does component sharing help or hurt reliability? an ...faculty.london.edu/kramdas/ramdas and...
TRANSCRIPT
Does Component Sharing Help or Hurt Reliability? An Empirical Study in the AutomotiveIndustryAuthor(s): Kamalini Ramdas and Taylor RandallSource: Management Science, Vol. 54, No. 5 (May, 2008), pp. 922-938Published by: INFORMSStable URL: http://www.jstor.org/stable/20122441 .
Accessed: 01/10/2014 08:16
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .http://www.jstor.org/page/info/about/policies/terms.jsp
.JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range ofcontent in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new formsof scholarship. For more information about JSTOR, please contact [email protected].
.
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.
http://www.jstor.org
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
MANAGEMENT SCIENCE hCK Vol. 54, No. 5, May 2008, pp. 922-938
DOI i0.i287/mnsc.l070.0791 issN 0025-19091 EissN 1526-55011081540510922 ? 2008 INFORMS
Does Component Sharing Help or Hurt Reliability? An Empirical Study in the Automotive Industry
Kamalini Ramdas Darden Graduate School of Business, University of Virginia, Charlottesville, Virginia 22906,
Taylor Randall David Eccles School of Business, University of Utah, Salt Lake City, Utah 84112,
Component sharing?the use of a component on multiple products within a firm's product line?is widely
practiced as a means of offering high variety at low cost. Although many researchers have examined trade
offs involved in component sharing, little research has focused on the impact of component sharing on quality. In this paper, we examine how component sharing impacts one dimension of quality?reliability?defined as
mean time to failure. Design considerations suggest that a component designed uniquely for a product will result in higher reliability due to the better fit of the component within the architecture of the product. On the other hand, the learning
curve literature suggests that greater experience with a component can improve
conformance quality, and can increase reliability via learning from end-user feedback. The engineering literature
suggests that improved conformance in turn increases reliability. Sharing a component across
multiple products increases experience, and hence, should increase reliability Using data from the automotive industry,
we find
support for the hypothesis that higher component reliability is associated with higher cumulative experience with a component. Further, we find support for the hypothesis that higher component reliability is associated with a component that has been designed uniquely for a product. This finding suggests that the popular design strategy of component sharing
can in some cases compromise product quality, via reduced reliability
Key words : empirical study; benefits of specific design; quality; component sharing History: Accepted by Christoph Loch, R&D and product development; received June 11, 2004. This paper was
with the authors 2 years and 4 months for 2 revisions. Published online in Articles in Advance March 27, 2008.
1. Introduction In this paper, we examine the impact of component
sharing?the use of a component across multiple end
products within a firm's product line?on component
reliability. A component's reliability is defined as its mean time to failure.
Consider the following design scenario relating a
firm's component-sharing strategy to the reliability of
the components used in its products. When designing an assembled product, designers repeatedly evaluate
whether to create a unique component specifically for
the product or to reuse an existing component. Spe cific design allows the designer greater flexibility in
tailoring the component specifications to the needs of the product, which should lead to higher reliability.
For any component, however, unanticipated defects
may arise in its manufacture, assembly, or use. Over
time, the occurrence of such defects is reduced via
improvements in manufacturing and assembly pro cesses for the component, and engineering improve
ments to the component itself based on feedback from
downstream firm functions and from end users. The benefit of using an existing component is that many
reliability problems may already have been identi fied and corrected via this improvement process. The downside to using an existing component is that
design fit can be compromised, leading to decreased
reliability. The tension described in this scenario leads us to the central questions of this paper. Does greater
reliability result when a component is used in a prod uct for which it was specifically designed? Is higher component reliability associated with increased expe rience with a particular component, in manufacturing, assembly, and field use? Finally, if both the specific design effect and the experience effect exist, which
might have the greatest relative impact, and in what situations?
We believe that these are important questions both for practitioners engaging in component shar
ing and for researchers studying this topic. Auto
industry executives we interviewed claimed that the economic incentive to share components increases
when after-sales warranty costs are factored into the
design decision. Our work examines whether, and in what situations, component sharing has the poten tial to reduce failures, and hence warranty costs. If
922
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry Management Science 54(5), pp. 922-938, ?2008 INFORMS 923
component sharing compromises product reliability, this detrimental effect on quality could negate the
widely touted benefits of component sharing, which
include reduction in product design, manufactur
ing and distribution costs, and increased responsive ness to consumer demand, as documented by Ulrich
(1995), Swaminathan and Tayur (1998), Ramdas and
Sawhney (2001), Ramdas (2003), Rutenberg (1971), Fisher et al. (1999), Gupta and Krishnan (1999), Krishnan and Gupta (2001), Desai et al. (2001), Kim
and Chhajed (2001), Ramdas et al. (2003), Thonemann
and Brandeau (2000), and Yano and Dobson (1998).
Conversely, if component sharing enhances product
reliability in addition to its more widely known ben
efits, this could lead companies to encourage more
component sharing.
Although a few researchers have prescriptively modeled how component sharing can be effectively used to satisfy market segments with differing qual
ity needs (Desai et al. 2001, Heese and Swaminathan
2006, Ramdas et al. 2003), the reliability trade-off we
described above remains unaddressed. Using empiri cal data in the domain of automotive braking systems,
we examine how component sharing impacts reliabil
ity. Examining the impact on reliability of tailoring a
component's design specifically to a product applica tion is one major contribution of our work. Another
key contribution is that we shed light on how differ
ent ways to increase the cumulative experience with a
component?experience on a single car model versus
experience via sharing the component across multi
ple models?improve reliability. Finally, we are able
to highlight design contexts where either the unique
design or experience effects dominate.
In ?2, we describe the industry context for our study. In ?3, we develop our hypotheses. In ?4, we discuss
the data and variable definitions. In ?5, we present
methodology, and in ?6, we present results. In ?7, we
discuss our findings and their implications for prac tice. Section 8 contains limitations and conclusions.
2. Industry Context The context for our study is the automotive industry.
We focus specifically on one component of the auto
motive braking system?the brake rotor?and study brake rotor sharing strategies at Ford Motor Com
pany. We chose to focus on braking systems because
brake reliability is of critical importance to the con
sumer, brake-related quality issues are a critical deter
minant of warranty costs (Automotive News, June
1999), and there is considerable sharing of braking
system components at Ford and other auto makers.
Within braking systems, we chose to focus on brake
rotors because, based on our discussions with indus
try experts, reliability problems associated with brake
rotors are most easily identified as occurring due to
rotor design decisions, as opposed to decisions about some other components in the braking system or the
automobile.
An automotive braking system is a hydraulic sys tem that converts human foot pressure applied at the
pedal into a much higher braking pressure applied at the wheels via the braking system components. The pressure applied at the wheels forces stationary brake components?brake pads that are attached to
the calipers?to rub against rotating components?the brake rotor, thus converting the kinetic energy of a
moving car into heat energy via friction. The pressure
applied to the rotor by the brake pads causes wear
on the brake rotor. Common rotor problems that arise
from use include warping, scoring, or even cracking of the rotor.
Automotive braking system design is initiated only after vehicle design has been broadly specified, via
"system-level parameters" such as vehicle weight, top
speed, and stopping distance. Given these inputs, the
components of the braking system must be designed so as to provide adequate torque to stop a car
from top speed within the desired stopping distance. In addition, all braking components are designed for "maximum loading" conditions: for example, the
brake pedal should not crack if the driver steps excep
tionally hard on it in a panic stop. Further, sev
eral constraints arise due to the interaction between
braking components: for example, the hydraulic ratio
(ratio of areas of master cylinder and caliper pis tons) must lie within prespecified limits to eliminate
excessive pedal "travel," which could cause the brake
pedal to hit the floor of the car. Braking system design parameters like rotor radius, desired pedal force, and area of the caliper pistons and master cylinder piston are manipulated to meet these different ends. In the context of our study, the challenge for the designer is to balance the constraints of each unique braking sys tem with the potential benefits of component sharing.
3. Theory and Development of
Hypotheses Our hypotheses rely on prior research in design the
ory and learning curve theory, as well as insights
gained from interviewing senior executives, man
agers, and engineers in the auto industry. The
first hypothesis addresses the question of whether
uniquely designed components produce better relia
bility outcomes. The second hypothesis addresses the
question of whether and how greater experience with a component impacts reliability.
Design theory suggests that designing unique com
ponents for each specific product application will
result in higher product quality. Ulrich (1995) dis cussed the role of components within modular and
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 924 Management Science 54(5), pp. 922-938, ?2008 INFORMS
integral product architectures. In a pure modular
architecture, the interfaces among components are
standardized, and so multiple products can be config ured by mixing and matching from a base set of com
ponents. In a pure integral product architecture, the
complex interactions among components require that
components be specifically designed for each prod uct (Ulrich and Ellison 1999). In practice, products are
often a blend of modular and integral architectures.
The elements of the architecture that are integral lead
designers to uniquely design components for each
end product rather than share components across end
products. An implication of this theory is that sharing
components inappropriately will result in poor fit of
the component with the product, and hence will hurt
quality. In probing exactly how quality may be affected
when design fit is compromised, it is useful to think
of the impact of design fit on multiple dimensions
of quality. We posit that a compromise in design fit
can negatively impact both performance and reliabil
ity. As a simple example, suppose that while design
ing a scissors intended for cutting cloth, designers at
a stationery company borrow blades from an exist
ing scissors that was designed to cut paper. (In actu
ality, scissors manufacturers typically advocate that
paper scissors not be used for cutting cloth, and vice
versa.) One possible outcome is that the scissors per forms poorly from the start, because the blade is just not what is needed for cutting cloth. Another possible outcome is that the scissors performance is acceptable at first, but deteriorates over time, so that at some
point the performance is longer acceptable. This effect
is called "perf or manee degradation/' and it is more
likely to occur if the fit of the design to the particular
application in which it is used is compromised. Such
performance degradation in turn impacts reliability. To our knowledge, this important effect of compo
nent reuse has not been documented in the literature.
One possible reason for this oversight is that in the
traditional engineering literature, reliability has had
a binary definition: a product has either failed, or it
has not. Only recently have researchers proposed a
continuous-state reliability model, in which gradual
degradation in performance occurs over time (Yang and Kapur 1997). Degradation can occur even in cases
where there is a distinct switch over to the "failure"
state?for example, a light bulb fails because its ele
ment is gradually evaporating over time. It is reason
able to expect that lack of design fit of a component to
the specific application in which it is used will accel
erate such performance degradation over time, thus
reducing reliability. This basic notion is complicated by industrial prac
tice. Auto manufacturers often choose to design a new
rotor for use on multiple models at once, rather than
on a single new model. We argue that doing this
dilutes the specificity of the design to any particular model in the use set. Therefore, for any vehicle, if the
rotor used on it was designed for the specific model
and model year of that vehicle, and furthermore if
that was the only vehicle for which it was designed, we expect the fit and therefore the reliability to be
higher than if that rotor was designed for use on sev
eral different models in that particular model year. In our empirical model, we expect that the positive
impact on reliability associated with using a rotor that
was designed for the model and model year of a vehi
cle will be moderated by the total number of models
on which that rotor was used in its first year.1 These
arguments lead to our first hypothesis.
Hypothesis 1. The greater the specificity of a compo nent's design to a product application, the greater the reli
ability of the component in that application. However, the
larger the set of products a component was designed for, the
poorer the specificity of the design to any one product in
that set, and hence the poorer the component's reliability.
In the automotive braking context, we expect that
the reliability of a brake rotor in a vehicle will be
higher if the rotor was designed for the specific model
and model year of the vehicle than if it had been orig
inally designed for another model or even another
model year of the same model. This effect will be neg
atively moderated by the number of models for which
the brake rotor was designed. The literature on learning curves provides empiri
cal evidence that product costs decrease in the cumu
lative production volume of a product.2 More recently, researchers have examined the impact of learning on
different aspects of quality. Reagans et al. (2005) con
sider the impact of learning on surgical outcomes.
In an airline industry study, Lapr? and Tsikriktsis
(2006) link learning to customer dissatisfaction, an
outcome measure of quality evaluated by consumers.
Fine (1986) and Ittner (1996) link learning curve
theory to improvements in conformance quality. Fine (1986) models conformance quality improve
ment over time as a function of both "learning-by
doing," as described by Anzoni and Simon (1979), and induced learning; that is, consciously engaging in
improvement activities. Lapr? and van Wassenhove
(2001) empirically identify factors that facilitate "for
mal" or induced learning. Mukherjee et al. (1998)
1 We assume that if a rotor was used on multiple models in its first
year, it was designed for use on all of these models.
2 See Hatch and Mowery (1998) for a comprehensive list of the
many studies that document the learning curve phenomena in
various industries. See, as examples, Wright (1936), airplanes; Baloff (1971), automobiles; Dick (1991), semiconductors; Lieberman
(1984), semiconductors.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry
Management Science 54(5), pp. 922-938, ?2008 INFORMS ^
925
report on empirically observing two types of learn
ing: operational learning, which comes from develop
ing specific solutions to specific problems faced in a
factory, and conceptual learning, which relies on con
ceptual models to understand why problems occur,
resulting in more general solutions.
The engineering literature suggests that an improv ement in conformance quality in turn improves prod uct reliability. Kececioglu (1991) states that quality con
trol assures conformance quality, i.e., conformance to
specifications. This reduces manufacturing variance, which can degrade reliability. Kececioglu (1991, p. 35) maintains that "no product can perform reliably with
out the inputs of quality control because quality parts and components are needed to go into the product so that its reliability is assured." Dasgupta and Pecht
(1991) indicate that a component can fail if its man
ufacturing tolerance requirements are not met. Wong (1995) graphs production quality of semiconductors
against their reliability in the field, and finds a clear
positive relationship. To understand why production experience and
conformance quality impact reliability, it is useful to
consider how product flaws affect reliability. For a
product in use, reliability problems can very often be
attributed to "built-in flaws," which are introduced
in the course of the manufacturing or assembly pro cess (Wong 1995). Wong distinguishes such built-in
flaws from "design flaws," which are introduced in
the course of the design process. Design flaws can be
very damaging to reliability, but are typically discov
ered early and corrected before the product is in use
(see Figure 1). The engineering literature suggests that reliability
improves with greater manufacturing experience, due
to a reduction in built-in flaws. Specifically, in the
context of brake rotors, an example of a built-in flaw
is overtightening of wheel lug nuts during assembly. Such overtightening can distort the rotor, inducing lateral runout, a condition where the surface of the
Figure 1 The Impact of Design Flaws and Built-in Flaws on Quality
Design flaw Built-in flaw
Where is flaw is
introduced? Design process Manufacturing
process
What dimensions
of quality does the flaw affect?
Performance
j (over time) Reliability
Conformance
I Performance
| (overtime)
Reliability
Where is the
flaw identified? Design, manufacturing,
manufacturing testing, or in field use
Manufacturing,
manufacturing
testing, or in field use
rotor wobbles from side to side as it rotates. Accord
ing to General Motors Corporation, undertorquing a
single lug nut can create as much as 0.003 inch of lat
eral runout (Carley 2002). A rotor with lateral runout
will not wear evenly, increasing the likelihood of fail ure over time. With improved manufacturing based on greater experience and learning by doing, con
formance quality, and in turn reliability, are likely to improve. Kececioglu (1991, p. 192), notes that
improvements in conformance quality reduce infant
mortality, and also improve reliability in the normal
life and end-of-life portions of a product's life cycle. As experience with a product increases, aside from
learning by doing in manufacturing, the firm also
has an opportunity to learn from end-user feedback.
For example, as products are used in the field, failed
products under warranty are returned to the firm, and undergo FMEA3 analysis, which can expose both
design flaws and built-in flaws. Modifications made to products based on this type of feedback and anal
ysis can improve component reliability. Thus, greater experience should improve reliabil
ity, be it via learning by doing in manufacturing, or
learning from end-user feedback.
A stream of research on learning transfer focuses on whether and how the learning accumulated in the course of production at one production site transfers to other production sites or to other products at the same site. Hatch and Mowery (1998) report that losses in semiconductor yield occur as processes are trans
ferred from development facilities to manufacturing facilities. Adler (1990) shows that not all productiv
ity gains associated with learning by doing are trans
ferred to new manufacturing situations.
Similarly, the empirical research that investigates
knowledge transfer across products suggests that
knowledge transfer occurs?but to an incomplete degree (Argote 1999). For example, Benkard (2000) finds in the context of aircraft production that only a
partial transfer of knowledge occurs across different models built at the same production site.
Based on these arguments, we expect that although the use of a component on a new model will increase
overall experience with the component, the shift to
the new model will likely be accompanied with a
discontinuity in the learning curve, in the form of an initial loss of quality. Multiple models that use
the same component may have slight differences that
make some of the learning from one model irrel
evant to another. Models that share a component
may be assembled on different lines within a plant, or in totally different geographical locations, further
impeding knowledge transfer. Argote (1999) provides an example of how an automotive plant using a
3 Failure modes and effects analysis.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 926 Management Science 54(5), pp. 922-938, ?2008 INFORMS
Figure 2 Annual Volume and Cumulative Volume for a Brake Rotor Used on the Ford Mustang from 1994 to 2001
200,000
180,000
160,000
O 140,000 E = 120,000 O
? 100,000
i 80,000
60,000
40,000
20,000
O
Annual volume
Cumulative total
1,200,000
1,000,000
-1-800,000 i O >
600,000 >
3
400,000 E O
200,000
1994 1995 1996 1997 1998 1999 2000 2001
Year
particular assembly system may discover a way to
improve its performance, but if sister plants do not use the same system, the improvements may not be
transferable. Of course, the efficiency of knowledge transfer is likely to vary across firms, because it is a
function of organizational structure and knowledge management capability
Focusing on brake rotors, Figures 2 and 3 show
examples in our data from Ford of how the cumu
lative experience with a specific brake rotor can be
attained in different ways?via sharing the rotor over
time on a single model, or via sharing the rotor over
multiple models, over time. In practice, different mod
els that share a component are often made at different
plants or on different assembly lines, and this trend
will increase as companies move towards implement
ing global product platforms.4 We expect that the
cumulative experience gained with a specific brake
rotor via its use on a single model will have a greater
positive impact on reliability than experience gained via its use on multiple models.
The above discussion leads us to our second
hypothesis.
Hypothesis 2. Higher component reliability will be
associated with higher cumulative production experience, but the magnitude of the relationship will decrease in the
number of products used to attain any particular level of
experience.
4. Data and Variables We obtained data from several sources. The data on reliability focuses specifically on the reliability of
automotive brake rotors. This data was obtained from
the Department of Transportation of the U.S. govern ment, and records consumer complaints about front
brake rotor quality problems. This data is used by the U.S. government to guide product recall deci
sions. Like the data on customer dissatisfaction used
by Lapr? and Tsikriktsis (2006), our data is based on
performance in the field, as opposed to internal eval
uation of quality within the firm. Our data on brakes
component sharing was obtained from an automo
tive research company. From this data, we can deter
mine all of the unique braking components in use
on vehicles sold on the U.S. market in the period of our study, and what vehicles used each unique brake. We obtained data on individual model vol umes and vehicle characteristics such as weight and
horsepower from Ward's Automotive (Ward's Automo
tive Yearbooks 1965-2003) and Automotive News, and
data on brake characteristics via direct measurement, Internet sources, and Motor Vehicle Manufacturer's
Association specification sheets. We obtained data on
control variables from public sources. For example,
precipitation data was obtained from the National
Weather Service, and population data from the U.S.
Census.
The unit of analysis in our study is an individ
ual vehicle on the road, identified by its vehicle
identification number (VIN). The VIN label, which
is often printed on the underside of the dash
board, uniquely identifies every vehicle on the road.
It can be used to track ownership, standard and
optional factory-installed equipment, and other indi
vidual vehicle-specific details. The original data set
from the government on brake failures contained 990
observations from Ford Motor Company for individ
uals reporting a first-time rotor failure. Of these, 297
observations were dropped due to incomplete data.
The primary reason for the incomplete data was lack
of precise rotor information to match the VIN. The
pattern of missing rotor information appears random,
reducing the potential for sample bias. This leaves 693
observations on which to test our hypotheses. Fifty six different brake rotors were used on the vehicles in
our data set. We describe below the variables used in our hypothesis tests. Refer to Table 1 for descriptive statistics.
4.1. Dependent Variable
4.1.1. Miles to Failure. We use miles to failure of
the brake rotor as our measure of rotor reliability, the dependent variable in our study. Miles to failure
is the number of miles driven in a particular vehi
cle from zero miles to the miles at the time of the
first reported failure of the brake rotor. A rotor fail ure is defined as a case in which the rotor requires
repair or replacement because it appears abnormal, either in the way in which it is operating or in its
4 A platform is defined most broadly as a set of resources that is
shared across products, ranging from components to production
processes.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry Management Science 54(5), pp. 922-938, ?2008 INFORMS 927
Figure 3 Annual Volume by Model and Cumulative Volume for a Brake Rotor Shared over Multiple Models and over Multiple Years
500,000 i-r 8,000,000
000,000
1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994
Year
physical condition. Our data indicates that rotor fail ure results from problems such as scoring, warping, or cracking of the rotor. The mean miles to failure is
31,673 (median = 27,000) with the minimum miles to
failure reported at 2 miles and the maximum miles to
failure reported at 134,000 miles. Figure 4 presents a
histogram of miles to failure.
Table 1 Descriptive Statistics {N = 693)
Mean Median Std. dev. Min. Max.
Continuous variables
Miles to failure 31,673 27,000 24,330 2 134,000 Cumulative volume 845,205 511,395 1,112,285 7,869 7,568,420
Time in use 1.75 1 2.55 0 15
Number of models 2.33 2.00 1.12 1.00 9.00
Horsepower 151.76 140 33.21 84 310
Swept area per ton 119.34 120 12.18 85.19 198.79
Precipitation 38.48 41 14.85 7 110
Population density 3,300.57 2,478 3,198.27 12.8 34,916
% Positive weight 7.72 6.21 6.59 0.03 30.13
difference
(/7 = 554)(%) % Negative weight 7.62 7.12 6.42 0.05 32.78
difference
(A7 = 82) (%) Suggested retail 13,587 12,440 4,328 5,893 28,987
price
Dichotomous variables (% of sample)
Year-specific design 36%
Carryover design 49%
Vented 88% Trucks 24%
4.2. Study Variables
4.2.1. Measures of Cumulative Experience. Exist
ing literature uses cumulative production volume as
a proxy for cumulative experience effects (Macher
2003). Although alternative measures such as time
and engineering resources have been explored (Hatch and Mowery 1998), cumulative volume is the stan
dard proxy. We measure the cumulative volume of a rotor used in a VIN as the total volume from its
time of first use up to the model year of that VIN.5
We calculated cumulative volumes for each brake
rotor by summing the volumes of all the models over
which the rotor had been used, up to and including the model year of the car. We obtained individual
model volumes from Ward's Automotive and Auto
motive News, and adjusted for cases in which mul
tiple rotors had been used on a model. The mean
cumulative volume for brake rotors is 845,205 rotors
(median =
511.395), with a minimum cumulative vol ume of 7,869 and a maximum of 7,568,420. Figure 5
shows the cumulative volumes over time for the 56
unique Ford brake rotors in our study. It is common
convention in the learning curve literature to take the
natural log of cumulative volume for estimation pro cedures (Hatch and Mowery 1998). We follow this
convention.
5 As a robustness check, we also measure volume using midyear
production volumes rather than end-of-year volumes. We report no significant differences in results using this alternative volume
measure.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 928 Management Science 54(5), pp. 922-938, ?2008 INFORMS
Figure 4 Histogram of Miles to Failure
25,000 50,000 75,000 100,000 125,000
Miles to failure
Aside from cumulative volume, we also use a mea
sure of learning based on the time that has passed since the first use of the component. The time in use
for each rotor-VIN combination is defined as the time
from when that rotor was first introduced until the
model year of the VIN in question. The mean time
in use is 1.75 years (median =
1). We note a rela
tively high correlation between cumulative volume
and time in use (Table 2, r ? 0.67, p < 0.01). We believe it is an empirical question as to whether
it is cumulative volume or time in use that better cap tures learning effects in the data. We show results
with both measures in our model specification.
4.2.2. Number of Models. The number of models
represents the number of different vehicle models on
which a particular brake was shared up to the time of
manufacture of the vehicle in question. This variable serves as the moderating variable in our hypothe ses. Rotors are shared on an average of 2.33 models
(median =
2) with a minimum of 1 model (no shar
ing) and a maximum of 9 models. We defined each
unique name
plate?e.g., Ford Taurus versus Mercury
Sable?as a distinct model. Although these vehicles
share a platform, there are some significant differ
ences; for example, in vehicle weight, which is an
important factor in brakes design. This is the reason
for our choice of definition. However, we obtain sim
ilar results when we classify models such as the Ford
Taurus and Mercury Sable as being the same model.
4.2.3. Measures of Product Specificity of Com
ponent's Design. We measure whether a brake was
designed specifically for a model in two different
ways. An indicator variable, year-specific design, cap tures whether or not the brake rotor used in a VIN
was designed specifically for the model and model
year of that VIN. If the rotor used on a VIN was intro
duced prior to the model year corresponding to that
VIN, then year-specific design equals zero. If instead
the rotor used on a VIN was introduced in the same
model year as that VIN, year-specific design equals one regardless of how many other models the rotor
was used on in its first year of use. In our sample, 36%
of the observations had rotors designed specifically for the model and model year of the VIN in ques tion. Another indicator, carryover design, is set equal to one if the brake used in a VIN is designed specifi
cally for the model corresponding to that VIN, but not
for its specific model year. This variable enables us
to examine the reliability benefits of sharing rotors on
the same model across time. For example, if a rotor is
Figure 5 Cumulative Volumes for the 56 Unique Ford Front Brake Rotors in Our Study
8,000,000
1983 1985 1987 1989 1991 1993 Model year
1995 1997 1999 2001
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry
Management Science 54(5), pp. 922-938, ?2008 INFORMS 929
Table 2 Correlation Matrix (N = 693)
Miles to failure
Cumulative
volume
Time in use
Number
of
models Horsepower
Swept area
per ton Precipitation Population
density
% Positive weight
difference
% Negative Suggested weight retail
difference price
Miles to failure Cumulative volume
Time in use
Number of models
Horsepower
Swept area per ton
Precipitation Population density % Positive weight
difference % Negative weight
difference
Suggested retail
price
1 -0.03
-0.13***
0.03
-0.19***
0.03
0.12***
-0.06
0.04
0.01
-0.09**
1
0.67*'
0.68*'
-0.35*'
0.08*'
0.04
0.03
0.23*'
-0.05
-0.26*'
1
-0.54*'
-0.19*'
0.07*'
0.03
0.03
0.12*=
0.05
-0.15*'
1
-0.30*'
0.02
0.09*'
0.02
0.28*'
0.10*=
-0.11*'
1
-0.24
-0.09
-0.10*
-0.30
0.08
0.56
1
0.03
0.04
0.37
-0.23
-0.34
1
-0.08*'
0.01
0.01
0.01
1
0.07*
-0.01
-0.11*
1
-0.26
-0.48
1
0.24*=
, *
coefficient significant at the p < 0.01, 0.05, and 0.10 levels, respectively.
used on a model in 1993 (the design year), 1994, and
1995, carryover design equals 0 in 1993 (the design
year) and 1 in each subsequent year. Forty-nine per cent of our observations used a carryover design.
4.3. Control Variables
4.3.1. Vented. Rotors may be vented or solid.
Vents alleviate heat generated as calipers come in con
tact with the rotor during braking. The reduced heat
keeps rotors from warping easily. We use an indicator
variable to indicate a vented rotor. We expect vented
rotors to have lower hazard rates than solid rotors, other things being equal. Eighty-eight percent of the
observations in our sample have vented rotors.
4.3.2. Horsepower. Automobiles with higher
horsepower can accelerate faster, placing higher
requirements on the brakes. We expect automobiles with higher horsepower to have a higher failure rate.
The mean horsepower was 151.76 HP (median =
140 HP).
4.3.3. Swept Area per Ton. The swept area is the area of contact between a brake caliper and brake rotor. All else being equal, the larger this swept area, the greater the braking ability. Heavier cars typically
have a larger rotor swept area. The rotor swept area
per ton controls for the relation between the weight of the vehicle and the compensating swept area, and
is a commonly used metric of braking potential. We
expect a higher swept area per ton to result in higher miles to failure. We report a mean swept area per ton
of 119.34 square inches per ton (median = 120 square
inches per ton).
4.3.4. Population Density. We use the population
density of the county of driver residence as a proxy for general driving conditions. We expect the hazard
rate to increase with population density because city
driving results in greater use of the brakes. We report a mean population density of 3,300 people per square
mile (median =
2,478). Population density exhibits considerable skewness. We mitigate this problem by taking the natural log of population density.
4.3.5. Precipitation. Interviews with brake ex
perts suggested that weather conditions associated with precipitation can reduce the longevity of brakes and brake rotors. We use the average precipitation of the county of driver residence to control for wet
weather conditions. The mean precipitation per year is 38 inches (median
= 41 inches).
4.3.6. Percent Positive Weight Difference. Inter views with industry personnel suggest that when a brake rotor is reused across multiple models, to ensure reliability the rotor selected might be overspec ified. In this case, we would expect the rotor per formance to increase when reused. Although there are many dimensions of performance specification, our interviews indicated that it is easier to share a
rotor across models if the new automobile weighs less than the weight of the car that the rotor was origi
nally designed for (design weight6). We use the pos itive percent weight difference, calculated below, as a proxy for the potential overspecification of shared
brake rotors. For each VIN,
Percent positive weight difference
(design weight ?
weight of VIN)
design weight = if design weight >
weight of VIN,
0 else.
6 If a rotor was used on multiple models in its first year of use, we
define design weight as the weight of the heaviest of these models.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 930 Management Science 54(5), pp. 922-938, ?2008 INFORMS
We report an average positive percent weight dif
ference of 7.72% (median =
6.21%). Note that 554, or
80%, of our observations exhibit a positive difference,
indicating that rotors are typically reused on models
that weigh less than their design weight.
4.3.7. Percent Negative Weight Difference. Inter
views with industry personnel also indicated that a
brake might also be reused in instances where the
weight of the automobile is higher than the original
design weight. To control for potential quality prob lems caused due to underspecification, we include a
variable that captures underspecification. We define
the percent negative weight difference as follows.
Percent negative weight difference
[ (weight of VIN ? design weight)
design weight = i if design weight < weight of VIN,
[ 0 else.
We report an average percent negative weight dif
ference of 7.62% (median =
7.12%). Note that only 82 observations have a negative weight difference,
whereas 57 observations display no weight difference.
4.3.8. Suggested Retail Price. We proxy for the
overall quality of the automobile design by using sug
gested retail price adjusted for inflation. We expect cars of higher price to have a higher quality. The
average inflation-adjusted price is $13,587 (median
$12,440).
4.3.9. Early Epoch. To control for potential
changes in brake technology over time, we divide
the brake introduction time period into epochs. We
present results with an early epoch (prior to 1985). However, results are robust to different specifica tions of the epoch variable. Eight percent of our
observations occur on brakes designed before 1985.
Finally, we use indicator variables to control for
the specific assembly plant that produced a VIN.
MacDuffie et al. (1996) reported significant differ ences in quality across automobile assembly plants.
Although we found significant coefficients for several
of the assembly plant dummies, for brevity we omit
these variables in reporting our results. We also use
an indicator variable to control for whether the vehi
cle is a passenger car or a truck. Twenty-four percent
of the observations are trucks.
Table 2 presents Pearson correlation coefficients
among the study variables. We highlight several of
the significant correlations. We note that our depen dent variable, miles to failure, is negatively asso
ciated with horsepower and suggested retail price, but positively associated with precipitation. It is not
significantly correlated with our variables of interest.
Second, we notice that cumulative volume is highly correlated with the number of models.
5. Methods Because we are dealing with survival data, a haz
ard rate model should be used rather than standard
regression analysis (Helsen and Schmittlein 1993). We
specify the hazard function used to test our hypothe ses as follows:
hi(t/X) =
A0(f).exp{a1(experience/) + a2(number of models,)
+ a3(experience^ (number of models/)
+ a4(specific design^)
+ a5(specific designz)(number of models,)
4- ? (control variables, )},
where h^t/X) represents the hazard of failure for the
zth observation (VIN in our study) at mileage t, given a set of covariates X, A0(?) represents a baseline haz
ard that is a function of miles driven but does not
vary by individual VIN, the as are the coefficients of
the study variables, "control variables" represents a
vector of control variables, and ? represents a vector
of coefficients of the control variables. The variable
"experience" refers to either ln(cumulative volume) or time in use, discussed earlier. The variable "spe cific design" refers to either year-specific design or
carryover design, discussed earlier.
An important issue we considered was what type of hazard model to estimate. There are a wide array of
models and the choice of models depends to a great extent on the purpose of the study (Allison 1995). In
our study, we are interested in understanding what
drives the differences in failure rates across different
models at any mileage, rather than how the hazard
rate varies with mileage. In other words, we are not
interested in estimating A0(f). One hazard model, the
Cox proportional hazards model, requires no strict
assumption about A0(f) (Cox 1972). The Cox model
assumes that the ratio of hazards for any two indi
viduals is constant over time. Using this assump tion, a method of partial likelihood estimation may be
used. During estimation the baseline hazard, A0(?), is
eliminated. Practically speaking, the model eliminates
potential biases introduced by assuming a specific functional form for the distribution of the baseline
hazard. In our application, it is reasonable to assume
that all brakes share a similar pattern of hazard A0(?) with regard to miles driven, and that relative differ
ences in reliability at any mileage are a function of
differences in the study variables and control vari
ables included in our analysis. For this reason, we
believe the Cox model to be most appropriate for our
situation. Further, Allison (1995) states that the Cox
model is commonly used for hazard rate analysis and
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry
Management Science 54(5), pp. 922-938, ?2008 INFORMS 931
has become a standard in hazard regressions. Regres sion coefficients in the Cox model have a relatively
straightforward interpretation. The coefficient is inter
preted as a risk ratio. By subtracting 1 from the risk
ratio and then multiplying by 100, one can interpret the coefficient as the percentage change in the haz
ard ratio for a one unit change in the independent variable.
One further methodological issue arises due to our
specific sample and self-reported nature of the data.
Because we have data only on those individuals who
experienced a rotor failure and who also chose to
report this failure, the hazard distribution for our data
could be different from that in the general popula tion. However, under the reasonable assumption that
the probability of reporting a failure is not a function
of any of the variables of interest in our study, it can
be shown that the coefficients of our study variables
have the same interpretation as they would if we had
in fact analyzed data representing both reported and
unreported failures (see the appendix).
6. Results Table 3 presents results of estimates obtained using the Cox proportional hazard model. Note that higher
reliability is associated with a lower hazard (nega tive coefficient) and lower reliability is associated with a higher hazard (positive coefficient). The chi-square statistics for all models in the table are significant at
the 0.01 level. The r-square values calculated based
on Allison (1995) range from 20.69% to 23.00%. The
first column presents results of a model with con
trol variables only As expected, higher proportional hazard rates are associated with higher horsepower,
higher population density, and brakes designed in the
early epoch. We see no significant association between
vented rotors, positive or negative percent weight dif
ference, the truck dummy variable, and the hazard
rate. Unexpectedly we see a significant negative asso
ciation between our precipitation variable and the
hazard rate. This result is not a function of influen
tial points in the data set. Further research yielded an alternative explanation for this result. When rain
ing, drivers often slow average speed. This results
in lower braking demands leading to lower failure
rates rather than the higher hypothesized rates of fail ure. We also see an unexpected positive association
between suggested retail price and the hazard rate.
These results suggest that cars with higher price are
of lower brake rotor reliability. The price variable may be associated with the propensity to report: owners
of higher-priced cars may report at lower mileages than owners of lower-priced cars (see the appendix).
Unfortunately, we cannot distinguish between these
two alternative explanations. We also notice a posi tive association between swept area per ton and the
Table 3 Cox Proportional Hazard Regression Estimating the Hazard of Brake Rotor Failure
I II III IV
Horsepower 0.005*** 0.004** 0.004** 0.004**
Swept area per ton 0.01** 0.01 0.01 0.01 Vented -0.13 -0.06 -0.12 -0.04
% Positive weight 0.32 0.56 0.92 0.76 difference
% Negative weight 0.77 0.68 1.36 1.12 difference
Suggested retail price 0.0001*** 0.0001*** 0.0001*** 0.0001***
Early epoch 0.74*** 0.81*** 0.88*** 0.82***
Truck -0.05 -0.12 -0.26 -0.14
Precipitation -0.006** -0.006** -0.007** -0.006**
ln(population density) 0.10** 0.10** 0.10** 0.10**
ln(cumulative volume) -0.23** ?
-0.23**
Time in use 0.01
Number of models -1.60** -0.24*** -1.62*
ln(cumulative volume) 0.10** ?
0.10*
xnumber of models
Time in use 0.01
xnumber of models
Year-specific design -1.18*** -0.65** -1.17*
Carryover design ? ?
-0.004
Year-specific design 0.53*** 0.31** 0.55**
xnumber of models
Carryover design ?
0.04
xnumber of models
Wald chi-square 160.20*** 179.72*** 175.18*** 181.13***
statistic
fl-squared (%) 20.69 22.84 22.34 23.00 N 693 693 693 693
***, **, *
coefficient significant at the p < 0.01, 0.05, and 0.10 levels,
respectively.
hazard rate. However, this variable is not significant in subsequent models.
Columns II and III present models containing our
two alternative measures of experience, ln(cumulative
volume) and time in use. In column II, we report results of a model using ln(cumulative volume) as
a measure of learning. We report a negative and
significant coefficient for year-specific design and a
positive and significant coefficient for the interaction
between year-specific design and number of models.
This finding is consistent with Hypothesis 1, which
states that the reliability of a brake will be higher on cars that are among the set of models the brake
was designed for, but that this effect will decrease
as the number of models the brake was designed for increases. We also report a negative and signifi cant coefficient for ln(cumulative volume) and a posi tive significant coefficient for the interaction between
ln(cumulative volume) and the number of models.
This finding is consistent with Hypothesis 2, which
states that reliability is increasing in cumulative expe rience, but that this effect is mitigated by the num
ber of models over which the experience is gained. In column III, we report results of a model in which
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 932 Management Science 54(5), pp. 922-938, ?2008 INFORMS
time in use is the measure of experience. Consistent
with column II and Hypothesis 1, we report a sig nificant negative coefficient for year-specific design and a positive significant coefficient for the interac
tion between year-specific design and the number of
models. However, we find no significant coefficients
for our second measure of experience, time in use, or
for the interaction between time in use and the num
ber of models.7 This suggests that cumulative volume
better captures the effect of experience in our data.
Schoonhoven (1981) suggests analyzing the deriva
tive of the estimated equation for more careful exam
ination of interaction effects when the main effect and interaction effect have opposite signs. For ease
of exposition, we ignore the subscript / that refers to
specific VINs, in our regression model. Our regression model using ln(cumulative volume) is
h(t/X) =
A0(f). expja^ ln(cumulative volume)
+ a2(numDer of models)
+ a3 In (cumulative volume)
(number of models) + a4 (specific design)
+ a5 (specific design) (number of models)
+ /3(control variables)}.
Taking the derivative of the hazard rate h with respect to volume, and setting this to zero, we have
dh(t/X) ?(cumulative volume)
= h(t/X)(_^_ i a3(number of models)\ =Q
\ cumulative volume cumulative volume /
Solving for the number of models at which dh(t/X)/d cumulative volume =
0, number of models = ? ax/a3.
For our data set, with ax = ?0.23 and a3
= 0.10,
the proportional hazard rate decreases with cumula
tive volume when the number of models over which
the volume is spread is less than 2.30, and starts to
increase with cumulative volume when the number
of models is greater than 2.30. In our data, roughly 50% of all observations have brake rotors that are
shared across more than two models. By similar logic, we analyze the interaction term between year-specific
design and number of models. Here, the benefits of
designing for a specific set of models are eliminated
when the number of models equals 2.22. In our data,
roughly 50% of all observations have brake rotors that
were designed for more than two models.
Column IV reports results of tests including carry over design as an additional measure of design speci
ficity. Consistent with the results of column II, we
report significant effects for year-specific design and its interaction with models. However, the coefficient on carryover design and its interaction with num
ber of models is not significant. These results sug
gest that the benefits of design specificity are found
only in the year of brake rotor introduction and not in subsequent years. A model (not reported) combin
ing year-specific design and carryover design into a
single variable yielded insignificant results. We checked outlier diagnostics and found that the
results in Table 3 are robust to the influence of sin
gle data points. Further, we noted no problems with
multicollinearity, except for the interaction terms, for
which the variance inflation factor is quite high.8 Our data also exhibit uneven distribution of observations over brake rotors. We tested our results for robust ness to this distribution by controlling for brake clus ters as suggested by Allison (1995) and note that the
results are consistent with those reported in Table 3.
We reestimated our model using different subsets of the data to test the robustness of our results. These
model runs are reported in Table 4. We only report results for models using the year-specific design and
the cumulative volume variables.
Our first concern deals with a potential endoge nous association between quality and the number of
models on which a brake rotor is used. We have no
direct hypothesis about the number of models and
quality. However, if designers are given feedback on
the quality of a brake, a designer may choose to reuse a brake because of its observed quality as opposed to the number of models a new brake is designed for affecting quality by altering design specificity. Our
industry interviews yielded mixed opinions as to the existence and strength of this feedback loop. Also,
reliability data only becomes available a few years after a brake's introduction. Nevertheless, to alleviate
this potential concern, we eliminate instances where
quality feedback may have led to a reuse decision
by eliminating all observations where a brake was
not designed for the specific model or models.9 Col umn I of Table 4 reports the results of this analysis.
We observe results consistent with those in Table 3.
A second concern raised by discussions with industry
7 We also tried a specification including both cumulative volume
and time in use to capture experience effects. Our results were qual
itatively similar in this case.
8 We observe that in models that include the main effects of cumu
lative volume and number of models, but exclude the interaction
term, the signs of the main effects do not change from what is
reported in columns II and III.
9 By placing this constraint on the data, we not only reduce the
opportunity for endogenous effects, but also reduce the variance in
the number of models. Importantly, this reduction in the number
of models may occur in instances where the detrimental effects of
sharing may be most likely.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry
Management Science 54(5), pp. 922-938, ?2008 INFORMS 933
Table 4 Robustness Tests of Cox Proportional Hazard Models
I First time
usage
No early failures
No late failures
IV
North American
volume
V
Passenger cars
VI
Trucks
VII Eliminate rotors
with caliper failure
Horsepower 0.002 0.004**
Swept area per ton 0.007 0.006
Precipitation -0.003 -0.006*
ln(population density) 0.11** 0.11*** Vented 0.28 0.02 % Positive weight diff. 3.01 ** 0.64 % Negative weight diff. 1.41 0.52
Suggested retail price 0.0001*** 0.0001*
Early epoch -0.31 0.94*** Truck -0.20 -0.07
ln(cumulative volume) -0.25* -0.26***
Number of models -2.45** -1.73***
ln(cumulative volume) 0.16** 0.11** x number of models
Year-specific design -1.29*** -1.21***
Year-specific design 0.61*** 0.54***
x number of models
Chi-square statistic 181.66*** 183.77***
ff-squared (%) 26.54 24.14 N 589 665
0.005**
0.007
-0.007**
0.09**
-0.06
0.28
-0.00001
0.0001*?
0.80***
-0.09
-0.19**
-1.26***
0.08*
-1.07***
0.48**
166.65***
21.68
682
0.004*
0.007
-0.006**
0.09**
-0.07
0.76
0.55
0.0001*
0.61*
-0.57*
-0.41***
-2.99***
0.20***
-1.49***
0.69***
176.52***
23.34
664
0.002
0.02***
-0.009**
0.06
-0.88*
1.87
1.64
0.0001*
-0.07
-0.39**
-3.36***
0.22**
-1.67***
0.81***
158.59***
26.07
525
0.003
-0.002
-0.004
0.21**
-0.03
0.70
1.53
0.0001 1.19***
-1.05***
-7.23**
0.53**
-2.23***
1.20**
43.03
22.60
168
0.005*^
0.007
-0.006**
0.09**
-0.16
0.18
1.28
0.0001*
0.79***
-0.22
-0.23**
-1.48**
0.09*
-1.24***
0.54***
178.62***
23.49
667
, **, *
coefficient significant at the p < 0.01, 0.05, and 0.10 levels, respectively.
experts deals with a potential "lemon" effect, mean
ing that a certain number of parts fail at the out
set of an automobile's introduction. These failures
may not have any association with the factors in this
study. To examine the robustness of our results to
"lemon-" related failures, we eliminated observations
with failures that had occurred at less than 500,1,000, and 2,000 miles. The results of the 1,000-mile cutoff are qualitatively similar to the 500- and 2,000-mile results and are shown in column II of Table 4. Third,
we observe in Figure 1 that failures are reported on
brakes up to 134,000 miles. There exists the possi
bility that failures on high-mileage observations are
not failures on original equipment, but on replace ment brakes. We reran our models after truncating
the failures at different levels down to 100,000 miles.
The results of the regressions after eliminating all fail
ures with mileage over 100,000 miles is shown in col
umn III of Table 4. Fourth, we noted a limitation of the
data, which is that cumulative volumes are calculated
for North American auto sales only. Thus, volumes for
brakes shared globally will be understated, creating a
potential bias in the association between volume and
the hazard rate. We were unable to acquire precise volume estimates for brakes shared globally. How
ever, through interviews with company representa tives and automotive industry experts we were able to
identify several models where this effect was likely to
be most severe, and we reran the analysis eliminating these models. These results are shown in column IV
of Table 4. Fifth, our data contain both passenger cars
and trucks. It is possible that the magnitude of these
factors differs greatly across the types of vehicles. Col
umn V of Table 4 shows results for passenger cars
and column VI shows results for trucks. Sixth, there
exists the possibility that the failure of the rotor was
due to failure of other parts of the braking system. To mitigate this concern, we obtained detailed notes
on each reported failure. Using these notes, we iden
tified failures for which the rotor failure was men
tioned in conjunction with failure of another part of
the braking system. We estimate models after elimi
nating these observations in column VII of Table 4.
We note in Table 4 that the results of regressions esti
mated using subsets of our data consistently show
support for Hypotheses 1 and 2.
7. Discussion We find support for the two questions we address
in our research. Rotor reliability is higher when the
model using a rotor is among the set of models that
the rotor was designed for, than if this is not the case.
However, this specific-design effect is dissipated as
the number of models that the rotor was designed for increases. Further, greater experience with a brake
rotor is associated with higher reliability. However, the positive impact of cumulative experience on the
reliability of a brake rotor is moderated by the num
ber of models over which this experience is attained.
The inherent tension in these findings merits further
analysis and interpretation. To facilitate this process, we compare the ratios of relative hazards between dif
ferent design scenarios. The use of ratio comparison
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 934 Management Science 54(5), pp. 922-938, ?2008 INFORMS
with a Cox proportional hazards model isolates the
effects of variables of interest.
7.1. Scenario 1: Reuse of an Existing Rotor on a
Continuing Model
Consider a scenario where an automobile model (e.g.,
the Ford Escort) is being redesigned. To maximize
reliability, should the rotor be carried over from the
previous year's Escort model or should a new rotor
be designed for the new model version? Using the estimates in Table 3, column II and a general hazard
function, we calculate the ratio of hazard between a
newly designed brake rotor and a brake rotor that was used in the previous model year. Let V denote
the expected volume on the new brake, CV the cumu
lative volume of the reused brake rotor excluding the
volume from the new model, and N the number of models on which the existing rotor is in use (e.g., the
existing model might be used on the Escort and Lynx). After simplification the resulting ratio is as follows:10
hazard of newly designed rotor
hazard of reusing a rotor with a given cumulative volume
?-1.18+0.53-l-0.23-ln(V)-1.6-l+0.10-ln(V>l _
?-0.23-ln(V+CV)-1.6-N+0.10-N-ln(V+CV) '
Intuitively, we weigh the reliability gains from
product-specific design associated with a new rotor
against the reliability gains from larger production volume, and hence greater experience with the reused rotor. The case for new design will dominate the case
for reusing the existing rotor when the above ratio
is less than one. Figure 6 illustrates how this ratio
changes with V and CV. For purposes of discussion, the calculations in Figure 6 assume that N is equal to
two, which is the median number of models sharing a brake rotor in our data set. The x-axis shows rep
resentative values from our data set for anticipated
first-year volume V of a new brake rotor that is used on a single model. The curves are for the minimum, first quartile, median, third quartile, and maximum
values of cumulative volume for a rotor that is shared across two models in our data set. The figure illus trates that as the anticipated volume of the new brake rotor increases, the effect of specific design tends to
dominate the learning effects from cumulative vol ume of the reused rotor. This effect occurs because as
the volume V of the new rotor increases, the learning effects obtained from the new rotor, combined with
the benefit for specific design, more than compen sate for the learning effects lost when choosing not to
Figure 6 Ratio of Hazard of New Design to Hazard of Reused Design as a Function of Volume (V) of New Rotor, and Cumulative
Volume (CV) of Reused Rotor
c
S 1 2_
x (7,869 Min V, 1.21) I _+_ Min cv = 57k ?
'
\ ^~~ 1stQtlCV = 327K
S j\ ~-*~- Median CV = 577 K
S 4\ ~^~ 3rd Qtl CV = 961 K
Anticipated volume (V) of the new brake rotor
reuse the existing brake rotor. Across the entire range of cumulative volume (CV) for an existing rotor in our data set, if the new rotor volume is close to the
minimum value seen in our data set, reuse results in
a lower expected hazard. If the new rotor volume is
above the first quartile of volumes from our data set, the new design results in a lower expected hazard.
7.2. Scenario 2: Design for Multiple Models
Consider a scenario where a designer needs to design brake rotors for a series of models in some model
year?for example, while designing a new platform. The designer can capture the benefits of design speci
ficity by creating a unique brake rotor for each model or can capture the potential learning effects associated
with greater experience by using a universal design that serves all models. In this scenario, we assume
that the anticipated volume of each individual model
equals V. The resulting ratio is as follows:
hazard of newly designed rotor
hazard of universal design
?-1.18+0.53(l)-0.23-ln(V)-1.6(l)+0.104n(V>l ""
e-1.18+053(N)-0.23-]n(N-V-lMN)+0.10-(N)-]n(N-V)
'
Figure 7 shows values of this ratio for different val ues of V and N. Note that a set of uniquely designed rotors is more and more likely to dominate the uni
versal design as the anticipated volume per model (V) increases. With respect to the number of models, we
see an interesting effect. For volume per model, V
set at its minimum value in our data set, the ratio
of hazards of unique design to universal design first
increases and then decreases with the number of mod
els N. In this scenario, universal design dominates
10 For simplicity, we assume that the value of %-positive-weight
difference for the reused rotor is 7.72%, its mean value in our data
set. Also, we assume that the swept area per ton is the same for
the new and reused rotors. These assumptions do not change the
direction of our results.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry Management Science 54(5), pp. 922-938, ?2008 INFORMS 935
Figure 7 Ratio of Hazard of Model-Specific Designs to Hazard of Universal Design as a Function of Volume per Model (V) and Number of Models Sharing the Rotor (A/)
O 50,000 100,000 150,000 200,000 250,000 300,000
Anticipated volume per model (V)
if there are two to eight models, but for nine or
more models, unique design dominates. This relation
is driven by the interaction between volume and the
number of models. Beyond a crossover volume per model V of 12,500 units, the hazard ratio is monoton
ically decreasing in number of models N. The inter
action among the variables is not straightforward, and the model enables a manager to avoid simplis tic heuristics such as "at low V, universal design is
increasingly better for reliability as N increases/'
7.3. Scenario 3: Reuse on a Continuing Model
with Uncertain Demand Volume
Consider a situation similar to scenario 1, but where
the demand for the continuing model is uncertain.
Suppose mean demand volume for the continuing model is 100,000 units, based on an equal chance of
two demand outcomes: 100,000 + D and 100,000 - D.
Volume variability is higher for larger D. In this setup we consider three scenarios, for low, medium, and
high volume variability. With high volume variability, for a swing in volume D that could be either posi tive or negative with equal probability, the reduction
in reliability of the new rotor relative to the existing rotor is a lot worse if a downswing in volume materi
alizes, than the increase in reliability of the new rotor
relative to the existing rotor if an upswing in volume
of the same magnitude materializes. Figure 8 shows
expected reliability ratios for the new to the existing rotor for different levels of variability in demand, as
well as the reliability ratio based on expected vol ume. If we look only at expected volume, then regard less of variability, the new rotor appears to provide
higher reliability than the existing rotor, for all val ues of cumulative volume of the existing rotor. How
ever, if we factor in variability in volume of the new
rotor, we find that it is better to go with reuse if the
volume variability is high enough and the cumula
tive volume of the existing option is high enough.
Figure 8
c ?) 75 1-10 <D "O Ui 1.05
1.00
3 0.95
0.90
O > 0.85
"? 0.80
B ?-75^
'?= 0.70
Ratio of Reliability of New to Existing Design as a Function of Variability in New Design Volume and Cumulative Volume of Existing Design
(3,576 K, 1.04)
s^" (576 K, 0.99)
K (56 K, 0.94)
~~o~~ Ratio using expected volume ?o? Expected ratio, low volume variability -a?Expected ratio, medium volume variability ?x-~ Expected ratio, high volume variability
500 1,000 1,500 2,000 2,500 3,000 3,500 4,000
Cumulative volume of existing design in thousands
Although intuitively it makes sense to go with a con
servative option (reuse) in the face of unpredictable demand, our model helps understand how unpre dictable demand has to be for it to make sense to go
with the conservative approach. Although we have
examined a simple demand distribution for illustra
tion, our model can be used in a similar way to navi
gate more complex distributions.
7.4. Scenario 4: Investment in Increasing
Design Modularity We expect the positive impact of unique design on
reliability to be greatest in the case of highly inte
gral architectures, where design fit is most important. A more modular architecture would facilitate compo nent sharing by lessening the importance of design fit. Suppose a firm decides to invest design effort
in modularizing its product's design. One benefit
from doing this would be a reduction in the pos itive impact of unique design on reliability, i.e., a
reduction in the magnitude of the coefficient of year
specific design. Suppose the magnitude of this coef
ficient is reduced by 50% (i.e., moving from ?1.18
to ?1.18*0.5) via moving
to a more modular archi
tecture. Figure 9 shows how the ratio of hazards for
designing uniquely versus reusing a rotor varies over
volume of the new design, for a more integral ver
sus a more modular architecture. For simplicity, the
cumulative volume of the reused rotor is set at its
median volume in our data set. We find that with
the more modular setup, reuse dominates designing anew over the entire range of volumes for the new
rotor, whereas in the current (more integral) setup, reuse dominates designing anew only if the new
rotor's volume is at or below the first quartile for a
new model's first-year volume in our data set. Thus,
our model helps us see the reliability impact of reuse
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 936 Management Science 54(5), pp. 922-938, ?2008 INFORMS
Figure 9 Ratio of Reliability of New to Existing Design as a Function of Integral vs. Modular Design
c:
d) .r' "O
8 2.0
o 0)
SB10
o (0
CO
* (Min V, 2.07) \
Vm (IstQtIV, 1.73)
Median CV = 577 K, integral Median CV = 577 K, modular
-??S^L^ (Max V, 1.33)
-1-1-1 0 100,000 200,000 300,000
Anticipated volume (V) of the new brake rotor
versus designing uniquely, for a given level of modu
larity. Together with this information, the design costs
saved via reuse can be weighed against the design costs associated with making the design more modu
lar, to decide how modular the design needs to be.
7.5. Scenario 5: Investment in Improving
Knowledge Transfer
The extent to which learning is transferred across fac
tories varies with organizational and structural factors
(Argote 1999). Similarly, such factors should impact
learning across models, within one factory or across
factories. How can a firm evaluate the benefits from
investing resources to increase cross-model learning? Our model provides one approach. If learning trans
fer improves, the coefficient of N ln(y) in our model
should be reduced in magnitude. Similar to scenario 4
above, we will find that reuse over multiple models
becomes more feasible with better learning transfer,
and, importantly, the model helps estimate how many models can share a
component.
The scenario analyses above illustrate different de
cision-making contexts for component sharing, and
identify the conditions where the unique design effects dominate and the conditions where the learn
ing effects dominate, within our data. The scenarios
were chosen because they are common decisions for a
designer, not because they represent a comprehensive list of design scenarios. Other scenarios are certainly
possible, and our approach provides a way to evalu
ate these scenarios.
8. Limitations and Conclusions The data are limited to one automotive company and are the product of the design processes at that
company. Generalization of the results will depend, in part, on the similarities among design processes across
companies. Further, our analysis focuses on a
single component, the brake rotor. Despite this, we
believe our methods and results may be generalized
in the following ways. First, brake rotors are mechani
cally similar to many other components in an automo
bile and components in other products. Second, the
methods and measures used to estimate the impact of component sharing on failure rate of brake rotors
are quite general and are applicable to components in
other assembled products. Increased experience enhances both learning by
doing in manufacturing as well as learning from
end-user feedback in the field, and both of these
mechanisms can result in better reliability. Although we are able to capture the overall effect of experience on reliability, we are unable to separate the effects
of these two mechanisms. We also recognize the fact
that we have not been able to collect all driver-specific control variables. Age and gender are examples of
variables for which hypotheses related to brake failure
might be formed. To the extent these variables are cor
related with our test variables, our study suffers from
omitted variable bias. Another variable that we did
not observe is design effort, which could be higher for
brakes intended for high-volume cars. We also did not
observe whether engineers planned ahead for future
vehicles in which a rotor would be used, which might result in purposeful sharing.
This study provides the first empirical evi
dence relating one dimension of quality?component
reliability?to component-sharing strategies. We find
support for the hypothesis that higher reliability is
associated with a component designed specifically for a model in a given year, but that this effect is
reduced when the component is designed for simul
taneous use on other models. A component that is
designed specifically for a model is likely to have the
best design fit in terms of meeting the specific design
requirements and constraints of that model, result
ing in higher reliability This fit aspect can be diluted
if a component is designed simultaneously for mul
tiple models, because different models are likely to
have different design requirements and constraints.
This finding suggests that the popular design strat
egy of developing multiple products off a common
platform with shared components can in some cases
compromise product quality We also observe that
improved reliability is associated with greater experi ence, but this effect decreases as the number of mod
els over which this experience is attained increases.
This dilution effect can occur due to incomplete trans
fer of knowledge gained via experience with one
model to other models that use the same compo nent, for example, because these different models
are made at different manufacturing locations, thus
offering less opportunity for knowledge exchange.
Comparison of the specific design effect and the expe rience effect suggests that the dominance of one effect
over another will depend on the sharing situation
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry
Management Science 54(5), pp. 922-938, ?2008 INFORMS 937
being considered. These findings suggest that design ers and researchers should look beyond the inventory and cost savings associated with component-sharing strategies and also include the quality ramifications
of component sharing while formulating component
sharing strategies. Future research should incorpo rate these quality trade-offs in prescriptive models for
component sharing. Although the trade-offs we have
focused on are quite general, future empirical research
may identify nuances to these trade-offs in specific industrial settings.
Acknowledgments The authors are
grateful to executives, managers and engi neers at Ford and General Motors, several automotive
experts and race car enthusiasts, and the Department of
Transportation for providing them with information; and
to Noah Springer, Ryan Sundquist, Prasanna Velamuri,
and Baocheng Yang for research assistance. The authors
are grateful to Mary Margaret Frank, Richard Holubkov,
Nils Rudi, Christian Terwiesch, and seminar participants at
UCLA, Wharton, Harvard Business School, University of
Chicago, INFORMS, and the Wharton Technology Minicon ference for their valuable comments and, in particular, to
Steven Stern for his help with sampling related issues. This research was
supported in part by a grant from the Batten
Institute for Innovation and Entrepreneurship at the Uni
versity of Virginia.
Appendix Let X denote the vector of covariates?study variables
and control variables?with which the hazard of failure at
mileage t is assumed to vary. At any mileage t, we can
state the hazard of failure given X generally as h(t/X) =
A0(f)exp{?/3X}, where ? is a vector of coefficients. Let p(report/t, Xj) denote the probability that the owner
of a vehicle would file a report of a rotor failure if a failure were to occur at mileage t, and given
a vector of covari
ates Xl7 where Xl may partially overlap with X. The prob
ability of reporting if a failure were to occur is likely to
vary with the mileage at the time of failure, and also per
haps with some of the covariates in X. For example, if city drivers are more
likely to complain than country drivers,
then the population density of the area where the vehicle
was driven may impact the probability of reporting. Let
r(t/X, Xj) denote the hazard function for those individuals who both experienced
a rotor failure and chose to report it.
The hazard function that we can estimate from our data is
in fact r(f/X,X1). By conditioning on the occurrence of a
failure at time t, r{i/X,Xx) =
h(t/X)p(report/t,X1). Taking the derivative of the logarithm of r(t/X, X-?) with respect to
X, we have
^logr(f/X,X1) _ dlogh(t/X) dlogp(report/t,X1)dX1 dx
" sx +
?x? lx' Because an individual car driver likely does not know
the cumulative volume of the brake rotor used in his or
her car, the number of models on which that rotor was
used, or whether that rotor was designed specifically for
the car, it is reasonable to assume that none of the variables
of interest in our study
are in the vector X1. Therefore, the
second term in the right-hand side of the above equation does not impact
our study variables. In interpreting the Cox
regression coefficients for r(t/X, X^, the components of ? for our
study variables have the same interpretation
as they would in a Cox regression that estimates the unobserved
hazard h(t/X). Note that thinking in terms of the probability of report
ing if a failure were to occur at time t allows us to model
scenarios where a failure has not actually occurred by time
t. It also allows us to model scenarios where the vehicle was
taken off the road prior to time t due to some non-brake
related issue, precluding the possibility of a brake failure at time t. In this situation, if the reason the vehicle was taken
off the road has nothing to do with why the brakes might fail, then the components of ? for our
study variables would
continue to remain unbiased.
Brake rotors typically fail?i.e., require repair or
replace
ment?within the first few years of operation. Thus, except
for car models introduced in the last few years of our study,
most cars on the road during the period of our study would
have experienced their first rotor failure. Therefore, by con
ditioning on failure as above, we are
accounting for a large
part of the risk set.
References
Adler, P. 1990. Shared learning. Management Sei 36(8) 938-958.
Allison, P. D. 1995. Survival Analysis Using SAS: A Practical Guide.
SAS Institute, Cary, NC.
Anzoni, Y, H. A. Simon. 1979. The theory of learning by doing.
Psych. Rev. 86 124-140.
Argote, L. 1999. Organizational Teaming: Creating, Retaining and
transferring Knowledge. Kluwer Academic Publishers, Norwell, MA.
Baloff, N. 1971. The extension of the learning curve. Oper. Res.
Quart. 22 329-340.
Benkard, C. L. 2000. Learning and forgetting: The dynamics of air
craft production. Amer. Econom. Rev. 90(4) 1034-1054.
Carley, L. 2002. Big three brake symposium. Tech update. Brake and
Front End (November).
Cox, D. R. 1972. Regression models and life tables. /. Roy. Statist.
Soc.f Ser B. 34(2) 187-220.
Dasgupta, A., M. Pecht. 1991. Material failure mechanisms and
damage models. IEEE Trans. Reliability 40(5) 531-536.
Desai, P., S. Kekre, S. Radhakrishnan, K. Srinivasan. 2001. Product
differentiation and commonality in design: Balancing revenue
and cost drivers. Management Sei. 47(1) 37-51.
Dick, A. R. 1991. Learning by doing and dumping in the semicon
ductor industry. /. Eaw Econom. 34 662-669.
Fine, C. 1986. Quality improvement and learning in productive sys tems. Management Sei. 32(10) 1301-1315.
Fisher, M. L., K. Ramdas, K. T. Ulrich. 1999. Component sharing in the management of product variety: A study of automotive
braking systems. Management Sei. 45(5) 297-315.
Gupta S., V. Krishnan. 1999. A product family-based approach for
integrated component and supplier selection. Production Oper.
Management 8(2) 163-182.
Hatch, N. W., D. C. Mowery. 1998. Process innovation and learn
ing by doing in semiconductor manufacturing. Management Sei.
44(11) 1461-1477.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions
Ramdas and Randall: An Empirical Study in the Automotive Industry 938 Management Science 54(5), pp. 922-938, ?2008 INFORMS
Heese, H. S., J. M. Swaminathan. 2006. Product line design with
component commonality and cost-reduction effort. Manufactur
ing Service Oper. Management 8(2) 206-219.
Helsen, K., D. C. Schmittlein. 1993. Analyzing duration times in
marketing: Evidence for the effectiveness of hazard rate mod
els. Marketing Sei. 12(4) 395-414.
Ittner, C. 1996. Exploratory evidence on the behavior of quality costs. Oper. Res. 44(1) 114-130.
Kececioglu, D. 1991. Reliability Engineering Handbook, Vol. 1. Prentice
Hall, Englewood Cliffs, NJ.
Kim, K., D. Chhajed. 2001. An experimental investigation of valua
tion change due to commonality in vertical product line exten
sion. /. Product Innovation Management 18(4) 219-230.
Krishnan, V., S. Gupta. 2001. Appropriateness and impact of
platform-based product development. Management Sei. 47(1) 52-68.
Lapr?, M. A., N. Tsikriktsis. 2006. Organizational learning curves
for customer dissatisfaction: Heterogeneity across airlines.
Management Sei. 52(3) 352-366.
Lapr?, M. A., L. N. Van Wassenhove. 2001. Creating and trans
ferring knowledge for productivity improvement in factories.
Management Sei. 47(10) 1311-1325.
Lieberman, M. 1984. The learning curve and pricing in the chemical
processing industries. RAND J. Econom. 15 213-228.
MacDuffie, J. P., K. Sethuraman, M. Fisher. 1996. Product variety and manufacturing performance: Evidence from the interna
tional automotive assembly plant study. Management Set 42(3) 350-369.
Macher, J. T. 2003. Vertical disintegration and process innova
tion in semiconductor manufacturing: Foundries vs. inte
grated producers. Working paper, Georgetown University,
Washington, D.C.
Mukherjee, A. S., M. A. Lapr?, L. N. Van Wassenhove. 1998. Knowl
edge driven quality improvement. Management Sei. 44(11) S35-S59.
Ramdas, K. 2003. Managing product variety: An integrative review
and research directions. Production Oper. Management 12(1) 79-101.
Ramdas, K., M. Sawhney. 2001. A cross-functional approach to
designing multiple line extensions for assembled products.
Management Sei. 47(1) 22-36.
Ramdas, K., M. L. Fisher, K. T. Ulrich. 2003. Managing variety for assembled products: Modeling component systems sharing.
Manufacturing Service Oper. Management 5(2) 142-156.
Reagans, R., L. Argote, D. Brooks. 2005. Individual experience and
experience working together: Predicting learning rates from
knowing who knows what and knowing how to work together.
Management Sei. 51(6) 869-881.
Rutenberg, D. P. 1971. Design commonality to reduce multi-item
inventory: Optimal depth of a product line. Oper. Res. 19(2) 491-509.
Schoonhoven, C. B. 1981. Problems with contingency theory: Test
ing assumptions hidden within the language of contingency
theory. Admin. Sei. Quart. 26(3) 349-377.
Swaminathan, J., S. Tayur. 1998. Managing broader product lines
through delayed differentiation using vanilla boxes. Manage ment Set. 44(12) S161-S172.
Thonemann, U. W., M. L. Brandeau. 2000. Optimal commonality in
component design. Oper. Res. 48(1) 1-19.
Ulrich, K. T. 1995. The role of product architecture in the manufac
turing firm. Res. Policy 24(3) 419^40.
Ulrich, K. T., D. J. Ellison. 1999. Holistic customer require ments and the design-select decision. Management Sei. 45(5) 641-658.
Wong, K. L. 1995. A new framework for part failure-rate prediction models. IEEE Trans. Reliability 44(1) 139-146.
Wright, T. 1936. Factors affecting the cost of airplanes. /. Aeronautical
Sei. 3(4) 122-128.
Yang, K., K. C. Kapur. 1997. Customer driven reliability: Integration of QFD and robust design. IEEE Annual Reliability Maintainabil
ity Sympos. Proc., Philadelphia, 339-344.
Yano, C. A., G. Dobson. 1998. Profit optimizing product line
design, selection and pricing with manufacturing cost consid
eration. T. H. Ho, C. S. Tang, eds. Product Variety Manage ment: Research Advances. Kluwer Academic Publishers, Boston/
Dordrecht/London, 145-176.
This content downloaded from 163.119.96.156 on Wed, 1 Oct 2014 08:16:41 AMAll use subject to JSTOR Terms and Conditions