Organizational Learning and the Transfer of Knowledge: An Investigation of Quality Improvement

Whereas most prior research on the learning curve has focused on improvements in efficiency, this paper deals with the impact of learning on product quality. The key data are measures of automobile reliability published in Consumer Reports. Analysis yields three findings: (1) Quality improves over the production life of a car model with the same kind of regularity as an efficiency learning curve. Thus, there is a quality learning curve. (2) Unlike in the efficiency domain, however, learning in the domain of product reliability is primarily a function of time, and not of how many cars have gone down the assembly line. Thus, quality depends not on the accumulation of production experience per se, but on the intensity of ‘‘off-line’’ quality improvement activities and on the transfer of knowledge from the general environment over time. (3) In contrast to the traditional injunction, ‘‘do not buy a new car in its first year of production,’’ the opposite advice actually seems to apply: In any given year, the newest car models have the best quality. That is, new car-model designs typically include significant quality improvements that are more than enough to outweigh any disruption created in manufacturing by the new model’s introduction and that even surpass the incremental improvements made to older, existing car models. (Learning Curve; Knowledge Transfer; Product Reliability; Quality; Cars; Auto Industry) The fact that people get better at doing things as they gain experience, but at a decreasing rate, is captured in what is often called a learning curve. This type of improvement has been well documented at many different levels of analysis: for individuals (Mazur and Hastie 1978), small groups (Leavitt 1951), work shifts within a factory (Epple et al. 1991), entire factories (Argote et al. 1990), multifactory organizations (Irwin and Klenow 1993), and industries (Udayagiri and Balakrishnan 1993). Yet despite the long history of research on organization-based learning curves going back at least to Wright (1936), the phenomenon’s theoretical underpinnings have remained murky (Kantor and Zangwill 1991). The present study contributes to the effort to identify the factors underlying the learning curve—to show that the learning curve does not just ‘‘happen’’ (Adler and Clark 1991, Zangwill and Kantor 1998). To help outline these theoretical underpinnings, it may be useful to review briefly the ‘‘five W’s’’ (who, what, where, when, and why) of organizational learning curves. This study suggests we need to rethink in particular the what and when of learning curves. Who does the learning? Argote (1993) focuses on three broad categories: individual employees, organizational systems (e.g., coordination, technology), and actors in the organization’s environment (e.g., suppliers, competitors). According to this classification, then, the ‘‘learning’’ underlying a learning curve is achieved by a combination of employees, organizational systems, and outside actors. What is learned? In most learning curve studies, the measure that shows improvement is productivity, which is an important element of organizational effectiveness, but only one element. A key goal of this study is to extend the boundaries of the learning curve concept to test its applicability to another crucial element of performance, quality improvement. While some might assume that whatever research applies to productivity learning must also apply to all other aspects of organizational performance, such an assumption is questionable. After all, organizational performance itself is not a single coherent construct; rather, it is an ‘‘umbrella’’ construct that includes many distinct organizational goals (Hirsch and Levin 1999). Thus, what we know about productivitybased learning may or may not carry over into the quality domain. DANIEL Z. LEVIN Organizational Learning and the Transfer of Knowledge ORGANIZATION SCIENCE/Vol. 11, No. 6, November–December 2000 631 Where does the learning take place? In answering this question, researchers have emphasized the important role of cognitive factors, such as search algorithms (Muth 1986), and of behavioral factors, such as engineering changes and training (Adler and Clark 1991). More recently, Zangwill and Kantor (1998, p. 913) have proposed that learning and improvement, though often observed at the aggregate level, take place in ‘‘numerous small activities [such as] entering sales data, billing, contacting clients, training, telephoning.’’ When does organizational learning take place? This question has not been widely or systematically addressed by the literature. This study tries to fill this gap by extending the theoretical argument advanced by Zangwill and Kantor (1998). Specifically, this paper argues that, when an organization’s main learning processes for noticing and improving problems are a function of time rather than accumulated production experience, then the overall learning curve will be a function of time rather than cumulative experience. This argument, presented in more detail below, runs counter to the usual finding in the learning curve literature, where ‘‘studies have found that calendar time becomes statistically insignificant once cumulative output is included in the analysis’’ (Lieberman 1987, p. 442). This study will not only explore learning during the learning curve, but also any learning before the learning curve even begins. For many years it has been routinely recommended never to buy a new car model in its first year of production (Consumer Reports 1982, p. 11; Harbour and Associates 1990, p. 52). This advice implies that quality-related learning proceeds faster during production than in the new car model design process. Consistent with the arguments of Tyre and Orlikowski (1994), however, I find evidence that the opposite is true: namely, that managers typically make the largest quality improvements during certain ‘‘windows of opportunity,’’ and one of those key windows is the period before product introduction. This early knowledge determines the starting point of a product’s eventual learning curve. The current study extends this idea of a window of opportunity by proposing that the larger the opportunity, the greater the amount of learning and improvement activities that can occur. For a car model’s minor update or facelift every year (the learning curve), the improvement will be relatively small; for a major model change or all-new car model introduction, the learning will be larger. Why does learning occur? Few researchers have asked this question because it is hard to imagine why people would not want to improve. As long as learning curve research focuses only on costs, researchers will only find instances in which the motivation to learn is present. Yet, while firms typically have a strong economic incentive to reduce their costs, they do not always have an incentive to improve product reliability. However, by the 1980s, U.S. automakers finally had a major strategic incentive in the form of Japanese competition. As one industry analyst noted, ‘‘The decade [of the 1980s] started with product quality being the domestic industry’s number one problem. It was a major reason why it was losing market share to Japanese imports at an alarming rate’’ (Harbour and Associates, 1990 p. 142). The auto industry, then, should be an especially good context in which to find and examine learning curves for quality. In sum, this study extends our understanding of learning curves in three main ways. First, it expands the what of learning curves beyond efficiency to include a new area, quality improvement. Second, in terms of when (and perhaps how) learning occurs, it provides evidence that at least some kinds of learning are time-based instead of experience-based. Third, by comparing the two times when learning can occur, this study shows that considerably more improvement occurs as a result of new product introduction than during any year of subsequent production (i.e., the learning curve). Hypotheses Quality improvement in the car industry arises out of two processes occurring at the product level. For a given car model at a given point in time, quality is a function of (1) improvement made during the car model’s production life, and (2) the car model’s initial baseline level of quality. Most learning curve research focuses on the first process (‘‘the learning curve’’ itself), but, especially in the car industry, the second process (new product development) is also important. I begin by investigating the learning curve itself, then step back to see the effects of new product introduction, and, finally, I compare the effects of each process. The first three hypotheses (H1–H3) address the rate of improvement for the learning curve: Does a learning curve for quality exist? If so, what form does it take? What factors influence it? Then, separate from these questions of what happens during the learning curve itself, Hypothesis 4 asks what happens during new product introduction, before the learning curve even starts: Is there improvement in the ‘‘starting points’’ of these learning curves? Finally, Hypothesis 5 compares these two types of learning—annual learning-curve improvements versus annual starting-point improvements— to see which has a greater impact. Quality can be defined in many different ways (Garvin 1988). The definition of quality used in this study is a product’s reliability, measured by frequency-of-repair DANIEL Z. LEVIN Organizational Learning and the Transfer of Knowledge 632 ORGANIZATION SCIENCE/Vol. 11, No. 6, November–December 2000 rates. While it is interesting to look at ‘‘internal’’ measures of quality, such as conformance to specifications, it is arguably even more interesting to take the customers’ point of view and focus on their experience of quality. Improvements in quality, and more specifically auto repair rates, should take the form of a learning curve. After all, ‘‘we might expect that when a car has been in production several years, management should have invested in identifying and solving most of the product’s quality problems. In other words, there ought to be a learning curve’’ (Cole 1990, p. 80). We expect this quality improvement to take the shape of a learning curve for the same reason that we see a learning curve for productivity improvement: namely that, after a while, ‘‘the easiest gains have already been made—the cream has been skimmed. New gains in product quality [then come] more slowly and appear to [get] more expensive’’ (Cole 1990, p. 77). The implication is that quality improvement in automobiles will show a learning curve for quality, as measured by repair rates. HYPOTHESIS 1. The average car model improves in quality during its production life, but in decreasing increments. Note that Hypothesis 1 (H1) leaves open the question of whether this quality-improvement learning curve is a function primarily of experience gained over time or as a result of cumulative production volume experience. Granted, these two possibilities are highly correlated, and their effects are not always so easy to tease apart (Adler 1990). Yet, from a theoretical perspective, this open question is an interesting one because it helps us to get at some of the underlying sources of organizational learning. In the well-documented domain of productivity-based learning curves, Argote (1993, p. 41) notes, ‘‘empirical studies have demonstrated that calendar time is generally not a significant predictor of organizational learning [once one takes into account the role of] cumulative output at the particular group or organization.’’ We might therefore expect a similar effect, in which cumulative experience (based on production volume) is more important than the ‘‘mere’’ passage of time, for quality as well. After all, the growing popularity in the manufacturing domain of continuous improvement and statistical process control as a means of improving quality in the auto industry suggests that cumulative production experience should be the driver of quality improvement. HYPOTHESIS 2. Quality improvement during the average car model’s production life is a function of cumulative production output. There is, however, reason to question this expectation, for in the quality domain, one might expect differences in (1) how people notice problems of poor reliability and (2) how they solve such problems. There is, by definition, a time lag between when a product is produced and when producers start to notice how reliable it is. That is, producers get less immediate feedback for reliability than they do for efficiency, or even for defects. Fortunately for automakers, it does not always take years to get feedback on product reliability. For example, they can get early warning signs from warranty claims, high-mileage vehicles like taxis and rental cars, and from performing their own endurance tests. However, this type of feedback does not depend especially on how many units of a given car model were made. Rather, it depends on how intensively, during a given period of time, people focus on gaining this knowledge. The use of this feedback to solve problems of poor reliability may also be more a function of time than of production experience. The total quality management (TQM) movement has relied heavily on off-line improvement teams. These teams first identify the root causes of problems, then propose, design, test, and implement solutions to the most important root cause until it is corrected; then move on to the next most important source of problems. Indeed, this improvement cycle ‘‘is the iterative learning loop at the heart of TQM’’ (Sterman et al. 1997, p. 504). As a consequence, learning in the quality domain is likely to come not so much from how many cars have gone down the assembly line, but from the intensity of ‘‘off-line’’ activities. And indeed, past studies of TQM have assumed that the form of improvement for quality improvement activities is largely a function of time (Schneiderman 1988, Sterman et al. 1997). Given the importance of time-dependent qualityimprovement cycles for solving reliability problems, as well as the time lag for how these problems get noticed in the first place, a competing hypothesis to H2 would be: HYPOTHESIS 2–ALT. Quality improvement during the average car model’s production life is a function of time. Learning in an organization can be a function not only of the organization’s own internally generated knowledge, but also of knowledge transferred from the organization’s environment (Argote 1993). What are some likely sources of transferred knowledge that would enhance a car model’s learning curve for quality? One likely source might be the makers of other, closely related car models. Prior research, after all, has found that knowledge transfer occurs within similar categories of nuclear power plants (Lester and McCabe 1993), pharmaceutical research programs (Henderson and Cockburn 1996), and semiconductors (Udayagiri and Balakrishnan 1993). In DANIEL Z. LEVIN Organizational Learning and the Transfer of Knowledge ORGANIZATION SCIENCE/Vol. 11, No. 6, November–December 2000 633 the automobile industry, the opportunity to share similar information and technology could potentially occur among car models within the same product ‘‘family,’’ called a vehicle platform. Even car models from different product families often share technological processes, suppliers, and certain common parts like seatbelts. It is also easier to transfer knowledge across groups or organizations if they have a shared culture (Argote 1993), so among car models in the same division or company, we might expect to see some knowledge transfer. Among U.S. automakers, too, there is a shared ‘‘Detroit’’ culture of sorts (Ingrassia and White 1994), as well as some shared suppliers (Harbour and Associates 1990). Along these lines, several prior studies have investigated cumulative output-based knowledge transfer (also called ‘‘spillovers’’), where the cumulative production output of one group or organization benefits another (Argote 1993, Darr et al. 1995, Irwin and Klenow 1993). Thus, it may be possible to detect production-based knowledge transfers in the quality domain from the following sources: HYPOTHESIS 3. The average car model’s quality improves as a function of the cumulative production output of its (a) platform, (b) division, (c) company, and (d) industry. We should note an alternative view of these spillover effects: It is also possible that the external knowledge available to stimulate quality improvements accumulates in a manner more related to the passage of time; e.g., by reverse engineering a competitor’s product, or by learning from a supplier’s technological breakthrough, or by avoiding the mistakes of earlier projects within the firm. We might therefore imagine an H3–ALT, focusing on time. This process, however, is already captured by H2– ALT, because car models from the same platform typically debut at the same time and because U.S. automakers and their divisions began operations too many decades ago for any differences in founding dates to be meaningful in this context. Indeed, prior research (Argote 1993) has identified the passage of time (e.g., as in H2–ALT) with a generalized knowledge transfer, or ‘‘march of progress.’’ H2–ALT thus reflects, in a general way, the benefits to car model producers from others’ knowledge of how to solve reliability problems. So far we have focused on the rate of improvement for existing products; i.e., the learning curve. Over time, however, companies introduce new products with major—even dramatic—improvements already built in. Surely, then, no picture of quality improvement and learning in the auto industry is complete without considering the learning that takes place as a result of new product introductions, before a car model’s learning curve even begins. For example, previous studies (Consumer Reports 1991, p. 248; Harbour and Associates 1990) have documented the fact that each of the Big Three U.S. automakers—Chrysler, Ford, and General Motors (GM)— made significant overall strides in improving quality during the 1980s. But, if we return to the ‘‘five W’s’’ of learning curves, we must ask when this improvement takes place. That is, do these overall improvements come from (1) incremental improvements made to existing car models (the learning curve examined in H1–H3), or from (2) the introduction of all-new car models to replace poorquality ones? This second type of improvement affects the ‘‘starting point,’’ or baseline performance level, of the learning curve. The null hypothesis would be that, over time, all of an automaker’s new car models, on average, have the same baseline quality level; i.e., a debuting car model must start from scratch, with no carryover of learning from previous car models and with no benefit from any newly created knowledge. In contrast, a ‘‘march of progress’’ view would suggest that it is likely that makers of debuting car models learn from the past. Moreover, in this view, the debut of a new model provides a window of opportunity (Tyre and Orlikowski 1994) for manufacturers to incorporate the very latest knowledge on design, materials, assembly, etc., because the usual constraints (e.g., the large costs in changing any production line technologies) are temporarily lifted when a new product line debuts. HYPOTHESIS 4. The later a car model debuts, the better is its baseline quality level. This hypothesis is tested in terms of time; i.e., it focuses on the year a car model debuts. (Note that it might also be interesting, as part of a future research study with the benefit of a longer time series, to test a comparable H4– ALT, based on cumulative production output, as well.) If H4 is supported, then we can say that there is at least some learning present before a new learning curve begins. But how does this new-and-improved starting point compare to the benefits of being at the end of a long learning curve? Is this ‘‘debut-year’’ learning enough to compensate for the lack of production experience? Prior research (e.g., Hayes and Clark 1986) tells us that the introduction of a new product into a repetitive manufacturing context can be highly disruptive, impairing efficiency and sometimes dramatically so. Such disruption is also likely to impair quality. This ‘‘disruptive learning’’ perspective would suggest that, even if the baseline quality improves, it will still take time before product quality reaches (and eventually surpasses) its previous level. After all, one could argue, any learning requires at least some ‘‘unlearning,’’ and so some disruptions are probably inevitable. DANIEL Z. LEVIN Organizational Learning and the Transfer of Knowledge 634 ORGANIZATION SCIENCE/Vol. 11, No. 6, November–December 2000 Figure 1 Three Alternatives for How a New Car Model’s Debut Might Affect the New Learning Curve’s Starting Point This view is expressed in the old industry adage, ‘‘Don’t buy cars in their first year’’ (Consumer Reports 1982, p. 11; Harbour and Associates 1990, p. 52). This saying is based on the premise that the learning that ultimately occurs during production will outweigh any reliability enhancements embedded in a brand-new product. By contrast with this adage and with the cases discussed by Hayes and Clark (1986), Garvin (1988), who investigated the defect rates of factories making room air conditioners, found that although defects did worsen for a day or so after the annual shutdown period, no effect was detectable in the quarterly data. That is, new models pick up the quality levels right where the older models leave off. According to this ‘‘constant learning’’ view, debut-year improvements and learning-curve improvements are equal, so that there is a constant rate of learning for automakers as they transition from older car models to newer ones. A third alternative derives from a theoretical perspective we might call ‘‘enhanced learning.’’ This perspective suggests that debuting car models will benefit from some combination of the carryover from previous learning, minus a few minor disruptions, plus the exploitation of the larger window of opportunity provided by the debut. This last point implies that more extraordinary or exceptional learning will occur during product debuts, and have a greater effect, than the more incremental changes made during a product’s production life (Lant and Mezias 1992, Tushman and Romanelli 1985). Although the basic design of cars has not changed fundamentally during the past century, the introduction of a new product is nonetheless an important discontinuity in design, manufacturing, and sales. Thus, relative to day-to-day changes, or even to annual model changes, ‘‘major’’ model changes or debuts can indeed involve more ‘‘radical’’ rethinking. Assuming firms are motivated to improve quality, as U.S. automakers surely were in the 1980s, then it follows that the larger the opportunity to improve, the greater the learning and improvement will be. That is, although the window of opportunity to improve a car model during its production life occurs at least every year, these opportunities are relatively small in scope. Change too much and factory costs will rise; take too long to make changes while the assembly line is shut down, and lost product sales will mount. But the window of opportunity is much larger when a new product first debuts. So much is changing anyway that the marginal cost of making major quality improvements is likely to be relatively low. Furthermore, to minimize costs and disruptions, organizations may wait until the next new product is introduced to incorporate new manufacturing and product technologies that will enhance reliability. Thus, the large window of opportunity makes possible a greater improvement in product quality from the organizational learning embedded in new products. So is this debut-year learning enough to compensate for the associated disruption? Figure 1 shows the three possible answers we have discussed. (Note that the vertical axis in Figure 1 is in logarithmic form, so the learning curves appear linear.) The final hypothesis, which is based on the ‘‘enhanced learning’’ perspective, frames the key test: HYPOTHESIS 5. In any given year, car models with the latest debuts will have the best quality, despite having a shorter production life in which to make improvements.