1. The Shipyard Study
Our first empirical study of knowledge depreciation was based on data from the construction of the Liberty Ship during World War II (Fisher, 1949). We learned about these data from Rapping’s (1965) study, in which he found evidence of learn- ing in the shipyards. Rapping’s study provided particularly compelling evidence of learning because he controlled for important factors such as economies of scale and technological progress associated with the passage of time in his analysis. Rapping found significant evidence of learning when these important additional factors were taken into account.
The Liberty Ship production program is particularly attractive for studying orga- nizational learning and forgetting. The Liberty Ship was built in 16 different ship- yards in the USA.1 Each of the yards producing the Liberty Ship began production during 1941 or 1942. The yards were new yards, the Emergency Shipyards, con- structed under the authority of the US Maritime Commission. A central agency was responsible for purchasing raw materials and equipment, approving each yard’s lay- out and technology, and supervising its construction (Lane, 1951). A standard design was adopted and produced with minor variation in all of the yards (Lane, 1951). The overwhelming majority of workers employed in the Emergency Shipyards had no prior experience in shipbuilding (Fisher, 1949). The yards con- structed a very large number (almost 3,000) of Liberty Ships. On average, 2 months were required to build each Liberty Ship.
Thus, the Liberty Ship production program is ideally suited for studying organi- zational learning, forgetting and transfer (See Chap. 6 for a discussion of the trans- fer results). Many ships were produced from homogeneous raw materials in different organizations by workers without prior industry experience. These features of the Liberty Ship production program control naturally for many important factors, such as prior experience of workers, which are difficult to control for statistically in many production environments. Further, the shipyards began production at different times, produced at very different scales of operation, and experienced different rates of labor turnover. Thus, there was variance on important factors that allowed us to examine whether knowledge depreciated and whether it transferred. We also explored the role of labor turnover in the acquisition and depreciation of knowledge. The following section describes in depth the method we used to assess whether knowledge depreciated. Readers less interested in the details of the estimation might wish to skip to the results or discussion sections.
Table 3.2 Variables used in the shipyard study Symbol
1.1. Method and Sources of Data
Our general approach was to estimate production functions in which output pro- duced in a given period depended on the inputs of labor, capital, organizational experience, and other variables. Variables used in our analysis of the shipyards data and the symbols used to designate them are listed in Table 3.2.
Our primary dependent variable was tonnage of ships produced per month. Tonnage produced per month refers to the weight of all vessels or portions of vessels produced during a month. Womer (1984) demonstrated that inappropri- ate inferences can be drawn from empirical analyses of learning if the measure of output is based on units finished in a given month, and the period of produc- tion exceeds a month. This problem did not arise in our analysis of the shipyard data because our output measure was the tonnage actually constructed in a given month.
Independent variables included measures of labor and capital inputs. Following Rapping (1965), shipways in use was our measure of capital inputs. Shipways are the structures upon which the ships were built.
We estimated models in which output in a given period depended on the inputs of labor (labor hours), capital (shipways), experience and other variables. Specifically, we estimated production functions of the following form:
As noted previously, Womer (1979) emphasized the importance of integrating the neoclassical production function approach and analyses of learning by doing. Because the Liberty Ship data were from several organizations that differed significantly in their scale of operation, we were able to integrate the production function approach and learning by doing. By controlling for inputs of labor and capital, we were able to separate increases in productivity due to learning from increases in productivity due to increasing exploitation of economies of scale. In addition, we controlled for calendar time in order to separate the effect of technical progress associated with the passage of time in the general environment from pro- ductivity improvements associated with increasing experience at a particular ship- yard. The vector Zit in Eq. (3.1) varies from regression to regression and represents these other variables that could influence productivity.
Because our empirical analysis of knowledge depreciation is a new contribution, an explanation of how we measured depreciation is developed. Our approach to measuring depreciation is formalized in Eq. (3.2). Variable Kit is the stock of experi- ence accumulated by yard i at date t. Equation (3.2) allows for the possibility that the stock depreciates over time by including the parameter lambda, λ. Lambda is estimated through a scanning procedure for maximum likelihood estimation (Dhrymes, 1966; Goldfeld & Quandt, 1972). If λ is estimated to be one, the accu- mulated stock of knowledge is simply equal to lagged cumulative output, the classic measure of experience. If λ is estimated to be less than one, however, there is evi- dence of depreciation because output from the distant past receives less weight than recent output.
As an example of the implications of estimates of lambda, consider a case where lambda is estimated to be 0.90. If the estimate of lambda (0.90) is substituted into Eq. (3.2), output from the previous period would be weighted by 0.90, output from the period before that would be weighted by 0.81 (λ2 = 0.902), output from the period before that would be weighted by 0.73 (λ3 = 0.903), and so on. Thus, if lambda is estimated to be less than one, output from the distant past receives progressively less weight than recent output in predicting current productivity. Further, as lambda becomes smaller, output from the distant past receives even less weight. Thus, the smaller lambda is, the less knowledge is retained, and the more depreciation of knowledge occurs.
Our measure of knowledge computed from Eq. (3.2) is substituted as a predictor variable into Eq. (3.1).2 As Eq. (3.1) indicates, knowledge acquired through the end of the preceding month, Kit-1, appears in the production function for month t. Thus, past but not current output appears on the right-hand side of Eq. (3.1). The coefficient on gamma (γ) in Eq. (3.1) indicates whether organizational learning occurred. If gamma is significantly different from zero, organizational performance changed as a function of experience. Thus, organizational learning occurred; the organization acquired knowledge. If the lambda (λ) in Eq. (3.2) is significantly less than one, depreciation occurred. The knowledge acquired from recent experience is a more significant predictor of current performance than knowledge acquired from experience in the more distant past.
We controlled for unmeasured yard-specific factors that could affect the produc- tivity of the yards. In Eq. (3.1), the Di are “dummy” variables for each shipyard. These dummy variables are included to capture unmeasured yard-specific factors such as land that could affect production and are relatively constant through time.
The error uit is assumed to be serially correlated as shown in Eq. (3.3). The serial correlation coefficients are assumed to be the same across all shipyards. The choice of a third-order autoregressive error was based on our analysis of the data. Third- order serial correlation coefficients all reached at least the .05 level of significance. This is not surprising because production of a ship required an average of 2 months, and longer periods were required early in the operation of the yards.4
Our findings concerning whether knowledge depreciated in the shipyards are pre- sented in Table 3.3. Results obtained from estimating five different models of cur- rent production are shown in Table 3.3. The dependent variable in all of these models is the logarithm of current production; the predictor variables for each model vary from column to column. The models were estimated by maximum likelihood. The coefficients of the yard-specific dummy variables are not of particular interest and are therefore not reported. A joint test of the null hypothesis that there were no yard- specific effects was rejected at a high significance level (p < .001), so important yard-specific effects were present. Yard-specific dummy variables were included in all analyses shown in Table 3.3.
Estimation was done using the following search procedure. Values of lambda, l, in the interval [0,1] were chosen. With lambda fixed, the remaining parameters were estimated by standard procedures for estimating regression models with autocorre- lated errors. Hence, for each chosen value of lambda the remaining parameters were estimated. We began with a search over values of lambda at increments of 0.05 in the interval [0, 1] to identify the subinterval in which the function reached a maximum and then located the maximum by searching that subinterval at increments of lambda of 0.01. The maximum likelihood estimates for the overall model were the value of lambda and the values of the associated coefficients that yielded the largest value of the likelihood function. This procedure is equivalent to nonlinear search procedures that vary all parameters simultaneously, but is computationally easier to implement. The maximum likelihood estimate of the depreciation parameter, lambda, for the model shown in Column 1 of Table 3.3 was 0.75. The estimation procedure did not yield a standard error for lambda. The standard errors of the remaining coefficients were computed treating lambda as a known parameter. This could result in some understatement of the standard errors of the coefficients and a corresponding over- statement of the t-statistics. Therefore, all conclusions regarding significance of alternative measures of learning were based on likelihood ratio tests. Using the dis- tribution of the likelihood ratio, we determined that a 93% confidence interval for lambda was approximately (0.65, 0.85). Thus, the hypothesis of no depreciation (λ = 1.0) was very strongly rejected by these data. Hence, the classic measure, cumu- lative output, significantly overstated the persistence of knowledge. Depreciation of knowledge was found to occur in this production environment.
These results indicate a rapid rate of depreciation. As noted previously, the results were obtained from monthly data. A value of λ = 0.75 implies that, from a stock of knowledge available at the beginning of a year, only 3.2% (λ12 = 0.7512) would remain 1 year later. Thus, if the stock of knowledge is not replenished by continuing production, it depreciates rapidly.
Third-order serial correlation coefficients all reached at least the 0.05 level of significance. Estimates of the other parameters were not sensitive to the order of serial correlation. For example, with either first- or second-order autocorrelation, the maximum likelihood estimate of lambda for the model in Column 1 of Table 3.3 was 0.80. Because autocorrelation coefficients up to third-order were significant, we adopted the third-order specification for the remaining analyses.
We investigated alternative models to see if a specification of the learning pro- cess that allowed for depreciation was more satisfying than one that did not. Column 2 of Table 3.3 is identical to Column 1 except that Column 2 included the classic cumulative output measure whereas Column 1 included the knowledge variable that allows for depreciation. The log of the likelihood function in Column 1 was significantly greater than the log of the likelihood function in Column 2, x2 = 41.59, df = 1, p < .0001. The contrast of Columns 1 and 2 provides further evidence that learning depreciated and that our knowledge variable was a better measure than the conventional one, cumulative output.
Results presented in Column 3 of Table 3.3 show the effect of including the con- ventional measure, cumulative output, and our knowledge variable in the same model. When cumulative output is included, the value of lambda that maximized the likelihood function was 0.75, as in Column 1. The conventional measure, cumula- tive output, had a small and statistically insignificant coefficient, whereas the knowl- edge variable was highly significant in this model. This is further evidence that the knowledge variable that allows for depreciation is better than the conventional mea- sure, cumulative output.
In Column 4, calendar time was introduced to capture the possibility that techni- cal change associated with the passage of time rather than learning in the shipyards was responsible for productivity improvements in shipbuilding. The negative coefficient for the time variable indicated that the mere passage of time did not explain the productivity gains. When the more general translog specification of the production function (Berndt & Christensen, 1973) was used, the coefficient of the calendar time variable was smaller in magnitude and statistically insignificant. Thus, there was no evidence that the shipyards became more productive simply as a function of the passage of time.
What are the implications of depreciation for future productivity? Figure 3.1 depicts results from a simulation to illustrate the effect of depreciation on unit costs. In this simulation, input levels were held constant at the sample mean for the first 13 months of operation of a shipyard. As can be seen from Fig. 3.1, during this period, unit costs declined at a decreasing rate. The levels of inputs were halved in month 14 and held constant thereafter. The reduction in input caused an immediate increase in unit costs, indicated by the vertical line at the date of the reduction. This jump in unit costs at the vertical line was due to scale economies. The subsequent increase in unit costs was due to the depreciation of knowledge. The reduction in inputs led to a reduction in output which in turn reduced the knowledge variable, K. Gains in knowledge from current production were not sufficient to offset the losses in knowl- edge from depreciation of the previous period’s stock. Thus, depreciation of knowl- edge led to an increase in the unit cost of production.
Why did knowledge depreciate? One very plausible cause for knowledge depre- ciation was personnel turnover. If knowledge were embedded in individuals and those individuals were “separated” or departed, their turnover might hurt organiza- tional performance. Therefore, we investigated the role of labor turnover in the pro- ductivity gains. Labor turnover was included in the model shown in Column 5 of Table 3.3. The rate of new hires and the rate of departures were included as explana- tory variables. This model shown in Column 5 has fewer observations than previous runs because of missing data for the rates of hires and departures. Hence, the likeli- hood function value for this equation cannot be compared to those for other equa- tions in Table 3.3. As can be seen from Column 5, these variables together did not contribute significantly to explaining changes in productivity. Additional analyses revealed that neither variable made a significant contribution when included sepa- rately. The estimate of the depreciation parameter in Column 5, λ = 0.70, indicates that knowledge depreciated rapidly, even after the effects of labor movement were taken into account.
Fig. 3.1 The relationship between unit costs and cumulative output when inputs are reduced to half their initial levels. Note: Reprinted by permission from L. Argote,
Conditions at the shipyards could have buffered them from the effects of turnover. A war was on—jobs were designed to be low in skill requirements so that inexperi- enced workers could be brought up to speed very quickly. Therefore, there was tre- mendous emphasis on standardization and formalization (Lane, 1951). Improvements made in one part of the yard were quickly codified and transmitted to others.
In summary, the results presented in Table 3.3 indicate that organizational learn- ing occurred: productivity improved markedly as the organizations gained experi- ence in production. As discussed in Chap. 1, learning curves are often characterized in terms of a progress ratio, p. The progress ratio describes what happens to unit costs with each doubling of cumulative output. The parameter, b, in Eq. (1.1) is related to the progress ratio, p, by the expression p = 2-b. The progress ratio derived from the estimate of b in Column 2 of Table 3.3 is 2-44%. This is a remarkable rate of productivity. With each doubling of the cumulative number of ships produced, the unit cost of production declined to 74% of its former value. This rapid rate of productivity growth is generally consistent with other analyses of productivity in the Liberty Ship production program (Searle & Gody, 1945).
The results shown in Table 3.3 also indicate that learning did not persist—knowl- edge acquired through production depreciated rapidly. With the exception of the model that included calendar time, estimates of lambda, the depreciation parameter, were all significantly less than one. When the model with calendar time was esti- mated using the more general translog specification, the depreciation parameter was significantly less than one.
We explored other potential explanations of our findings. For example, costs of changing the rate of output are often emphasized in discussions of production activ- ities (e.g., Asher, 1956). As noted previously, Lockheed executives frequently men- tioned the difficulties encountered when they increased the rate of production of the L-1011 (“Lockheed losses on TriStar,” 1980). We investigated the importance of these adjustment costs in the shipbuilding program by including variables measur- ing the rate of change in input levels from one period to the next. While there is evidence that adjustment costs were present in the Liberty Ship program, our results on the depreciation of knowledge were unchanged by the inclusion of the adjust- ment-cost variables.
We also investigated whether different results would be obtained with a more general specification of the production function. We performed additional analyses using the more general translog specification (Berndt & Christensen, 1973) that includes (ln H)2, (ln W)2 and ln H ln W in addition to the terms appearing in the Cobb– Douglas production function shown in Table 3.3. We estimated the translog model for all of the models shown in Table 3.3. In all cases except Column 2 of Table 3.3, the additional terms introduced for the translog function were significant (p < .05). Estimates of lambda, the depreciation parameter, in these alternative models ranged from 0.62 to 0.80 and were all significantly less than one. Thus, the results with the more general translog model reinforced the results on knowledge depreciation.
We also allowed for the possibility that the rate of learning slowed down or lev- eled off. Cumulative output is the conventional measure of experience. Using cumu- lative output in logarithmic form, as in Eq. (1.2), implies that unit costs converge to zero as cumulative output increases. It may be that cumulative output is the correct measure but that unit costs converge to a positive number rather than to zero. To investigate whether the rate of learning levels off, we estimated Cobb–Douglas and translog production functions with both ln K and (ln K)2 as predictors. This quadratic function, evaluated at values of K less than the value at which the function reaches a minimum, approximates a function with a positive asymptote—even with no depreciation of learning. The maximum likelihood estimate of lambda was significantly less than one when knowledge was included as a quadratic function, providing further evidence that knowledge acquired through learning by doing depreciated.
We performed additional analyses to deal with the possibility that there might be simultaneity in the choice of inputs and outputs. This simultaneity might occur, for example, if shipyards that were less productive scheduled more labor hours. To deal with this issue, we estimated several of the models shown in Table 3.3 using the nonlinear two-stage least-squares procedure of Amemiya (1974). As instruments, we used current and lagged values of real wages, shipyard dummy variables, time and time squared, lagged endogenous variables (output, shipways, labor hours), and current and lagged exogenous variables. The results from the nonlinear two-stage least-squares procedure were virtually identical to those obtained using ordinary least squares (see Table 3.3).
The results consistently exhibit evidence of economies of scale in shipbuilding. For example, the results in Column 1 of Table 3.3 indicate that an increase in hours and shipways of one percent would result in a 1.31% increase in output, other things constant. The results indicate that an increase in shipways would result in a more than proportionate increase in output, other things constant. When measures of labor hours and shipways were included in the models, the knowledge variable, K, was highly significant and the estimate of lambda, l, that best fits the data was significantly less than one. Thus, the results shown in Table 3.3 indicate that when input effects and economies-of-scale effects are controlled for, strong evidence of learning and knowledge depreciation remain.
1.3. An Application
My colleagues and I used publicly available data to apply our findings on organiza- tional learning and knowledge depreciation to Lockheed’s production of the L-1011 (Argote et al., 1990). Table 3.1 displays the yearly production of the L-1011. As Table 3.1 indicates, the rate of output of the L-1011 varied enormously over time. For a wide range of values of the depreciation parameter, Lockheed’s production rates imply that the knowledge variable for the L-1011 peaked in late 1974 or early 1975 and then declined. A high level of the knowledge variable would be associated with low production costs.
Examining reports of Lockheed’s cost data, we found that they were consistent with our hypothesis of depreciation. Lockheed reported in 1975 that production costs were less than the price at which each plane was sold ($20 million). Thus, production costs were low during the period when the knowledge variable was high. As can be seen from Table 3.1, cuts in production occurred in late 1975. Costs rose to exceed price and appeared to remain above price for the rest of the production program. In 1982, the L-1011 sold for between $29 and $35 million (converted to 1975 dollars for comparison with the earlier period). Thus, the unit cost of produc- tion was less than $20 million in real terms in 1975 when the knowledge variable was at its highest but greater than $29 million in real terms in 1982. These data are consistent with the hypothesis that depreciation occurred at Lockheed.
Benkard (2000) obtained detailed data from Lockheed on the production of the L-1011. Benkard demonstrated very convincingly that depreciation occurred in the L-1011 production program. Indeed, the attractiveness of the model including depreciation is that it explains both the first half of production, which is consistent with the classic learning curve model, and the last half of production, in which costs rose rather than receded with increasing experience.
Benkard (2000) also investigated whether knowledge transferred completely between two different versions of the L-1011 produced by Lockheed. Although there was significant transfer across the two models, it was incomplete. That is, the second model benefited from some—but not all—of the production experience acquired on the first model. Evidence of depreciation remained—even when one allowed for incomplete knowledge transfer across the two models. This finding is very important because it suggests that the depreciation results are not due to prod- uct changes that render previous knowledge obsolete. Evidence of depreciation remained strong when incomplete knowledge transfer across the different models of the product was taken into account.
When we presented our results on depreciation along with follow-up work in aerospace, automotive and service industries to managers, the results seemed to strike a chord with them. Indeed, one manager referred to our results as document- ing a phenomenon he fervently believed in and called “industrial amnesia.”
1.4. Discussion of Causes of Knowledge Depreciation
What causes industrial amnesia? That is, why might knowledge depreciate? Knowledge could decay because products or processes change and thereby render old knowledge obsolete. The incomplete transfer of knowledge observed across the two models of the L-1011 by Benkard (2000) could reflect this obsolescence. Some of the knowledge acquired on the first model may not have been relevant for the second. The depreciation observed in the Lockheed case, however, was not due solely to this obsolescence. Evidence for depreciation remained strong even when incomplete transfer of knowledge across the models was taken into account. Future research aimed at assessing depreciation would benefit from allowing for incom- plete transfer across models, if different models are manufactured. This would enable one to determine if depreciation occurred, while allowing for the possibility that knowledge acquired on one model might not be relevant for another.
Knowledge could also decay because organizational records are lost or become difficult to access. This phenomenon occurred at Steinway piano company. When the firm decided to put a discontinued piano back into production, Steinway dis- covered that it did not have any records or blueprints at its New York facility about how to produce the piano (Lenehan, 1982). Similarly, almost all of the information collected and stored before 1979 by Landsat, an earth surveillance program, is no longer accessible: the data were recorded by equipment that no longer exists or cannot be operated, and the magnetic images have “bled” over time (Marshal, 1989).
In a similar vein, the editors who restored the “Star Wars” trilogy discovered that the original prints of the film had seriously decayed (Morgenstern, 1997). The col- ors seemed wrong; the print looked faded. Although “Star Wars” had originally been shot on four different varieties of film stock, all of the varieties were subject to fading and various color shifts. Attempts had been made to preserve the completed negative of “Star Wars” on a pair of “protective masters.” The preservation effort, however, was not successful. The negative was not cleaned properly before it was copied. The results of the copy attempt were never inspected. According to Morgenstern (1997), this problem is not unique to “Star Wars” but rather character- izes many films made as recently as the 1980s: “Many of our most cherished mod- ern movies…are already in deep decay and could be lost to theatrical audiences forever” (p. A16). These examples illustrate that organizational knowledge can exhibit decay: having knowledge at one point in time does not guarantee that the organization will have it in the future.
Another possible cause of knowledge depreciation is member turnover. Organizational members can leave and take their knowledge with them. Turnover not only deprives the organization of the knowledge and skills of the departing member, turnover also disrupts the performance of members who were interdepen- dent with the departing member and had developed relationship-specific assets. In a study of ambulance companies, David and Brachet (2011) found evidence of orga- nizational forgetting. Further, the researchers distinguished the effect of member turnover from the effect of skill decay of individual members caused by inactivity or task interference. Results indicated that the contribution of member turnover to organizational forgetting was about twice the effect of skill decay. Studies of the effect of turnover on organizational learning and forgetting are discussed in Chap. 4, on organizational memory.
A question we examined in follow-up work was whether the departure of certain key people would affect organizational performance. Perhaps the departure of exceptional performers would affect organizational outcomes. Or perhaps the depar- ture of gatekeepers (Allen, 1977) who bridge social networks or of individuals who occupy key positions in an organization’s social network (Burt, 1992; Krackhardt & Hanson, 1993) would have more of an effect on organizational performance than the departure of individuals not occupying key structural positions (Shaw, Duffy, Johnson, & Lockhardt, 2005). Whether the effect of turnover on organizational learning and forgetting depends on who turns over is a focus of our second study of knowledge depreciation. This study is now described.
2. The Automotive Study
A major goal of our second study, which was conducted in the automotive industry, was to examine in more depth the role of labor turnover in the acquisition and depre- ciation of organizational knowledge. Although little empirical evidence existed about factors responsible for learning and depreciation in organizations, most discussions of organizational learning curves included individual learning as an important source of the gains in organizational productivity observed with increasing experience (Hayes & Wheelwright, 1984; Hirsch, 1952; Wright, 1936). Similarly, in their theoretical discussion of organizational learning, Walsh and Ungson (1991) pointed out that indi- viduals can act as “retention facilities” for organizational memory. To the extent that knowledge acquired through learning by doing is embedded in individuals, their turn- over would be harmful for organizational learning. Similarly, Huber (1991) and Simon (1991) suggested that turnover would be harmful for organizational memory.
2.1. Research on Consequences of Turnover
Although there has been a long tradition of research on predictors of turnover, research devoted to determining the consequences of turnover is more recent (Dalton & Todor, 1979; Mobley, 1982; Mowday, Porter, & Steers, 1982; Staw, 1980). The automotive study analyzed whether the effect of turnover depended on who departed. Price (1977) suggested that the effect of turnover on organizational effectiveness depends on the performance levels of departing members. Similarly, Mowday et al. (1982) hypothesized that characteristics of individuals leaving the organization moderate the effect of turnover on organizational performance. Boudreau and Berger (1985) developed a multivariate decision-theoretic utility model for assess- ing the consequences of employee movement into and out of an organization.
Empirical studies of the performance levels of departing employees have yielded mixed results. Although Price (1977) concluded that leavers are relatively more often high-performing employees than those remaining in the organization, Dalton, Krackhardt, and Porter (1981), Dreher (1982), and Wells and Muchinsky (1985) all reported that the performance of employees who left an organization was significantly lower than the performance of those who remained. In their reviews of this litera- ture, Jacovsky (1984) and McEvoy and Cascio (1987) concluded that turnover was higher among poor than among good performers. This relationship held for both voluntary and involuntary turnover but, of course, was stronger for involuntary turn- over (McEvoy & Cascio, 1987).
Schwab (1991) suggested that whether it is the good or the poor performers who leave depends on several contingencies. Schwab (1991) found that the relationship between individual performance and turnover was positive for tenured faculty mem- bers and negative for untenured faculty. That is, for tenured faculty, higher perform- ers were more likely to leave, whereas for untenured faculty the reverse was true. Schwab (1991) concluded that the relationship between individual performance and turnover depends on several contingencies, such as whether performance is exter- nally visible, and whether there are external job opportunities. The mixed pattern of results on the relationship between individual performance and turnover suggests that it is difficult to predict a priori whether turnover is higher among good or among poor performers because the relationship depends on important contextual factors.
Nonetheless, whether it is the good or the bad performers who are leaving is likely to have important consequences for organizational outcomes. If it is the poor performers who are leaving an organization, we would not expect turnover to have a negative effect on organizational performance. Conversely, if it is the good per- formers who are departing, we would expect turnover to have a negative effect on performance. Thus, the performance of who is turning over should be taken into account when predicting the effect of turnover on organizational learning and pro- ductivity gains.
The automotive study examined empirically the role of turnover in organiza- tional learning curves. We investigated whether the effect of turnover depended on the performance of those who departed. We also investigated the effect of move- ment of employees into the plant. Our expectation was that movement of employees into the plant at moderate levels would have a positive effect on productivity. March (1991) found a non-monotonic inverted-U relationship between the number of indi- viduals moving into an organization and its performance in a simulation. New employees may bring new ideas and new skills and be more highly motivated than employees with longer tenure (Abelson & Baysinger, 1984; Mowday et al., 1982; Staw, 1980). At some point, however, the cost of integrating too many new employ- ees becomes disruptive for organizational performance. Thus, we expected a non- monotonic inverted-U shaped relationship between movement of individuals into the plant and productivity gains.
2.2. Method and Sources of Data
We collected data from a North American truck plant (Argote, Epple, Rao, & Murphy, 1997). The workforce at the plant, which was unionized, numbered approx- imately 3,000. The technology at the plant was extremely advanced. We collected weekly data over a 2-year period from the start of production at the plant. Our data included measures of the number of trucks produced, total direct labor hours worked, number of shifts worked, and movement of employees into and out of the plant.
The turnover data were disaggregated according to the various reasons employ- ees left the plant. Based on theoretical reasons, the effects of two types of turnover were investigated: turnover of high-performing employees who left the plant because they were promoted (promotion) and turnover of employees who were discharged for poor performance (discharge). A third turnover variable included all other reasons employees departed that were not a function of performance (e.g., retired, deceased, quit, laid off, and so on). We performed sensitivity analyses which disaggregated these other types of turnover and investigated their separate effects as well.
We took the same general approach to estimation described for the shipyards study. Our general approach was to estimate production functions. Because we ana- lyzed data from only one plant and its physical facilities were relatively unchanged throughout the course of the study, there was no need to control for physical facilities.
The same approach to assessing depreciation was used for the automotive study as for the shipyards study. As in the previous study, if λ = 1, the accumulated stock of knowledge equaled cumulative output, the measure of experience in the conven- tional learning curve formulation. Thus, if λ = 1, there was no evidence of deprecia- tion. If λ < 1, evidence of depreciation existed because past output received less weight than recent output in predicting current productivity.
We first estimated the classic learning curve. In this analysis, the value of l, the depreciation parameter, was constrained to equal one. Thus, this model is the con- ventional learning curve that assumes knowledge is cumulative and persists through time. The results from estimating this model provided strong evidence of learning at the automotive plant: production increased significantly with rising cumulative out- put. The progress ratio derived from the estimate of the learning rate in this study was 83%. Thus, each doubling of cumulative output at the plant led to a 17% reduc- tion in unit cost. Results also indicated that there were constant returns to labor hours and that output went up proportionately with the number of shifts worked.
We then estimated a model that did not constrain the depreciation parameter to equal one. The value of l, 0.989, obtained from estimating this model, was significantly less than one. In subsequent analyses, we also included time as an explanatory variable to investigate the extent to which technical progress associated with the passage of time was responsible for productivity gains. There was evidence that the plant became more productive as time passed. The experience variable remained highly significant when time was included as an explanatory variable. Further, the depreciation parameter was significantly less than one in this analysis that included the time variable. Indeed, the estimated value of the depreciation parameter was lower for this model than for the previous ones.
These results have interesting implications for organizational memory. The results suggest that there is a relatively permanent component to organizational memory as well as a more transitory component. The permanent component, which the time variable is picking up, could correspond to knowledge embedded in the organization’s procedures and routines. This permanent or procedural component does not evidence depreciation. The more transitory component of organizational memory is reflected in the faster depreciation rate found in the model that accounts for the permanent component of organizational knowledge. This transitory compo- nent may be analogous to declarative knowledge, or knowledge of facts (Singley & Anderson, 1989). Cohen and Bacdayan (1994) extended the distinction between procedural and declarative knowledge made in analyses of individual cognition to the dyadic level of analysis. In an interesting laboratory study, Cohen and Bacdayan (1994) found that procedural knowledge exhibited less forgetting than declarative knowledge. Our results from the field suggest that there may be both a permanent and a transitory component to organizational memory.
Another major focus of the automotive study was analyzing the effect of person- nel movement into the plant on productivity. We found an inverted-U relationship between the number of new hires moving into the plant and the plant’s productivity. Further analyses revealed that the maximum of this function was reached at 38 per- sons per week. Thus, increases in the number of new hires moving into the plant up to approximately 38 persons (between 1 and 2% of the workforce) per week were associated with increases in productivity. Beyond that point, decreases in productiv- ity were observed.
To investigate our hypothesis that the effect of turnover depended on the perfor- mance of the departing employees, additional analyses were performed in which the turnover variable was disaggregated as a function of the reason employees departed. In these analyses, the coefficient of the variable representing employees who were promoted out of the plant as a result of their good performance was negative and significant, as predicted. Thus, the turnover of these high-performing employees negatively affected the organization’s productivity. The coefficient of the variable representing employees who were discharged for poor performance was positive, consistent with the expectation that their removal would improve organizational performance. The discharge variable, however, was not consistently significant. The variable representing other types of turnover did not approach significance.
Sensitivity analyses were performed in which the types of turnover included in the “other” category were disaggregated and their effects estimated, either sepa- rately or in combination with other types of turnover. With the exception of the “lay-off” variable, these other types of turnover were not significantly related to productivity. In a few analyses, the coefficient of the lay-off variable was negative and marginally significant. In a few regressions, the positive coefficient of the dis- charge variable was significant. These effects, however, were not consistently significant. Additional analyses were performed to investigate whether the rate of learning plateaued or slowed down over time. We investigated this possibility by including a quadratic term, the square of the knowledge variable. The coefficient of the square of the knowledge variable was extremely small in magnitude and did not approach statistical significance, suggesting that the rate of learning did not change in this production environment.
Because some product options may require more labor content than others, we also investigated the effect of product mix on productivity by including a variable repre- senting different product options in the model. The coefficient of the product mix variable was small in magnitude and statistically insignificant. Including the product mix variable did not affect the coefficients of the other variables in the model.
The results obtained from these additional analyses reinforced our previous con- clusions regarding learning, depreciation, and the role of turnover. The evidence for learning remained strong. The depreciation parameter was significantly less than one. Movement of new employees into the plant at moderate levels was consistently shown to help productivity. Turnover of high-performing employees due to promo- tions appeared to hurt the plant’s productivity.
3. The Franchise Study
We also investigated whether knowledge depreciated in a study of fast food fran- chises (Darr, Argote, & Epple, 1995). Because our primary goal in the franchise study was to investigate the transfer of knowledge across the various stores, this study is described in depth in Chap. 6. We also investigated whether knowledge depreciated in the study because knowledge that depreciates rapidly can be difficult to transfer.
Our analyses of learning in the fast food franchises were based on weekly data. Estimates of the depreciation parameter obtained in the fast food study ranged from 0.80 to 0.83 (Darr et al., 1995). This is an incredibly rapid rate of depreciation. A value of the depreciation parameter equal to 0.83 implies that roughly one half (l4 = 0.844) of the stock of knowledge available at the beginning of a month would remain at the end of the month. From a stock of knowledge available at the beginning of a year, a very negligible amount (l52 = 0.8452) would remain 1 year later. Without continuing production to replenish the stock of knowledge, virtually all production knowledge would be lost by mid-year. The rate of depreciation found in the fast food study is the most rapid we have found.
4. Other Analyses of Depreciation
In addition to our work and that of Benkard (2000), several other studies have inves- tigated knowledge depreciation. Thompson (2007) constructed a detailed data set about Liberty ship production from primary sources at the National Archives. Analyses of this data set, which included measures of product mix, resulted in more modest estimates of the rate of depreciation than we found. Further, Thompson (2007) found that labor turnover appeared to explain a significant amount of the depreciation. Kim and Seo (2009) analyzed the productivity of the shipyard that produced the largest number of Liberty ships. Based on a different model from ours that used the elapsed time between units to explain knowledge depreciation, Kim and Seo (2009) found rapid rates of depreciation similar to our estimates: only about three quarters of the knowledge available at the beginning of a month remained at its end. Thus, all three studies of the Liberty ship production program have found evidence of depreciation; the amount estimated is sensitive to model specification.
In a study of learning by and between firms in contractual relationships, Kellogg (2011) found significant evidence of knowledge depreciation. Further, Kellogg (2011) suggested that the depreciation was due to the loss of relationship-specific capital between firms.
A couple of studies contrasted the rate of depreciation for different types of expe- rience. In an analysis of accident experience in coal mines, Madsen (2009) found that the effect of experience from major accidents in which lives were lost decayed at a slower rate than the effect of minor accident experience. Similarly, in a study of orbital launches, Madsen and Desai (2010) found that failure experience decayed more slowly than success experience.
Rather than estimate the extent of knowledge depreciation using the method described here, several studies have used various discount factors which assume varying patterns of knowledge depreciation. These studies typically find evidence of some depreciation (e.g., see Ingram & Baum, 1997). An exception is Ingram and Simons (2002) who did not find evidence of depreciation in their analysis of kibbutz agriculture. Ingram and Simons suggested that the persistence of knowledge observed in kibbutzim may be due to the very stable and motivated membership of those organizations.
Source: Argote Linda (2013), Organizational Learning: Creating, Retaining and Transferring Knowledge, Springer; 2nd ed. 2013 edition.