|
Article Excerpt
1. Introduction
Seventy years have passed since Wright's (1936) pioneering study, yet in all that time the empirical study of organizational learning curves has experienced remarkably little technical change. Wright observed that the unit labor requirement in airframe manufacturing declined at a constant rate with each doubling of cumulative output. This phenomenon, consistent with a loglinear relationship between productivity and cumulative output, has subsequently been found to apply in numerous settings, although reported rates of learning vary widely across industries and across firms within industries (Dutton and Thomas 1984, Yelle 1979).
Among the few noteworthy innovations in the study of learning curves has been the introduction of the notion of organizational forgetting. Several researchers (Argote et al. 1990, 1997; Benkard 2000; Darr et al. 1995; Epple et al. 1991, 1995) have noted that observed costs may actually increase during certain periods of a product life cycle. In a neat appeal to symmetry with the literal interpretation of the learning curve, these researchers have argued that reversals in productivity can be attributed to organizational forgetting. (1)
There is considerable evidence that interruptions to production may be associated with organizational knowledge loss (e.g., Hirsch 1952, Baloff 1970). But the more recent studies have made the rather stronger claim that organizational forgetting occurs even under conditions of continuous production. (2) The evidence, drawn from a diverse set of industries, suggests that knowledge depreciation can be economically significant, although it varies widely across cases. Among pizza franchises, for example, Darr et al. (1995, p. 1758) found that knowledge depreciates at the astonishing rate of 17% per week, implying that "roughly one half of the stock of knowledge at the beginning of the month would remain at the end of the month." In wartime construction of Liberty cargo vessels, Argote et al. report that knowledge depreciated at the rate of 25% per month. Benkard's (2000) study of aircraft manufacturing by Lockheed generated an annual rate of depreciation of about 40%, or about 3% per month.
Although numerous explanations have been given for why organizations appear to forget in the face of interruptions to production (e.g., Anderlohr 1969), few have been given for continuous depreciation of knowledge. One--that technological change makes past experience increasingly irrelevant--is perhaps a misnomer. (3) A second explanation is that organizations often fail to record experiences because of inadequately designed organizational memory systems (Landry 1999). A third explanation is that tacit knowledge embodied in employees is lost to labor turnover. Only the third explanation has been subject to direct testing, and the evidence is a little thin. In wartime shipbuilding, Argote et al. (1990) found that labor turnover rates averaging 10% per month did not appear to affect productivity. Argote et al. (1997) found a u-shaped relationship between productivity and turnover in an American truck plant. However, Argote (1999) has noted that the rank order of knowledge depreciation rates in several studies matches the rank order of labor turnover rates.
Having accepted the evidence along with the few explanations on offer, many authors have drawn attention to the consequences of organizational forgetting for firm profitability and some have proposed strategies to help firms retain their hard-won knowledge (e.g., Belason 2000, Cross and Baird 2000, Kransdorrf 1997). Perhaps just to be contrary, Peters (1999) has gone so far as to argue that forgetting is more valuable than learning and has proposed methods by which firms can increase the amount they forget. More considered arguments along the same lines can be found in Huber (1991) and Walsh and Ungson (1991). Yet other research has shown how forgetting undermines attempts to institute flexible production schedules and just-in-time manufacturing processes (Smunt 1987). Interest is now turning toward the broader consequences of organizational forgetting. Benkard (2000), for example, has called for new theoretical efforts to explain its strategic implications.
In view of this broad interest, this paper offers another look at a familiar case study--the Liberty shipbuilding program of World War II. The episode has long been a classic case study of learning (Searle 1945, Rapping 1965, Lucas 1993, Thompson 2001, Thornton and Thompson 2001) and it is also a seminal case study of forgetting (Argote et al. 1990). The paper exploits recently discovered data on unit labor requirements (ULRs), collected by the author from primary sources at the National Archives, to produce new estimates of the rate of forgetting. The data were originally collected by auditors because cost-plus contracts paid out on individual ships demanded accurate and contemporaneous records of labor utilization at the individual product level. The quality of the data in this study is therefore comparable to those used in Benkard's (2000) analysis of aircraft manufacturing.
The unusually rich data make it possible to address two possible pitfalls in estimating forgetting rates that I think have not received sufficient attention to date. The first of these concerns problems of aggregation bias induced by unobserved changes in the product mix that can complicate inference about learning and forgetting from more aggregate data than those used here. The second concerns the relationship that can exist between assumptions about the nature of learning and inferences about the rate of forgetting. In particular, if learning is bounded, the dominant log-linear specification can produce high imputed rates of forgetting even when none exists. These issues are discussed in [section][section]2 and 3.
The main empirical results are reported in [section]4. A loglinear specification for learning with no attempt made to control for the product mix returns a rate of forgetting of 8.4% per month, about one-third of the rate reported in Argote et al. (1990). Accounting for product-mix changes reduces the estimate by half, to 4.2%. Somewhat surprisingly, replacing the loglinear specification with either of two bounded learning models has little effect on the results. Absent controls for changes in the product mix, the bounded learning models return rates of forgetting of 8.7% and 5.8% per month. In both cases, accounting for the product mix reduces these estimates significantly, to 5.7% and 3.6%, respectively; rates that are very close to those reported in Benkard (2000). The addition of controls for labor turnover also yields surprising results. When added in the usual way as a level effect, turnover rates are found to be positively correlated with productivity, while the estimated monthly rate of forgetting declines further to between -1.4% and +2.0%. An alternative specification, in which the rate of forgetting is a function of labor turnover rather than time, also produces no evidence of organizational forgetting.
Section 5 concludes the paper. Changes in the specification of the learning curve had little effect on inferences about forgetting, although there is no reason to expect this finding to hold in other applications. In contrast, adequate controls for the product mix appeared to matter, and using them yielded modest rates of organizational forgetting, at least relative to most previous findings. Attempts to relate forgetting to labor turnover were unsuccessful. To the contrary, the inclusion of labor turnover data eliminates entirely any evidence of organizational forgetting. Although these findings are robust across three specifications for learning, the concluding section includes a caveat. A major motivation for incorporating organizational forgetting into our learning models is to improve our ability to track the data. While it is certainly true that the incorporation of forgetting achieves this aim, it turns out that a wide range of rates of forgetting does almost as well as any other in terms of model fit.
2. Aggregation and the Product Mix
A challenge in measuring learning and forgetting rates with firm- or plant-level data is that at this level of aggregation inferences from output levels can easily confound within-product variations in productivity with changes in a firm's product mix. (4) Even minor product changes, unobservable to the econometrician,...
|
