Improving Quality Through the Concept of Learning Curves
W. DALE COMPTON, MICHELLE D. DUNLAP, and JOSEPH A. HEIM
Quality is the hallmark of competitive products. Consumers reject products that are of inferior quality, and they shun companies who are perceived to provide products or services with less than competitive quality. A company cannot survive in the current world marketplace without providing a product or service that is of high quality. It would be hard to find a U.S.-based manufacturing enterprise that does not place quality near the top of a list of strategic or operating objectives that would also include cost, innovation, and customer focus.
This sensitivity to the customer demand for quality has not always been a dominant force in the operating strategies of U.S. companies. Having realized its importance, companies find that they must now direct their energies in ways that focus on this objective. A broad range of operating procedures must be modified or, in some cases, created: quality must be designed into products; a productive interaction must be stimulated among the design, manufacturing, and marketing activities; input from the customers must be obtained and used; and active participation by their employees in creating “quality” products and services must be encouraged.
With this new awareness of the importance of quality have come two specific needs. First, there is a need to measure the performance of the organization against that of its competitors and, second, there is the need to assess the trends in one's performance in order to take appropriate actions to ensure continuous improvement. In the first case, an absolute measure of
performance is needed. This is sometimes referred to as “benchmarking,” or measuring oneself against the world leader—the “best-of-the-best”—in a product or process arena (Compton, in this volume). In the second, progress over time is the prime concern, that is, how well the organization is achieving continuous improvement in performance. A proper combination of these two measures is critical. Without them an organization cannot properly evaluate its absolute competitive status, nor can it be assured of its ability to remain competitive over a long period of time. With “high quality” as a prerequisite for being competitive in the marketplace, achieving and maintaining high quality in all aspects of an operation is a critical component of the foundations of all manufacturing systems.
This article focuses on the second of the two needs identified above, assessing how an organization improves over time. We will be concerned, therefore, with assessing trends in quality. We will conclude with some observations about the need for continuous and careful collection of the type of data that are critical to a proper assessment of progress.
“Quality” is not a universal descriptor that has a unique definition under all circumstances. Garvin (1984) has described five approaches to defining quality. The appropriate metric for measuring the quality of a product or process will depend on the definition or circumstance that is of immediate interest. It can, for example, refer to defects arising from a production process, defective parts shipped to customers, or reliability of the product in service. Although we will not discuss the various measures of quality in this paper, we have obtained examples of each of the above measures. We offer examples of the first two in this paper.
Measures of the quality of the outputs of a system can be obtained in many ways. In the day-to-day operation of a manufacturing enterprise, the collection and use of process data to support statistical process control (SPC) is important to achieving high-quality manufacturing. SPC requires that measures of one or more attributes of the quality of a production system be regularly employed, and it provides a paradigm for the efficient use of those measures to control the process. There is ample evidence of the importance of this real-time control in improving the quality of the processes and the products that result from these processes. In every sense, the effective use of SPC and total quality control (TQC) have become important elements of the foundations of effective manufacturing systems.
The measures of quality that are implicit in the application of SPC necessarily concern shorter time periods; that is, they reflect the current status of the process or system that is producing the product. Although the importance of this near-term collection of data—and appropriate analysis to
accomplish SPC—is not questioned, it is also clear that an understanding and quantification of the longer-term trends in quality are also critical to achieving continuous improvement in quality.
THE LEARNING CURVE RELATED TO COSTS
A traditional approach to measuring the long-term cost performance in a manufacturing operation is to use the “experience” or “learning” curve concept (Henderson and Levy, 1965). This asserts that the fractional reduction in the average cumulative cost (in constant units of measure) of producing a product is proportional to the fractional increase in the quantity of the product that is produced and yields a power law representation that is similar to that first described by Wright (1936). A common formulation of this law relates the cost of production of the nth unit, Xn to the total production volume N
Xn= KN−b (1)
for large N.
Equation 1 has been used many times and has been shown to be valid for a wide variety of products in many different industries (see Argote and Epple, 1990, for a discussion of this form of the learning curve in manufacturing). The literature contains numerous discussions of circumstances in which an exponential law is the appropriate formulation for the learning curve (Buck et al., 1976; Pegals, 1969). A simpler formulation for a learning curve, seldom used in the literature and applicable only under limited circumstances, is the linear representation. Determination of the form that is most appropriate depends on many factors, including the nature of the data sampling protocol. In general, however, if it is not possible to determine which form is most appropriate, either because of an absence of a priori knowledge or because of a lack of sufficient high-quality data, the simplest formulation is probably best. Selecting the simplest formulation entails testing to determine whether the data are best fitted by a linear, an exponential, or a power law representation.
Irrespective of the formulation chosen, learning curves are not to be viewed as merely descriptive. They can be, and frequently have been, used as an aid in making predictions, in that early experience in the production of a product can be used to predict future manufacturing costs. Assuming that one has confidence in the form of the equation that is chosen—whether power, exponential, or linear—and that one can make a reasonable estimate of the constants that appear in them, one can readily predict the costs to produce a unit after some future cumulative production volume has been achieved. Even in the absence of detailed data on a given product, the experience of many manufacturers with many products is that manufacturing costs can be expected to decrease by 10 to 20 percent for each doubling
of production volume. Abernathy and Wayne (1974) have explored the limits of validity of the learning curve concept.
The improvement depicted by the experience curve is a result of conscious effort and attention on the part of the management and employees of the enterprise. It cannot be expected to continue without the attention and focus that accompanies a clearly accepted operating objective, in this case an objective of continuously reducing the costs to manufacture the product or to offer the service. A variety of actions combine to produce the desired cost reductions (Allan, 1975):
Improved efficiency in the use of labor through training and incentives.
Introduction of new and improved processes that reduce manufacturing costs.
Redesign of the product to reduce manufacturing costs.
Standardization of the product to reduce the variety of tasks demanded of the workers.
Scale effects resulting from large volume production.
Substitution of lower-cost materials while retaining product features.
THE LEARNING CURVE RELATED TO QUALITY
Just as competitive pressures have forced the management of U.S. companies to pay special attention to costs, so also is management being forced to pay special attention to improvement in quality. Although many approaches are taken to improve quality, these efforts have a few key actions in common:
Simplification of product design to enhance manufacturability.
Involvement of the employees in designing the manufacturing system.
Enhanced training of the employees.
Substitution of automated machinery in areas that are not conducive to human operation.
Collection of extensive data on each operation, and analysis to identify problems and trends in those operations.
Introduction of new or improved processes that are less sensitive to variation.
Although the specific actions taken to improve quality differ from those taken to reduce unit costs, a striking similarity exists between the two lists. In particular, both result from conscious actions taken by management and employees to accomplish a common strategic objective for the enterprise.
Both combine human commitment and training with technical improvements. Both require extensive knowledge of the processes being employed and the products being produced. Therefore, quality and costs might be expected to share a common representation. One might then speculate that quality should follow an experience curve similar to that of cost. By analogy, therefore, a quality learning curve might take one of three forms such that the quality index (QI) for the nth item is defined as follows:
In the above equations, (QI)n* is the asymptotic value of the quality index, (QI)o is related to (QI)l, the quality of the first unit produced and to (QI)n*, (QI)a and (QI)b are constants, and N is the cumulative volume of the products produced. In Equations 2, 3, and 4, the sign can be positive or negative—positive if the quality index is improving as cumulative production volume increases, for example, yield from a process; negative if the quality index reflects defects or defective parts, which will decrease as the cumulative production volume increases.
While the particular attribute of the product or process being considered will most likely be different for each product and process, the above formulations are independent of the specific attribute that can be related to the quality index. One should not expect, however, the numerical values of the constants to lie within a specific range or to have any particular relationship from one product to another, because the quality indices can differ depending on the attribute chosen for examination.
Schneiderman (1988) appears to be one of the first to treat production yields or the quality of products shipped according to a learning curve. Schneiderman offers a number of examples of quality learning curves that are presented as exponential formulations in which a measure of quality is plotted as a function of time from the start of production. It should be noted that this formulation is consistent with Equation 3 only in the case that production rates are constant over time—a circumstance that seldom occurs.
A test of the hypothesis that a quality index is describable by Equations 2, 3, or 4 can, in principle, be made by examining the quality of products or processes at various levels of production. For some dozen products—for which measures of quality and production volumes could be obtained—we have generally found that two of the three formulations are virtually indis-
tinguishable in their ability to represent the data. In some cases, the linear and exponential laws were indistinguishable—meaning that the coefficient of correlation for the two laws was nearly the same —while in others, the power and exponential laws were indistinguishable. We found no case in which all three representations were equally good.
Data are presented in Figure 1 Figure 2 through Figure 3 relating an index of quality to the cumulative volumes of production for three different products—light bulbs, a small electric motor, and grey iron castings. General Electric Company and The Dalton Foundries, Inc., graciously supplied the data contained in these figures.
A description of the quality index for each of the products is given in the figure captions. Having no a priori basis on which to choose the preferred formulation for representing the quality index, we examined each of the products using Equations 2, 3, and 4. Following an observation by Buck et al. (1976) that the exponential form of the learning curve is some-
what to be preferred for batch, or average, sampling of the metric in question, we have chosen to present three of these sets of data in terms of the exponential relationship of the quality index to the cumulative volumes of the product produced. In none of these cases was saturation apparent, implying either that (QI)n* was effectively zero or that the observed values are so far from the saturation value that the present representation is not adequate to display a saturation. It is of particular interest that the correlation coefficients for a linear plot of the data shown in Figure 1 and Figure 3 are essentially the same as shown for the curves as plotted. The correlation coefficients for the curves shown in the graphs are given in the captions, along with the correlation for the best alternative formulation. Each of the data points in these three figures represents an average of the quality metric for a period of one year. Thus, for Figure 1, the quality data are for 20 years of production, Figure 2 for 9 years, and Figure 3 for 12 years. In some cases the index is defined as defects in production; in others, the shipping of a faulty product to a customer.
In each of the cases examined, there is clear evidence that the quality index, although defined differently for each group of products, is related to cumulative volume of production. With the diversity in product type and processes represented here, the hypothesis that one or more forms of the learning curve exists for quality is supported by these data. Although the present data do not appear capable of distinguishing among the various forms for a learning curve for quality, it appears that one or more forms can easily be found to permit a reasonable extension for setting new goals or examining the impact of past actions on performance.
Because the time frame in which these products were in production is long, it is reasonably certain that the actions suggested earlier as being important for management and employees in achieving a continuous improvement in quality were taken throughout the life of these products. Each of the companies that provided these data has indicated that the quality trends demonstrated in these figures are the result of constant and consistent
attention to the importance of continuous improvement. Although this dependence of a quality index on cumulative production has been demonstrated for only this select group of products, we believe that this phenomenon is generally true. A collection of additional examples from other industries would help support this conclusion.
We have been surprised to find that few companies keep data in the form or with a consistency that allows the following of trends as described in the types of curves shown here. In our view, this is a shortcoming that should be addressed by all concerned with continuous improvement. The systematic collection of data on quality and the representation of these data in the form described by Equations 2, 3, or 4, offer a means of tracking progress on the “continuous improvement ” of quality and a means by which realistic expectations can be established for future goals. Above all, the existence of a learning curve for quality should be viewed as one more example of the need for careful collection of systematic data. Without good data, this important foundation cannot be used.
Creating high-quality products through high-quality manufacturing processes and systems is a critical element in the foundations of manufacturing. The learning curve for quality should therefore be viewed as an important element in the foundations of manufacturing. We believe that the learning curves can be an important contributor to achieving improved quality.
This work was supported in part by a grant from the Ford Motor Company Fund.