restricted to two decimals, a total of 101
n + 1/4
values are obtained for β and 100 values for α due to α < 1.
where Pi is the probability of failure for the ith ranked stress And then α = 0.999 is added. As a result, 1012 possible com-
datum, and n is the sample size. binations of the α and β-values are produced. In order to
By taking the logarithm twice, Eq. (1) can be rewritten in search for the combination leading to an unbiased estimate
a linear form. of the Weibull modulus, a Monte Carlo simulation was used,
  as shown in Fig. 1.
1
ln ln = m lnσ − m lnσ0 (4) Eq. (1) can be rewritten as
1−P

 1/m
The Weibull modulus can thus be obtained directly from the 1
σ = σ0 ln (6)
slope term in Eq. (4) and the scale parameter can be deduced 1−P
from the intercept term.
However, the estimators of the Weibull modulus are al- If we consider a large “specimen” population with prescribed
ways biased for both the ML and LR methods.5–7,15,16 In m and σ 0 values, i.e. mtrue and σ 0,true , random strength data
most cases the bias increases rapidly as the sample size can be obtained from Eq. (6) provided random numbers be-
decreases. In the LR method, different probability estima- tween 0 and 1 are substituted for the probability of fail-
tors also lead to different biases of the estimated Weibull ure P. For the sake of convenience, we let mtrue =10 and
modulus. It has been shown that Eq. (3b) gives the largest- σ 0,true = 1 throughout this study. A computer program was
biased estimate of the Weibull modulus, while Eq. (3a) written, which used a sample of random numbers to obtain
results in the least bias for n ≥ 20.9–12 To overcome this strength values σ 1 , σ 2 , . . ., σ i , . . ., σ n . This set of strength val-
shortcoming, several authors proposed the use of a cor- ues, regarded as a fictitious sample, was ranked in ascending
rection factor to adjust the bias of the estimated Weibull order and the probabilities of failure were calculated from Eq.
modulus.9,15,16 However, both analytical analyses and nu- (5) with each combination of the α and β-values. The sample
merical calculations revealed that each set of strength data was then analyzed with the LR method to give the estimated
gives the statistically correct Weibull parameters and that value of the Weibull modulus. This procedure was repeated
the bias arises only from the method of adding the param- 10,000 times. Consequently, a total of 10,000 samples were
eters, if one tries to obtain a mean value from a number generated and 10,000 estimated values of the Weibull modu-
of sets of strength data.17 In practice, when only one set of lus were obtained for each combination of the α and β-values.
strength data is available, the correction factor should not be Then the mean value m̄, standard deviation Sm and coefficient
applied. of variation CVm of these moduli were computed from
The probability estimator has a significant effect on the
4
bias of the estimated Weibull modulus in the LR method. 
10
mj
The objective of this paper is to try to find appropriate ex- m̄ = (7)
j=1
104
pressions to define the probability estimator, which leads to
 D. Wu et al. / Journal of the European Ceramic Society 26 (2006) 1099–1105 1101

3. Experimental

A sample of alumina agglomerates, often used as cata-

lyst supports in chemical industry, was investigated in this
work. The specimens are spherical in shape, with a diameter
of 5.1 ± 0.2 mm and a pellet density of 1.382 g/cm3 . They
are brittle materials fabricated deliberately to produce high
porosity and optimized pore distribution.
The tensile strength was measured using a diametral com-
pression test. To adequately characterize the strength prop-
erties of the population, a sufficiently large number of the
strength tests of 500 was performed. A ZQJ-II strength
tester made in Dalian, China, was used, described in detail
elsewhere.4,18 The fracture stress was calculated according
to the equation of Hiramatsu and Oka:19
2.8F
σ= (10)
πd 2
where F is the load at fracture and d is the specimen diameter.

4. Results and discussion

4.1. Probability estimator for the unbiased estimate of

the Weibull modulus

By Monte Carlo simulation in Section 2, the combination

of the α and β-values leading to an unbiased estimate of the
Weibull modulus was determined for each sample size, as
listed in Table 1. It is clear that α and β both are sensitive to
Fig. 1. Schematic flow diagram of the Monte Carlo simulation.
the sample size; however, there is no a distinct relationship
existing between α or β and the sample size.
4

10
(mj − m̄)2 4.2. Comparisons with the commonly-used probability
Sm 2
= (8) estimators
j=1
104 − 1
A Monte Carlo simulation procedure similar to that men-
Sm tioned in Section 2 was conducted for the commonly-used
CVm = (9)
m̄ probability estimators, Eqs. (3a)–(3d). Fig. 2 shows the de-
pendence of the normalized mean values of the estimated
where mj is the estimated Weibull modulus of the jth sample. Weibull moduli, m̄/mtrue , on the sample size n for five prob-
Clearly, the combination of the α and β-values, which makes ability estimators investigated. It can be seen that with the
m̄/mtrue be equal to unity, leads to an unbiased estimate of use of the probability estimator proposed in this work the
the Weibull modulus. normalized mean Weibull moduli are always equal to unity
For illustration of the effect of the sample size, the gen-
erated random samples were of size n = 10, 15, 20, 25, 30, Table 1
α and β-values in Eq. (5) leading to unbiased estimate
35, 40, 45 and 50. The combination of the α and β-values
leading to an unbiased estimate of the Weibull modulus was n α β
determined for each sample size. 10 0.37 0.24
It should be pointed out that, although the above simu- 15 0.54 0.85
20 0.49 0.32
lation was carried out for the arbitrarily chosen values of 25 0.47 0.13
mtrue =10 and σ 0,true = 1 its results are valid for any value of 30 0.53 0.41
mtrue and σ 0,true , for previous studies have shown that the 35 0.57 0.64
values of m̄/mtrue and Sm /mtrue , and the distribution of the 40 0.56 0.52
estimated m/mtrue are independent of the prescribed values of 45 0.51 0.14
50 0.56 0.42
mtrue and σ 0,true for the LR method.5,6,9,10
1102 D. Wu et al. / Journal of the European Ceramic Society 26 (2006) 1099–1105

Fig. 2. Estimated Weibull modulus as a function of the sample size. () Eq. Fig. 3. Dependence of the coefficient of variation of the estimated Weibull
(3a), (♦) Eq. (3b), () Eq. (3c), () Eq. (3d), and () unbiased. modulus on the sample size. () Eq. (3a), (♦) Eq. (3b), () Eq. (3c), ()
Eq. (3d), and () unbiased.

at all sample sizes examined, which reveals that this estima-

tor leads to the unbiased estimate of the Weibull modulus.
Among the commonly-used probability estimators, Eq. (3a) An overestimation of the Weibull modulus often leads to an
gives the least-biased estimate of the Weibull modulus for underestimation of the probability of failure at low stresses,
n ≥ 20. The next is Eq. (3d), followed by Eq. (3c). The esti- and hence a lower safety arises in reliability prediction.
mator, Eq. (3b) leads to the largest bias for all sample sizes Fig. 4 shows the occurrence probability of the Weibull
examined. It is also clear that as the sample size decreases the modulus overestimation, i.e. m/mtrue > 1 as a function of the
bias increases rapidly for Eqs. (3a) and (3d). These results are sample size. The higher the probability is, the lower the safety
in agreement with those reported by previous authors.9–12 is. It can be seen from Fig. 4 that the probability for the es-
In Fig. 3, the coefficient of variation of the Weibull mod- timator, Eq. (3a) is the highest; however, it is still less than
ulus is plotted as a function of the sample size. Clearly, the 50% for n ≥ 15. It implies that for the LR method with differ-
coefficient of variation is as expected decreasing with in- ent probability estimators, the underestimation of the Weibull
creasing sample size for all probability estimators. It is also modulus always occurs more frequently than the overestima-
shown that the coefficients of variation for the different esti- tion. It is also clear that the estimator for the unbiased esti-
mators are approximately equal at all sample sizes; however, mate of the Weibull modulus results in a higher safety than
it seems the estimator proposed in this work results in the Eq. (3a). However, the estimator leading to the highest safety
least coefficients of variation at most of the sample sizes. is Eq. (3b) though it gives the worst precision of estimation.
Statistics textbook tells us that the higher the probability of
computing an estimate near to the true value is, the higher the
estimation precision is. It indicates that the estimation pre-
cision of the Weibull modulus is related not only to the bias
but also to the coefficient of variation. The latter describes the
dispersion or breadth of the estimated data distribution, while
the former shows the centrality or location of the distribution.
The smaller both of them are, the more accurate the estimated
Weibull modulus is. Based on this criterion, the best probabil-
ity estimator can be judged from Figs. 2 and 3. Apparently,
the probability estimator proposed in this work, leading to
the unbiased estimate of the Weibull modulus, gives the most
accurate estimation for each sample size, and should, there-
fore, be preferred. Among the commonly-used probability
estimators, Eq. (3b) gives the worst precision of estimation,
while Eq. (3a) leads to the highest precision for n ≥ 20, which
was considered as the best probability estimator by many
authors.9–12 Fig. 4. Occurrence probability of the Weibull modulus overestimated as a
From an engineering point of view, the safety is of the function of the sample size. () Eq. (3a), (♦) Eq. (3b), () Eq. (3c), ()
first importance, while the estimation precision is the second. Eq. (3d), and () unbiased.
 D. Wu et al. / Journal of the European Ceramic Society 26 (2006) 1099–1105 1103

Table 2
Estimated Weibull parameters of experimental data
Methods m σ 0 (N/cm2 )
LR, Eq. (3a) 4.42 168.35
LR, Eq. (3b) 4.36 168.53
LR, Eq. (3c) 4.39 168.43
LR, Eq. (3d) 4.40 168.40
ML 4.43 168.23

4.3. Estimation of the scale parameter

Similar Monte Carlo simulations were conducted to pro-

duce 10,000 estimated values of the scale parameter, σ 0 , for
each probability estimator and for each sample size. The
mean value and the coefficient of variation were calculated.
And then the 10,000 estimated σ 0 were ranked in ascending
order and divided into 40 equally sized intervals. The num-
ber of estimated σ 0 falling into each interval was counted.
This number, normalized through division by 10,000, the to-
tal number of estimated σ 0 , produces the relative frequency
of occurrence, which was taken as the y-value. The midpoint
of the given interval was used as the x-value. The resulting
histogram can be regarded as an empirical probability den-
sity distribution.5 For the sake of comparison, the same data
processing was also carried out for the estimated Weibull
moduli.
As an example, the probability density distributions of the
estimated Weibull modulus and scale parameter at a sample
size of 20 are shown in Fig. 5, where the x-axes of two sub-
figures have the same range and the same scale. The results
of other sample sizes are similar to those shown in Fig. 5. It
can be seen that all probability estimators give a similar dis-
Fig. 5. Probability density distributions of the estimated Weibull modulus
tribution of σ 0 /σ 0,true that scatters in the vicinity of the true
(A) and scale parameter (B) at a sample size n = 20. () Eq. (3a), (♦) Eq.
value in a much smaller range, as compared with the Weibull (3b), () Eq. (3c), () Eq. (3d), and () unbiased.
modulus. In reality, for any sample size the coefficient of
variation of the estimated scale parameter is about one tenth
of that of the estimated Weibull modulus; therefore, the scale true values of the Weibull parameters can be obtained def-
parameter can be estimated with accuracy about an order of initely only for an infinite number of specimens. However,
magnitude higher than the Weibull modulus.5,14 From Fig. 5, a sufficiently large number of tested specimens will give a
it is also clear that for the estimation of the scale parameter, quite good approximation. Especially, the Weibull parame-
there is no significant difference between the probability es- ters estimated with the maximum likelihood method may be
timator proposed in this work and the commonly-used ones. regarded approximately as the true values.
Note that the distribution of m/mtrue is asymmetrical and The advantage of the unbiased estimate is that the mean
lightly skewed to the right; the distribution of σ 0 /σ 0,true , how- Weibull modulus from a number of sets of strength data
ever, takes an approximately symmetrical form. Similar re- should be close to its true value, which can be verified by
sults have also been reported by Khalili and Kromp.5 For any the following procedure. The 500 experimentally measured
sample size and any probability estimator, the mean value of fracture stresses were treated as a full set. To investigate the
the estimated scale parameter is always close to its true value, effects of the sample size, the subsets of the full set were se-
and its bias is negligible. Therefore, the estimator of the scale lected in groups of 10, 15, 20, 25, 30, 35, 40, 45 and 50. The
parameter obtained with any probability estimator is approx- subsets were selected randomly from the full set by assigning
imately unbiased. a computer-generated random number to each member of the
full set, sorting the stress data according to that number, and
4.4. Experimental validation then choosing the first 10, 15, etc., results in the full set. Each
subset was considered as an independent set of data, and its
Table 2 gives the Weibull parameters of the experimen- Weibull modulus was estimated with the LR method with
tal data, estimated with different methods. In principle, the the estimators, Eqs. (3a)–(3d), and the estimator proposed in
1104 D. Wu et al. / Journal of the European Ceramic Society 26 (2006) 1099–1105

In Fig. 7 the coefficient of variation of the estimated

Weibull modulus of experimental data is plotted as a function
of the sample size, where the dotted line is the curve fitting
result of the data points shown in Fig. 3. It can be seen that
the coefficient of variation of the Weibull modulus decreases
with increasing sample size, but independent of the prob-
ability estimators, as similar to the results of Monte Carlo
simulation. It is also clear that the coefficient of variation
from actual experimental data is always higher than that from
Monte Carlo simulation at any sample size. The reason for
this phenomenon is that the strength data from actual experi-
ments does not perfectly follow the Weibull statistics, which
aggravates the dispersion of the estimated Weibull modulus
of actual experimental data.

Fig. 6. Weibull modulus of experimental data as a function of the sample 5. Conclusions

size. () Eq. (3a), (♦) Eq. (3b), () Eq. (3c), () Eq. (3d), and () unbiased.
Using a Monte Carlo simulation, a probability estimator
for the unbiased estimate of the Weibull modulus was deter-
this work. By repeating this procedure, the total number of
mined. Compared with the commonly-used probability es-
the subsets for each sample size amounts to 10,000. Finally,
timators, the estimator proposed gives a more accurate es-
the mean value and the coefficient of variation of the 10,000
timation of the Weibull modulus, and the same estimation
Weibull moduli were calculated for each sample size and for
precision of the scale parameter, which is much higher than
each probability estimator.
that of the Weibull modulus. It is concluded that the estimated
Fig. 6 shows the dependence of the mean value of the
scale parameter is always approximately unbiased.
estimated Weibull modulus of actual experimental data on the
The probability estimators were also compared from an
sample size. Note that the only probability estimator which
engineering point of view. It is found that the probability
has leveled off as a function of the subset size is the one
estimator proposed results in a higher safety than Eq. (3a).
proposed in this work. And this estimator yields a Weibull
However, the estimator leading to the highest safety is Eq.
modulus, which is very close to the true value approximated
(3b) though it gives the worst precision of estimation.
by the ML method. Thus, it can be seen that the probability
It is shown that the estimated Weibull modulus from actual
estimator proposed in this work is certain to give the unbiased
experimental data is more dispersive than that from Monte
estimate of the Weibull modulus. However, the estimators
Carlo simulation, which arises from the fact that the strength
of the Weibull modulus, obtained with the commonly-used
data from actual experiments does not perfectly follow the
probability estimators, are always biased.
Weibull statistics. Finally, the unbiased property of the es-
timated Weibull modulus was validated with actual experi-
mental data. It is, therefore, recommended that the probabil-
ity estimator proposed, leading the unbiased estimate of the
Weibull parameters, should be used in the linear regression
method for estimating the Weibull parameters.

Acknowledgements

The corresponding author gratefully acknowledges finan-

cial support for this work by the Teaching and Research Pro-
gram of Southeast University, P.R.C. for Excellent Young
Teachers.

References

1. ASTM standard, C1239-00 Standard practice for reporting uniax-

Fig. 7. Coefficient of variation of the Weibull modulus of experimental data ial strength data and estimating Weibull distribution parameters for
vs. the sample size. () Eq. (3a), (♦) Eq. (3b), () Eq. (3c), () Eq. (3d), advanced ceramics. American Society for Testing and Materials,
and () unbiased. Philadelphia, PA, 2003.
 D. Wu et al. / Journal of the European Ceramic Society 26 (2006) 1099–1105 1105

2. Quinn, G. D., Flexure strength of advanced structural ceramics: a 11. Sullivan, J. D. and Lauzon, P. H., Experimental probability estimators
round robin. J. Am. Ceram. Soc., 1990, 73, 2374–2384. for Weibull plots. J. Mater. Sci. Lett., 1986, 5, 1245–1247.
3. Madjoubi, M. A., Bousbaa, C., Hamidouche, M. and Bouaouadja, 12. Papargyris, A. D., Estimator type and population size for estimating
N., Weibull statistical analysis of the mechanical strength of a glass the Weibull modulus in ceramics. J. Eur. Ceram. Soc., 1998, 18,
eroded by sand blasting. J. Eur. Ceram. Soc., 1999, 19, 2957– 451–455.
2962. 13. Wu, D. F. and Jiang, H., Comment on A new probability index for es-
4. Wu, D. F., Li, Y. D., Shi, Y. H., Fang, Z. P., Wu, D. H. and Chang, timating Weibull modulus for ceramics with the least-square method.
L., Effects of the calcination conditions on the mechanical properties J. Mater. Sci. Lett., 2003, 22, 1745–1746.
of a PCoMo/Al2 O3 hydrotreating catalyst. Chem. Eng. Sci., 2002, 57, 14. Wu, D. F., Lu, G. Z., Jiang, H. and Li, Y. D., Improved estimation of
3495–3504. Weibull parameters with the linear regression method. J. Am. Ceram.
5. Khalili, A. and Kromp, K., Statistical properties of Weibull estimators. Soc., 2004, 87, 1799–1802.
J. Mater. Sci., 1991, 26, 6741–6752. 15. Davies, I. J., Empirical correction factor for the best estimate of
6. Trustrum, K. and Jayatilaka, A. De S., On estimating the Weibull Weibull modulus obtained using linear least squares analysis. J. Mater.
modulus for a brittle material. J. Mater. Sci., 1979, 14, 1080–1084. Sci. Lett., 2001, 20, 997–999.
7. Langlois, R., Estimation of Weibull parameters. J. Mater. Sci. Lett., 16. Davies, I. J., Best estimate of Weibull modulus obtained using linear
1991, 10, 1049–1051. least squares analysis: An improved empirical correction factor. J.
8. Song, L. Y., Wu, D. F. and Li, Y. D., Optimal probability estimators Mater. Sci., 2004, 39, 1441–1444.
for determining Weibull parameters. J. Mater. Sci. Lett., 2003, 22, 17. Peterlik, H., The validity of Weibull estimators. J. Mater. Sci., 1995,
1651–1653. 30, 1972–1976.
9. Bergman, B., On the estimation of the Weibull modulus. J. Mater. 18. Li, Y. D., Wu, D. F., Zhang, J. P., Chang, L., Wu, D. H., Fang, Z. P.
Sci. Lett., 1984, 3, 689–692. et al., Measurement and statistics of single pellet mechanical strength
10. Wu, D. F., Li, Y. D., Zhang, J. P., Chang, L., Wu, D. H., Fang, Z. P. of differently shaped catalysts. Powder Technol., 2000, 113, 176–184.
et al., Effects of the number of testing specimens and the estimation 19. Hiramatsu, Y. and Oka, Y., Determination of the tensile strength of
methods on the Weibull parameters of solid catalysts. Chem. Eng. rock by a compression test of an irregular test piece. Int. J. Rock
Sci., 2001, 56, 7035–7044. Mech. Miner. Sci., 1966, 3, 89–99.