(FMEA) in Pharmaceutical Tabletting Tools - The University of Jordan

Ordinal Logistic Regression Model of Failure Mode and Effects Analysis
(FMEA) in Pharmaceutical Tabletting Tools
Mohammad D. AL-Tahat ⇑ , Abdul Kareem M. Abdul Jawwad, Yousef L. Abu Nahleh
University of Jordan, Faculty of Engineering and Technology, Industrial Engineering Department, Amman 11942, Jordan
article info
Article history:
Received 1 December 2011
Accepted 16 August 2012
Available online 17 September 2012
Keywords:
Pharmaceutical Tabletting Tools
FMEA
Ordinal Logistic Regression (OLR)
1. Introduction
abstract
The main objective of this paper is to use Ordinal Logistic Regression Modeling (OLRM) to
predict and to investigate the relationship(s) between the different types of failures
encountered in tableting tools of pharmaceutical industry and relevant tablet- and punch
attributes. This would help minimize the occurrence of such failures in and avoid potential
failure occurrences in future punch designs. Three punch attributes (punch diameter, location
and shape) and five product attributes (tablet mass (gm), hardness (Kp), thickness
(mm), moisture content (percent loss on drying (LOD %)) and sieve size (mm)) have been
investigated in terms of their relative contributions towards different failure types. The
present OLRM model has been successfully applied to the predict failure types according
to the aforementioned factors. Furthermore, OLRM quantitatively links and evaluates the
effects and contribution of each of these factors to the occurrence of different failure types.
The OLRM methodology has been validated conveniently and proved to be powerful prediction
tool. This is indicated by the marginal 2.4% error percentage encountered.
Ó 2012 Elsevier Ltd. All rights reserved.
In pharmaceutical industries, tablets are usually prepared by the instantaneous compression of powders between a twopunch
set in a cylindrical die Fig. 1. The compression force may be supplied by either the lower or upper punch. Tablets are
becoming complex and exotic in both shape and profile for brand identity marketing. As the tablets become more complex so
does the tooling. Therefore, tooling strength and durability is a major consideration when designing a tablet.
The complex design and function of tableting allows a limited number of tool materials; particularly steel, to be used for
this purpose among the many steels available. Compaction of granules into a solid form (tablet), with sufficient hardness and
bond strength, results in high forces on the tooling, not just static loading but cyclic loading inflicting both stresses and
strains. Pharmaceutical tabletting tools facing various types of tool breakage/failure; these include punch-tip breakage/chipping,
tip wear, punch distortion and shorter punch-lengths among others.
In this research, it is wanted to investigate the types of failures encountered in pharmaceutical tabletting punch sets and
what root causes are involved in these failures using Failure Mode and Effects Analysis (FMEA). This is expected to help in
choosing the right tool material, tablet profiles as well as appropriate operating and maintenance procedures to reduce and/
or eliminate punch failures.
Pharmaceutical tablets are solid, flat or biconvex dishes prepared by compressing a drugs or a mixture of drugs, with or
without diluents. They vary in shape and differ greatly in size and weight, depending on the amount of medicinal substances
and the intended mode of administration. It is the most popular dosage form accounting for 70% of the total produced med-
⇑ Corresponding author.
E-mail address: altahat@ju.edu.jo (M.D. AL-Tahat).
1350-6307/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.engfailanal.2012.08.017
Engineering Failure Analysis 27 (2013) 322–332
Contents lists available at SciVerse ScienceDirect
Engineering Failure Analysis
journal homepage: www.elsevier.com/locate/engfailanal

icaments. All medicaments are available in the tablet form except where it is difficult to formulate [1]. A tablet should have;
elegant product identity while free of defects like chips, cracks, discoloration, and contamination, sufficient strength to withstand
mechanical shock during its production packaging, shipping and dispensing, chemical and physical stability to maintain
its physical attributes over time, the ability to release medicinal agents in a predictable and reproducible manner, and
finally should have a chemical stability over time so as not to allow alteration of the medicinal agents [1].
Typical production sequence of medical tablets, Fig. 2, includes a filling stage, a compression stage and ejection stage. In
the filling stage the exact measurement is used to give accurate weight and dosage for the tablet while filling the bore of the
die with granules.
Compressing of granule is essential to provide the correct hardness and solidarity of the tablets. If a tablet is not compressed
properly it can crumble. Normally a pre-compression is used prior to the main compression stage, the purpose of
which is to remove excess air before the powder is fully compressed and the correct final pressure is reached. After compression
an ejection mechanism lifts both lower and upper punches and ejects tablets into a collection tray and the entire cycle is
repeated for each tablet.
2. Types and geometries of tabletting tools
As pointed to above, two punches are normally employed; an upper punch component to put the pressure on the powder
and to control penetration depth into the die and a lower punch which components for die overfill and achieve final tablet
weight, Fig. 3, [3]. The most important dimension of the tooling is the working length within a set of punches. Working
length, Fig. 3, is normally subject to very tight dimensional control (0.002–0.005 in.) [3]. Variations in this specification will
result in weight, thickness and hardness variation.
Different punch shapes depend on types of tablets as shown in Fig. 4. Punch shapes can be classified into: Round, Capsuleshape,
oval, geometric (such as square, rectangular and triangular) and irregular shape punch which are used produce tablets
with different shapes like fruits, animals, vehicles etc.
A proper tablet-shape selection is a key requirement for extended punch and die life. This, of course, has a great economical
implications on the whole pharmaceutical sector.
Dies are basically hollow cylindrical parts, as shown in Fig. 5, within which the compression of powder between upper
and lower punches takes place. Good tablet design is essential to prevent downstream production problems, consequently
producing high quality tablets with maximum efficiency of the tableting process.
3. Failure Mode and Effects Analysis (FMEA)
M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332 323
Fig. 1. A schematic illustration of a tableting compression station [2].
Fig. 2. A schematic illustration of typical production sequence of medical tablets; filling, compression and then ejection stage [2].
FMEA is a procedure that is performed to classify each potential failure effect according to its severity and probability of
occurrence. It is a systematic, proactive method for evaluating a process to identify where and how a part might fail and to
assess the relative impact of different failures, in order to identify the parts of the process that are most in need of change.
Several research efforts had been undertaken for the analysis of defects in manufacturing tools by many approaches but
mainly using FMEA [4].
In an attempt to increase die-set life, Alpagut et al. [5] investigated the possible use of Zirconia as a material for the manufacture
of punches and dies for use in tableting machines and its effect on ejection of tablets. An experimental study to

324 M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332
Fig. 3. Upper punch and lower punch components [3].
Fig. 4. Round shape punch die sets and Capsule shape [2].
Fig. 5. Die geometry and shape (I Holland) [2].
investigate the movement of powder during the axially symmetrical uni-axial compaction of microcrystalline cellulose has
been carried out by Brunel et al. [6], they concluded that changing the geometry of the punch and die set would inherently

change the direction of the resultant forces causing major changes in the direction of powder movement and thus, response
to wear and other failure mechanisms.
Dubar et al. [7] proposed a new approach to the degradation of cold forging tooling, where a tabletting die is analogous to
forging tools with powder flow replaced metal flow in their modeling. The approach is developed depending on samples
coming from actual work pieces, and contactors coming from actual tools at various stages of their lifetime. A study made
by Kanga et al. [8] on a die wear model considering thermal softening involved the construction of the wear model in order to
predict tool life by wear based on the general Archard’s wear model. In Archard’s model, the hardness of a die is usually considered
to be a function of both temperature and the operating time of the die. A study on die stress analysis of the orbital
forging process to obtain die stress distributions and predict die failures had been made by Sheu and Yu [9], the different
installation positions of punches, performs, and profiles of punch tip and tool configurations were proposed to imply the type
of die failure conditions. The results of Sheu and Yu [9] showed that installation errors and punch tip with long mandrel and
split die configurations would result in significant die stresses with the corresponding die failures.
Despite the widespread use of such tooling worldwide, and their versatile consequent failures, this subject has been dealt
with mainly in industrial pathways and has grasped very limited attention within the academic and research communities.
This has created a vital need for a systematic research program in order to tackle the different types of failures encountered
in these tools. Given the immense global production volume of tabletting tools, it has been envisaged that a quantitative
treatment of failure evaluation would be more beneficial to this subject than a ‘‘simply’’ basic qualitative failure mode analysis.
This has shaped the main theme of the present paper.
In this paper FMEA will be applied to investigate at the main types of failures in the punch set and what are the root
causes for these failures. OLR modeling has been used to quantitatively link and evaluate the effect and contribution of
the aforementioned factors to the occurrence of the different failure types.
4. Qualitative analysis of collected data
M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332 325
The data used in the current research has been provided by a major Jordanian pharmaceutical producer covering production
period 2007–2009 and accounting for a total of 381 punches. Three major types of failures have been found common in
these punches (i) concentricity failures (punch distortion leading to off center positioning), (ii) edge failures (either a cracked
or chipped edge) and (iii) length variation (shortening of punch length). These failures are denoted C, E and L, respectively,
Table 1. In addition to the main ‘‘single’’ failure modes, various punches have been found to suffer from a combination of two
or three of the above failure types, Table 1, leading to a total of six failure modes. For modeling purposes the different six
failure modes are assigned ‘‘codes’’ between 0 and 5, Table 1. A cause-and-effect analysis of the tabletting process, conducted
during this research, has led to three fishbone diagrams, Figs. 6–8, mapping the main failure modes (C, E and L, respectively)
to their probable causes and/or contributing factors.
Frequencies of occurrence for each type of failure according to punch shape and position are shown in Table 2. It could be
seen that edge failures and length variation are more frequent in the lower punches than in the upper punches. About 70% of
concentricity failures are seen to have taken place in the upper punches. Concave punches have the highest frequency of failures.
Similarly failure types were also related to punch shape and diameter as shown in Table 3 and Table 4 respectively. The
8.4 mm diameter punch is seen to have the highest frequency (80%) of these failures.
Generally the most frequent defect of tableting tools is edge breakage. Edge breakage may happen as a result of two
mechanisms during the compression stage of the tabletting process; (i) an excessive pressure in order to compensate for
a shorter punch length; or (ii) a sudden impact force due to an eccentric punch (presence of type-C failure). Type-C failure,
itself, could have been, also, caused by an excessive pressure leading to punch distortion. It could be concluded at this point
that edge breakage (type-E failure) is usually preceded by one or both of the other two failure types (L and C). Type-C failure
would, most probably, be preceded by type-L failure, which seems to be the first in failure sequence.
Edge breakage problem is more frequent in the lower punches than in the upper punches. This could be directly linked to
the higher frequency of length variation in the lower punch. The lower punch is more likely to experience more wear during
the tableting process. This comes from the fact that there are less low punches than upper punches as well as the fact that a
major portion of the pressure is experienced by the lower punch. For example in the concave 8.4 mm size, there are 13 upper
punches versus 3 lower punches only. Accordingly, the lower punches are likely to experience more concentricity failures. It
is noted, however, that 70% of this failure take place in the upper punch. This is due to the fact that the upper punch moves
Table 1
Classification of types of found failures.
Seq. Failure Code
1 Concentricity (C) 1
2 Edge (E) 2
3 Length variation (L) 3
4 Combined failure (L) + (E) + (C) 5
5 Combined failure (L) + (C) 4
6 Combined failure (C) + (E) 0

326 M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332
Fig. 6. Fishbone diagram for concentricity (C) failure.
Fig. 7. Fishbone diagram for edge (E) failure.
up outside the die and down inside the die, hence would not have the support from the die like the lower punch which stays
inside the die during the whole production cycle. This is expected to increase the chance of concentricity failures in the upper
punch. The concave shape has the highest frequency of failures that are present (72%). This is due to the fact that this shape
has the highest production level and the smallest punch sizes leading to more pressure concentration.

Table 2
Distribution of failures according to punch shape and position.
Among the combined failures it could be seen that the present intersections (instances of mutual presence) are
{(L) + (E) + (C)}, {(L) + (C)}, and {(C) + (E)} but no intersection between {(L) + (E)}. This implies a complete absence of the combined
failures {(L) + (E)}. Based on this, edge breakage failure seem to be a final occurrence and there would be two routes
leading this failure; either after both length variation and concentricity through excessive force (as explained above) or after
concentricity due to sudden impact loading. This also indicates that the length variation problems are not always detected by
the inspection workers.
5. Failure prediction using Logistic Regression Modeling
M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332 327
Fig. 8. Fishbone diagram for length variation (L) failure.
Punch Frequencies of occurrence of failure type Total
Type Shape Code Concentricity (C) Length variation (L) Edge (E) (C) + (E) (L) + (C) (C) + (E) + (L)
Upper punches Oval 10 36 2 38
Oblong 9 7 8 15
Flat 8 2 2
Concave 7 54 1 39 36 130
Sub-total 63 37 47 36 2 0
Lower punches Oval 6 9 7 16
Oblong 5 11 11
Flat 4 2 12 14
Diamond 3 5 5
Concave 2 11 40 72 25 148
Caps 1 2 2
Sub-total 26 54 79 0 12 25
Total 89 91 126 36 14 25 381
Logistic Regression Analysis (LRA) extends the techniques of Multiple Regression Analysis (MRA) to research situations in
which the outcome variable is categorical rather than a variable with a numerical value [10]; the fundamental model underlying
MRA posits that a continuous outcome variable is a linear combination of a set of predictors and error. Thus, for an
outcome variable, Y, and a set of C predictor variables, X1,...,Xp, the MRA model is of the form [11]:

328 M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332
Table 3
Distribution of failure according to tool diameter.
Punch diameter (mm) Frequencies of occurrence of failure type Total
Concentricity (C) Length variation (L) Edge (E) (C) + (E) (L) + (C) (C) + (E) + (L)
11.11 6 36 25 67
11.3 ⁄ 8.25 5 5
12.7 4 12 16
18.5 ⁄ 10.25 7 2 9
5 ⁄ 10 2 1 3
5.5 ⁄ 8 44 44
6 ⁄ 15 2 2
6.35 40 36 76
7 ⁄ 18.25 16 8 24
8.4 11 1 75 87
8 8 8
10.3 40 40
Y ¼ a þ b1X1 þ b2X2 þ þbpXp þ e ¼ a þ X
bjXj þ e ð1Þ
j¼1
where b j is a multiple (partial) regression coefficient and e is the error of prediction. If the error is omitted, the resulting model
represents the expected, or predicted, value of Y [11]:
eðYjX1; ...; XpÞ ¼Y 0 ¼ a þ Xp
j¼1
b jXj
Each observed score, Y, is made up of a predictable, component, Y 0 , that is a function of the predictor variables X1,...,Xp,
and an error, or unpredictable component, e, that represents error of measurement and/or error in the selection of the model
[12].
The model for LRA assumes that the outcome variable, Y, is categorical. For simplicity, and because it is the most case
commonly encountered in practice, it is assumed that Y is dichotomous, taking on values of 1 and 0. Information in Fig. 9
represents Minitab output for Ordinal Logistic Regression Model (OLRM) employed in this study.
6. Results and discussion of OLRM
Table 4
Samples punch specifications.
Specification Failure code
1 2 3
Diameter (mm) 15.5 12.75 7.5
Location 2 1 2
Mass (mg) 900 710 123
Hardness (Kp) 18 24 3.5
Thickness (mm) 6.4 5.5 3.45
Moisture (LOD %) 2.7 3 1
Sieve (mm) 1 1 1
OLRM described in Fig. 9 shows the estimates of the b coefficients, their standard deviations, z-values and p-values. The
odds ratio and 95% confidence intervals for each of the coefficients are also shown. All variables with p value less than 0.05
are significant in the model. Log Likelihood (LL) explains the probability of the observed results given the b coefficients. Since
the likelihood is always less than 1 it is customary to use the logarithm of the likelihood multiplied by 2. This value provides
a measure of how well the model fits the data. The LL statistic is analogous to the error sum of squares in multiple
linear regressions. As such it is an indicator of how much unexplained information remains after fitting the model. A large
value means a poorly fitting model and vice versa. Therefore, a reliable model would have a small LL value (and hence, a
small 2LL value). If a model fits perfectly, the LL value would be is 1 leading to a 2LL value of zero ( 2 log 1 = 0).
One advantage of multiplying the log likelihood by 2 is that 2LL has an approximately chi-square distribution making it
possible to compare values against those that arise by chance alone. The present observed LL value is 30.661, Fig. 9, indicating
a satisfactory modeling and fitting data. Significant p value at the bottom of the regression table indicates that there is
a significant association between at least one explanatory variable and the outcome by testing whether all slopes are equal to
zero. If the p value were not significant there would be no need to go further.
ð2Þ

M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332 329
Fig. 9. Minitab output of the proposed OLRM.
The p values for each term in the model indicate whether or not there is statistically significant association between a
particular explanatory variable and the outcome. Goodness-of-fit tests are usually general tests that assess the fitted model’s
overall departure from the observed data [13], Two classical global goodness of fit tests for the OLRM have been adapted; the
Pearson test and the residual deviance test [14], which compare the fit of the model with the saturated model.

330 M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332
Table 5
Results of applying the OLRM on three samples.
Sample number Constant number
a1 a2 a3 a4 a5 1 42.2346 14.0315 2.93722 34.33702 46.46002
2 30.3187 2.11562 14.85309 46.25288 58.37589
3 72.6083 44.4052 27.4365 3.963309 16.08631
These two are available for most of the OLRM(s) for categorical responses. These measures can be directly obtained from
most of the established statistical packages. However under the null hypothesis of a well fitted model, the distributions of
the goodness of fit statistics are approximately chi-square distributed only if most expected counts formed by the cross classification
of the response levels and all covariates are greater than 5 [15]. The goodness of fit hypothesis is;
H 0: The model fits well to the data.
H1: The model does not fit well to the data.
As shown in Fig. 9, p-values of the two tests are greater than 0.05 indicating that the model satisfies the goodness of fit at
5% significant level, the null hypothesis will not be rejected. It should be noted here that, contrary to the case of normal linear
regression, a p value larger than the confidence level is needed for model significance.
Fig. 9 displays also the numbers and percent of concordant, discordant and tied pairs, as well as the common correlation
tests. Large values for Somer’s D, Goodman–Kruskal–Gamma and Kendall’s Tau – indicate that the model has a good predictive
ability. In our case there were 99.2% pairs concordant and 0.7% discordant giving 99% better chance for pairs to be concordant
than discordant, indicating the model to have excellent predictive ability.
7. Model validation
While there are six different responses (failure modes), the OLRM in Fig. 9 shows that there are five constants only. This
comes from the fact that for G outcome categories, there is only one parameter (b) for each of the predictor variables (e.g., b1
for predictor X 1). However, there is still a separate intercept term (a g) for each of the G–1 comparisons making the number of
constants equal to the number of response-1 [16]. The shaded box in Fig. 9 shows Logistic Regression parameters. At this
stage the present model is ready to be tested and evaluated using punch specifications given in Table 4 and the observed
failure types. Comparison between prediction results and actual failures would indicate the predictive power of the model.
Fig. 10. Rotating-head type punch [2].

OLRM validation has been implemented as follows:
One of the five constants (g values) is chosen for further calculations with its corresponding a, i.e., when g =1a1 will be
used, when g =2a 2 will be used and so on. After the constant has been computed, the model will be applied as follow:
Y ¼ a þ 18:8983ðDiameterÞ 15:042ðLocationÞ 0:3381ðMassÞþ1:78037ðHardnessÞþ32:7696ðThicknessÞ
þ 11:4021ðMoistureÞ 317:81ðSieveÞ ð3Þ
Table 5 shows the results of applying this model for the three sample specifications
The Barrel of the punch did not experience any type of failure, the failure occurs on the tip and the head of the punch so
the design can be more flexible by using ‘‘insert-type’’ tips and heads, in which case these will be changed when ever failure
occurs. The number of upper punches is more than lower punches resulting in larger production volumes per punch for lower
sets. A future design may then look into more balance life cycles of upper and lower punches. As the first type of failure to
occur is length variation, and as this is mainly due to wear of punch head, it would be beneficial to use a more wear-resistant
material for this part of the punch.
8. Concluding remarks
In this research study Failure Mode and Effects Analysis (FMEA) has been applied to relate main types of failure that occur
in pharmaceutical punch sets to their root causes both qualitatively and quantitatively (through OLR modeling). Types of
failures experienced in tableting punches have been revealed to be affected by; punch diameter, punch location, punch
shape, product mass, product hardness, product thickness, moisture content and product sieve size. Edge-breakage problems
were found to account for 33% of failures; followed by length variation (23.9%) and then concentricity (23.4%). The most logical
failure sequence which seems to fit the data presented in this study seems to start with length variation followed by
concentricity and finally edge breakage/chipping. An OLR model has been obtained and was successfully applied for the prediction
of potential failure type(s) based on punch and product attributes.
The OLRM is validated conveniently; it proved to be powerful prediction tool with minor percentage error value of 2.4%.
9. Recommendations
Given the failure sequence postulated in this study, i.e., failure cycle stars with a shortening in the punch length, and the
fact that expected failure commencement sites are only at the tip (through breakage/chipping) and punch head (through
wear) with the rest of the punch body (barrel) being unaffected, a viable means of avoiding whole-punch replacement is
to utilize replaceable (insert-type) punch-head and tip in future designs. A rotating punch-head would also be a viable option
to minimize the amount of wear encountered, Fig. 10. In future work, chemical composition of the tablets may be considered
as an input variable to investigate its effect on the different types of failures occurring in pharmaceutical punch sets.
Acknowledgment
This work is extracted from a master’s thesis entitled ‘‘Failure Mode and Effects Analysis (FMEA) in The Pharmaceutical
Tableting Tools’’ which has been successfully defended, approved, and published by The University of Jordan in July 2010. It
was supervised by Dr. Mohammad D. Al-Tahat and Dr. Abdul Kareem M. Abdul Jawwad, and was prepared by graduate student
Yousef Lutfi A. Abu Nahleh, to whom this work primarily belongs.
References
M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332 331
[1] Sahoo PK. Pharmaceutical technology tablets. Delhi Institute of Pharmaceutical Sciences and Research, Pusph Vihar-III, M.B. Road, New Delhi 110 017;
2007.
[2] I Holland Troubleshooting in Tablet Tooling Seminar, I Holland Limited Meadow Lane, Long Eaton, Nottingham. NG10 2GD, England; 2010.
[3] Techceuticals. A tablet making training resource for tablet making professionals, 365 Red Cedar Street, Suite 202 Bluffton, South Carolina, USA 29910–
4519; 2011. .
[4] William Winchell. Continuous quality improvement: a manufacturing professional’s guide. society of manufacturing engineers, Product ID: BK91PUB3
ISBN: 978-087263405; 1991.
[5] Alpagut K, Michael JT, Ron S. An application for zirconia as a pharmaceutical die set. J Eur Ceram Soc 2003;24(10–11):3091–101.
[6] Eiliazadeh Brunel, Pitt Kendal, Briscoe Brian. Effects of punch geometry on powder movement during pharmaceutical tableting processes. Int J Solids
Struct 2004;41(21):5967–77.
[7] Dubar M, Dubar L, Dubois A. Wear analysis of tools in cold forging: PVD versus CVD TiN coatings. Wear 2005;259(1–2):1109–16.
[8] Kanga JH et al. A study on a die wear model considering thermal softening: (I) construction of the wear model. J Mater Process Technol
1998;96(1999):53–8.
[9] Sheu JJ, Yu CH. The die failure prediction and prevention of the orbital forging process. J Mater Process Technol 2008;20(1):9–13.
[10] Peng Chao-Ying Joanne, Lee Kuk-Lida, Gary M. An introduction to logistic regression analysis and reporting. J Educ Res 2002;96(1):3–14.
[11] Gude Justin A, Mitchell Michael S, Ausband David E, Sime Carolyn A, Bangs Edward E. Internal validation of predictive logistic regression models for
decision-making in wildlife management. Wildlife Biol 2009;15:352–69.
[12] Xian-Jin X, Jane P, William C. Increasing the power: a practical approach to goodness-of-fit test for logistic regression models with continuous
predictors. Comput Stat Data Anal 2008;52:2703–13.
[13] Archer Kellie J, Lemeshow Stanley. Goodness-of-fit test for a logistic regression model fitted using survey sample data. Stata J 2006;6(1):97–105.

332 M.D. AL-Tahat et al. / Engineering Failure Analysis 27 (2013) 322–332
[14] Kuss Oliver. Global Global Goodness-of-Fit Tests in Logistic Regression with Sparse Data. Institute of Medical Epidemiology: Biostatistics and
Informatics, Medical Faculty University of Halle-Wittenberg; 2002.
[15] Abeysekera W, Sooriyarachchi R. A novel method for testing goodness of fit of a proportional odds model: an application to an AIDS study. J Nat Sci
Found Sri Lanka 2008;36(2):125–35.
[16] David GK, Mitchel K. Logistic regression: a self-learning text. 2nd ed. Springer; 2002.