The Acquisition, Transfer, and Depreciation of Knowledge in Service Organizations: Productivity in Franchises Author(s): Eric D.

Darr, Linda Argote, Dennis Epple Source: Management Science, Vol. 41, No. 11, Focused Issue on Service Management (Nov., 1995), pp. 1750-1762 Published by: INFORMS Stable URL: http://www.jstor.org/stable/2632871 . Accessed: 01/02/2011 14:02
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=informs. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.

http://www.jstor.org

The of

Acquisition, Knowledge in

Transfer, Service

and

Depreciation

Organizations: Franchises

Productivity in

Eric D. Darr * Linda Argote * Dennis Epple

The Anderson GraduateSchool of Management, UCLA, 405 Hilgard Avenue, Los Angeles, California 90024-1481 GraduateSchool of Industrial Administration, CarnegieMellon University, Schenley Park, Pittsburgh, Pennsylvania 15213 GraduateSchool of Industrial Administration, CarnegieMellon University, Schenley Park, Pittsburgh, Pennsylvania 15213

inter orga

INTER ORGANNISATIONNELEL DANS CET ARTICLE

paper examines the acquisition, depreciation and transfer of knowledge acquired through learning by doing in service organizations. The analysis is based on weekly data collected over a one and a half year period from 36 pizza stores located in Southwestern Pennsylvania. The 36 stores, which are franchised from the same corporation, are owned by 10 different franchisees. We find evidence of learning-in these service organizations: as the organizations gain experience in production, the unit cost of production declines significantly. Knowledge acquired through learning by doing is found to depreciate rapidly in these organizations. Knowledge is found to transfer across stores owned by the same franchisee but not across stores owned by different franchisees. Theoretical and practical implications of the work are discussed. (Organizational Learning;LearningCurves; Productivity; KnowledgeTransfer)

The

1. Introduction
As organizations produce more of a product, the unit cost of production typically decreases at a decreasing rate. This phenomenon or close variants of it is referred to as a learning curve, a progress curve, an experience curve, or learning by doing. "Learning curves" have been found in many organizations, including those producing aircraft, ships, trucks, and refined petroleum products (Argote and Epple 1990). Reviews of the literature on organizational learning curves can be found in Argote (1993), Dutton and Thomas (1984), and Yelle (1979). Most previous research on learning curves has focused on manufacturing organizations. The current study extends this work by examining learning in service organizations. More specifically, we examine productivity gains in 36 pizza stores owned by 10 franchisees in Southwestern Pennsylvania. The current study also extends previous work on organizational learning by analyzing the transfer of

learning in service organizations. By transfer, we mean whether organizations learn from the experience of other organizations. Levitt and March (1988) and Huber (1991) have suggested that organizations learn from the experience of others as well as from their own direct experience. We empirically examine whether knowledge transfers across organizations by analyzing whether stores benefit from production experience at other stores. The study of learning transfer has important implications for firms planning for the start-up of multiple facilities, for competitive strategy, for antitrust policy (Spence 1981), for trade policy (Gruenspecht 1988), for the success of joint ventures (Kogut 1988), and for explaining interfirm and international differences in rates of learning (Mody 1989). A further contribution of the current study is its analysis of the depreciation of organizational knowledge. The classic learning curve formulation (e.g., see Yelle 1979) assumes that learning is cumulative and that it
0025-1909/95/411 1/1750$01.25
Copyright ?) 1995, Institute for Operations Research and the Management Sciences

1750

MANAGEMENT SCIENCE/VO1.41, No.

11, November 1995

DARR, ARGOTE, AND EPPLE Productivity in Franichises

persists through time. More recent work suggests, however, that knowledge acquired through learning by doing may depreciate (Argote et al. 1990). The current study empirically estimates whether depreciation occurs in these service organizations. Information about whether knowledge depreciates has important implications for forecasting production quality, costs and rates. Failure to allow for depreciation of learning may result in forecasts with large errors (e.g., see Argote et al. 1990 analysis of Lockheed's production of the L-1011 as reported in the Wall Street Journal, 1980-1981). Additionally, the extent of knowledge depreciation has implications for developing human resource policies concerning personnel retention in general and downsizing in particular. Information about depreciation also has implications for competitive strategy (Argote et al. 1990) and for explaining interfirm differences in rates of productivity gains (Argote and Epple 1990). In the sections that follow, we review empirical evidence on the transfer and depreciation of organizational learning. Two examples of knowledge transfer that occurred in the organizations we studied are discussed. We also review empirical evidence on the effectiveness of various transfer mechanisms to develop our research hypotheses.

firm-specific experience, architect/ engineer experience, and industry experience on the unit cost of construction. Results indicated that only firm-specific and architect/ engineer experience accounted for significant portions of variance in the unit cost of construction. Argote et al. (1990) found that shipyards which began production later were more productive initially than shipyards with earlier start dates. Once shipyards began production, however, they did not benefit from production experience at other yards. Epple et al. (1991) and Epple et al. (in press) analyzed transfer across shifts within two manufacturing facilities. The researchers found evidence of transfer at both sites. There were differences across sites, however, in the extent to which knowledge acquired on the first shift transferred to the second. Evidence of experience transfer has also been found for angioplasty surgery success rates at different hospitals (Kelsey et al. 1984). The researchers used calendar time as a proxy variable for technical progress in the environment. The researchers found that success rates improved with calendar time for only the first 20 or so operations performed by a surgeon. Thus, transfer of learning appeared to influence the early but not later performance of surgeons.

1.1. Empirical Evidence on Transfer of Learning Several researchers have empirically examined the transfer of organizational learning. Their collective results indicate that knowledge transfer is selective. Zimmerman (1982) examined the transfer of construction knowledge relating to 10 nuclear reactors built over a ten-year period. He analyzed the effects of firm experience and industry experience on the unit cost of construction. Results indicated that transfer of learning occurred: the industry experience variable accounted for a significant portion of variance in plant cost. Thus, learning accrued to individual firms as a result of industry-wide experience. Firm-specific experience, however, was more significant than industry experience. Alternatively, Joskow and Rose (1985) found no evidence of industry experience transfer for 411 coalburning steam-electric generating units built between 1960 and 1980. The researchers analyzed the effects of

1.2. Qualitative Evidence on Transfer of Learning in the Production of Pizzas Two events that occurred in the stores we studied illustrate that transfer of organizational knowledge is selective. The first incident concerned the development and transfer of an innovation for placing pepperoni. The usual procedure for placing pepperoni on a pizza is to distribute it evenly over the entire pie. When this procedure was used on pan pizzas, which are thickcrusted, it often resulted in finished pizzas with mounds of pepperoni pooled in the center. One solution to this problem was to monitor and modify pepperoni placement during baking. This was a difficult and timeconsuming task. Another solution to the problem was developed that involved a different initial method of placing pepperoni. Rather than distributing the pepperoni equally over the entire pie, pepperoni is placed in spokes around the pie. As the pan pizza cooks and the cheese flows, the

MANAGEMENT SCIENCE/VOl.

41, No. 11, November 1995

1751

DARR, ARGOTE, AND EPPLE Productivity in Franichises

pepperoni fans out across the pie, resulting in finished pan pizzas with equally distributed pepperoni. The innovation was discovered at a franchise store in western Pennsylvania. Initially, it transferred only to other stores owned by the same franchisee. The advantages of the placement innovation soon became evident to a visiting franchisee who carried it back to his franchise organization. The adopting franchisee was so impressed with the consistent results of the placement procedure, he recommended it at the next quarterly meeting of all western Pennsylvania franchisees. The placement procedure was soon in use at all stores within those franchise organizations. A consultant from the parent corporation visited western Pennsylvania at about this time, and was impressed with the pepperoni placement procedure. Within a year, the procedure was in use at 90% of the stores nation-wide. The pepperoni placement procedure for pan pizzas started as a storelevel innovation and eventually produced corporatewide benefits through extensive transfer of learning. The second incident concerned the layout for the workflow. The final step in the pizza production process involves placing a finished pizza pie in an appropriately sized box. Traditionally, the phone operator takes the order information (e.g., size and type of pizza, address) and records it on a box label. The labeled boxes are then arranged vertically on a shelf near the pizza oven. The vertical arrangement saves space, but it forces the pizza maker to read each label sideways and also means that the pizza maker has to open each box prior to placing a pie inside. While opening a box may not seem like a difficult or time-consuming task, it becomes more difficult while balancing a hot pizza in one hand. More than one pizza has been dropped on the floor while moving a pizza from the oven and opening a box at the same time. A better boxing arrangement was discovered that involved placing opened boxes horizontally on a large table near the pizza oven. This arrangement allows a pizza maker to read the label in its natural position and to move a finished pizza directly from the oven into the box. The new boxing arrangement saves time and reduces waste from dropped pizzas. The boxing innovation was discovered at a franchise store in western Pennsylvania. It transferred to other stores owned by the same franchisee but not to stores

owned by different franchisees. Thus, not all innovations produce benefits outside the store or franchise of origin. 1.3. Organizational Relationships and Transfer Mechanisms A potentially important factor that may explain whether knowledge transfers across organizations is the relationship that exists between the organizations involved in the transfer. Transfer of knowledge can occur between independent organizations or between subdivisions of a single organization. In our sample, stores owned by different franchisees are conceptualized as independent organizations, whereas stores owned by the same franchisee are seen as subdivisions of a single organization. Tushman and his colleagues theorized that interorganizational relationships between independent organizations differ from relationships between organizations owned by the same firm (e.g., Tushman 1977, Tichy et al. 1979). The researchers demonstrated that the extent of social networks and commonality of language were greater between subdivisions of a single organization than between independent organizations. Tushman (1977) conceptualized social networks to be regular communication and personal acquaintances. Regular communication and personal acquaintances have both been proposed as mechanisms for transfer of knowledge. This suggests that the process and rate of knowledge transfer between organizations owned by the same franchisee and those owned by different franchisees may differ. Two bodies of literature are related to the rate of knowledge transfer: the adoption/ diffusion of innovations literature and the technology transfer literature. Both literatures focus on the role of transfer mechanisms, conduits or agents through which transfer of knowledge takes place, in facilitating the knowledge transfer process. The transfer mechanisms particularly relevant for the present study are regular communication, personal acquaintances, and meetings. In general, the literatures suggest that high levels of transfer mechanism use are associated with high levels of technology transfer. Regular Communication. Regular communication refers to exchanges of information which occur at stan-

1752

SCIENCE/VOl. 41, No. 11, November MANAGEMENT

1995

DARR, ARGOTE, AND EPPLE Productivity in Franichises

dard and / or frequent intervals through status reports, phone calls and the like. Ghoshal and Bartlett (1988) found that diffusion of innovation was facilitated by high levels of communication. Regular communication has also been found to facilitate the transfer of technology in the textile machinery sector (Rothwell 1978). Transfer of knowledge through regular communication is more likely between commonly owned stores than between differently owned stores since franchisees communicate frequently with their own stores and collect weekly cost and profit data. On the other hand, franchisees in this sample are not required to report any information to other franchisees, but rather only to the corporation. Thus, transfer of learning through regular communication between stores owned by different franchisees is less likely. Personal Acquaintances. Huberman (1983) found that prior personal acquaintances served as a mechanism for transferringknowledge about organizational policies between universities and high school districts. Martilla (1971) found that interpersonal relationships between members of different organizations were a significant determinant of adoption decisions. Liebenz (1982) demonstrated that personal acquaintances were positively related to the ease of technology transfer between United States and Eastern European companies. Interfirm interpersonal relationships have also been shown to influence organizational decisions concerning adoption of philanthropic contribution activity (Galaskiewicz and Wasserman 1989). Because stores owned by the same franchisee have greater geographic proximity and more meetings, they are likely to have more personal ties than are stores owned by different franchisees. Thus, transfer of learning fostered by personal acquaintances is likely to be greater between stores owned by the same franchisee than between stores owned by different franchisees. Dutton and Starbuck (1978) demonMeetings. strated that face-to-face meetings and conferences were significantly related to the degree of diffusion of highway-related computer simulation technology. The franchise organizations in western Pennsylvania meet only once a month. On the other hand, meetings between the stores of a common franchisee occur at least once a week. This suggests that transfer of learning fostered by meetings is likely to be greater between stores

owned by the same franchisee than between stores owned by different franchisees. 1.4. Empirical Evidence on Depreciation of Learning A second potentially important factor affecting the extent of knowledge transfer is the rate of knowledge depreciation. Knowledge acquired through learning by doing may not persist indefinitely. Rather, knowledge may be lost through individual forgetting, misplaced manuals, personnel turnover, and the like. Examining the effect of a strike on the performance of an organization, Hirsch (1952) found that the average decline in labor requirements immediately after the strike was less than the prestrike average, but that the prestrike pattern was eventually restored. Similarly, Baloff (1970) found that when performance was resumed at an organization after an interruption, it was inferior to the level achieved prior to the interruption. Argote et al. (1990) estimated the amount of depreciation that occurred in 13 shipyards. The researchers used a geometric weighting of past output to allow for the possibility of knowledge depreciation. They found that recent production experience was a more important predictor of current productivity than past experience. Thus, evidence was found that knowledge depreciated. In the extreme very rapid depreciation may eliminate a stock of knowledge before it can be transferred. Even less extreme depreciation rates may reduce a stock of knowledge sufficiently to restrict transfer. Potential transfer partners may view a reduced stock of knowledge as less desirable because it may contain fewer novel ideas or incomplete ideas. In light of the potentially important impact of depreciation on knowledge transfer, we investigate the rate of transfer while allowing for depreciation. 1.5. Research Hypotheses Based on previous empirical research concerning organizational learning, we expect to find evidence of learning in service organizations. That is, we expect that as each store gains experience in production, productivity will improve. We also expect to find evidence of depreciation in these stores. And we expect to find evidence of transfer. We expect transfer to be greater across stores owned by the same franchisee than across stores owned by different franchisees and also that there will

MANAGEMENT SCIENCE/Vol.

41, No. 11, November 1995

1753

DARR, ARGOTE, AND EPPLE Productivity in Franchises

be more use of mechanisms for transferring knowledge across stores that belong to the same franchisee.

2. Data and Method

The conventional form of the learning curve is: y = ax-b where

(1)

y = the cost/ unit to produce the xth unit, a = the cost of producing the first unit, x = the cumulative number of units produced, and b = parameter measuring the rate costs are reduced as cumulative output increases For estimation purposes, the learning curve can be rewritten: log y = log a - b log x. (2)

average age of the franchise organizations was 3.75 years. The corporation's regional office provided data concerning pizzas sold and production costs for each store by week from January 1, 1989 through June 15, 1990. Structured interviews with the franchisees provided information on the frequency that phone calls, meetings, and personal acquaintances were used to transfer information. The data are from a very desirable situation. The inputs (i.e., the raw materials) are homogeneous. Therefore, input characteristics are controlled for naturally in the sample. Differences in technology across pizza stores are very small. Product mix and economies of scale will be controlled for in the analyses. 2.2. Analysis Plan The symbols used throughout the paper and the variables they represent are listed below. t-calendar time in weeks, J,1-number of stores in franchise n, qniit-pizzas produced by franchisee n in store i in week t, C,1it-costs (food and labor) for store i in franchisee n in week t, Qlit= s=Oq,1is-cumulative number of pizzas produced by store i in franchisee n through week t, = FQ,1t {=1 Q,,it-cumulative number of pizzas produced by franchisee n through week t, IQt = E 1?=1 FQ,,t-cumulative number of pizzas produced in all stores in all franchisees through week t, p,it-percentage of pan pizzas produced by franchisee n in store i in week t, and sni-dummy variables for each store. The variable Q is a proxy for store-specific knowledge. The variable FQ is a proxy for franchisee knowledge and IQ is a proxy for interfranchisee knowledge. The dummy variables capture variance associated with store specifics such as management style, age, and location. Several models are estimated in which the unit cost of production depends on store-specific experience, franchisee experience, interfranchisee experience, and other variables. The most basic model we estimated using least-squares regression (Column 1 in Table 1) was: + b2 Log FQ,1t-l Log(c,1jt/q,jt ) = bo + bi Log Q,1jt_j + b3 Log IQt-l + b,,is,,i + U,1it. (3)

In the above expression, the cumulative number of units produced is a proxy variable for knowledge acquired through production. If unit costs decrease as a function of this knowledge (i.e., the coefficient of cumulative output is statistically significant), other things held constant, organizational learning is said to occur. The learning curve format provides a method for evaluating organizational learning and its transfer (Argote and Epple 1990). Store-specific learning, intrafranchise transfer of learning and inter-franchise transfer of learning are of interest in this research. Storespecific knowledge may be measured by cumulating store pizza production through time. Franchisee knowledge may be measured by aggregating the cumulative output across all stores owned by a common franchisee. Interfranchise knowledge may then be measured by aggregating the cumulative output across all franchisees in our sample. 2.1. Source of Data Data for this research were collected from the entire set of stores in southwestern Pennsylvania that are franchised from one of the largest pizza corporations. The sample centered around the Pittsburgh area, and included 10 different franchisees who owned a total of 36 stores. The largest franchisee owned 11 stores, whereas five of the franchisees were single store owners. The oldest franchise organization had been in business for 11 years, and the youngest for just 3 months. The

1754

MANAGEMENT SCIENCE/VOl. 41, No. 11, November

1995

DARR, ARGOTE, AND EPPLE Productivity in Franchises

We allow for serial correlation of the error term in all equations we estimate. In Equation (3), if b1 is significant, store-specific learning has occurred. If b2 is significant, transfer of learning between stores owned by a common franchisee has occurred. If b3 is significant, transfer of learning between stores owned by different franchisees has occurred. In these analyses, the unit of time is a week. The variables Q, FQ and IQ are the cumulative pizza production through the end of the previous week. The lagged cumulative output is used on the right-hand side of Equation (3) because cumulative output serves as a proxy for experience acquired as a result of past output. Alternative explanations of our findings are investigated by estimating models with additional variables to represent possible technological change associated with the passage of time and economies of scale. The next model we estimated (Column 2 in Table 1) was:

Log( c,it/qnit ) = bo + bi Log

Qnit-1

+ b2 Log FQnt1

+ b3 Log IQt-l + b4t + b5qnit + b6q + Unit nit + bniSni (4)

We next estimated models with variables representing changes in the rate of learning and product mix (percentage pan pizzas). The third model we estimated (Column 3 in Table 1) was:
Log(Cnit/qnit =

bo+ b, Log Qnit-1 + b2 Log FQnt-1

+ b3 Log IQt-l + b4t + b5qnit
+ b6qnit2 + b7[Log(Qnit1)]2

+ b8pnit + bniSni + unit-

(5)

We also investigate whether knowledge persists through time or whether it depreciates by replacing cumulative output with the following knowledge variable

Table 1

EstimatedCoefficientsfor Models PredictingUnitCosta (1) (2) -0.098t (0.020) -0.066t (0.019) -0.008 (0.010) 0.003t (0.001) -0.0003t (0.1E-04) 0.6E-07t (0.4E-08) (3) -0.097t (0.020) -0.064t (0.020) -0.009 (0.010) 0.003t (0.001) -0.0003t (0.1E-04) 0.5E-07t (0.4E-08) 0.009 (0.007) 0.021 (0.017) (4) -0.104t (0.019) -0.059t (0.022) -0.004 (0.010) 0.004t (0.002) -0.0003t (0.1E-04) 0.5E-07t (0.4E-08) 0.009 (0.008) 0.022 (0.021) 0.80t (0.046) 0.512t (0.022) 0.593 (5) -0.106t (0.022) -0.094* (0.047) -0.001 (0.011) 0.002t (0.0008) -0.0004t (0.9E-05) 0.9E-07t (0.4E-08) 0.003 (0.009) 0.052 (0.048) 0.83t (0.042) 0.492t (0.024) 0.653

Store-specificLearning (b1) Transferbetweencommonlyowned stores (b2) Transferbetweendifferently owned stores (b3) Calendar Time(b4) Pizza Count(b5) Current Pizza Count(b6) Squareof Current Squareof Store-SpecificLearning (b7) PercentagePan Pizza (b8) of Knowledge Depreciation (X) Autocorrelation Coefficient R2
a

-0.117t (0.019) -0.104t (0.016) -0.015* (0.007)

0.569t (0.017) 0.237

0.5814 (0.014) 0.557

0.589t (0.015) 0.565

Standarderrorsare shown in parentheses. p < 0.05, and tP < 0.01, and tp < 0.001.

MANAGEMENT SCIENCE/VOI.

41, No. 11, November 1995

1755

DARR, ARGOTE, AND EPPLE Productivity in Franchises

in the next models (Columns 4 and 5 in Table 1):

Knit = XKnit-1 + qnit

Figure1

Number RelationBetweenTotalCostper Pizzaand Cumulative of Pizzas Produced

(6)

Equation 6 allows for the possibility that knowledge depreciates over time by the inclusion of the parameter X.If X = 1, the accumulated stock of knowledge is simply equal to lagged cumulative output, the conventional measure of learning, and there is no evidence of depreciation. If X < 1, there is evidence of depreciation: recent output is a more important predictor of current productivity than past output. An iterative search algorithm was used to estimate the parameter X and all other parameters to minimize the sum of squared residuals. The fourth model we estimated (Column 4 in Table 1) was:
Log(cnit / qnit) = bo +

A S A AA A AA A A jh A A ALAL A A AA A

b, Log Knit-1 + b2 Log FKnt-1

Cumulative PizzasProduced
Note: Thesedataarefroma singlestorefora period of 11 years.

+ b3 Log IKt-, + b4t + b5qnit + b6q nit + b7[Log(Knit-)]2

+ b8pnit + bniSni + Unit (7)

3. Results
A learning curve plotted from a single store is shown in Figure 1. This figure shows the characteristic learning curve pattern: the unit cost of producing pizza decreased at a decreasing rate as the cumulative number of pizzas produced increased. 3.1. Store-specific Learning and Transfer of Learning Effects Results concerning the effects of store-specific learning, transfer between commonly owned stores and transfer between differently owned stores on cost per unit' are presented in Table 1. Results of estimating Equation (3) using a maximum-likelihood estimation algorithm allowing for first-order autocorrelation of the residuals are presented in column 1 of Table 1.2 Column 1 shows
1 The constant term and the coefficients of the store-specific dummy variables are not of particular interest so are not reported to preserve the confidentiality of the data. A joint test of the null hypothesis that there are no store-specific effects is rejected at a high significance level (p < 0.001), so important store-specific effects appear to be present in the sample. A regression with just the store-specific dummy variables explained roughly half the variance (0.328) explained by our full model. Store-specific dummy variables are included in all analyses.
2 The presence of the autocorrelation coefficient, p, makes the model nonlinear in the parameters. Other parameters that we add below

where FKand IK are defined analogously to FQ and IQ except that K replaces Q in the summations. Thus, knowledge that transfers is allowed to depreciate. The most complex model we estimated included unknown production histories in the knowledge variables K, FK and IK. The fifth model we estimated (Column 5 in Table 1) was:
Log(Cnit/ qnit) =

bo + b, Log(vni +
FKnt-1)

Zni + Knit-i)

+ b2 Log(vn + Zn+

+ b3 Log(v + z + IKt-1)

+ b4t + b5qnit+ b6qnit2 + b7[Log(vni + Zni + Knit-1)]2

+ b8pnit+

bniSni + Unit,

(8)

where for store i in franchise n, Vni and Zniare respectively the known and unknown production histories prior to the first observation in our sample. Vn and Zn are obtained by aggregating the preceding variables across all stores in a franchise, and v and z are aggregates of these variables across all stores in the sample. When production history is known, the unknown production history Zni= 0. Similarly, when production history is 0 unknown, the known production history Vni O.

1756

MANAGEMENT SCIENCE/VOL41, No. 11, November

1995

DARR, ARGOTE, AND EPPLE Productivity in Franchises

the effect of the conventional measure of store-specific learning, lagged cumulative output for each store, on cost per unit. As can be seen from the table, the variable representing store-specific learning has a significant negative coefficient, supporting our hypothesis that the unit cost of production would decrease as the cumulative number of pizzas produced at each store increased. Learning curves are often characterized in terms of a progress ratio, p. The progress ratio, p, is related to the coefficient for store-specific learning, b1, as follows: p = 2b. (9)

While there is considerable variance in progress ratios found in different studies, the modal progress ratio for manufacturing firms is approximately 80% (Dutton and Thomas 1984). Thus, for each doubling of cumulative output a 20% reduction in unit cost is realized. Based on the results shown in column 1 of Table 1, a progress ratio for the entire sample was calculated to be p = 0.929. For each doubling of cumulative output, the unit cost of producing a pizza decreased approximately 7%. Thus, pizza stores in the sample demonstrated a much slower learning rate than the modal "80% learning curve" found in manufacturing firms. The effects of transfer between commonly owned stores and transfer between differently owned stores on cost per unit are also presented in column 1 of Table 1. The negative coefficients b2 and b3 suggest that both transfer between commonly owned stores and transfer between differently owned stores account for significant decreases in the unit cost of production. Analysis of the residuals from Equation (3) revealed first-order autocorrelation. There was no evidence of higher order autocorrelation. All of the models shown in Table 1 correct for first-order autocorrelation by jointly estimating the correlation coefficient with other coefficients of the models. Models with more control variables were estimated to explore alternative explanations for the results. We divided the control variables into two separate sets

including the depreciation parameter, X, and the unknown production histories, zni, also enter nonlinearly. Therefore, a nonlinear search algorithm was used to estimate all the results we report. This algorithm searches for the parameter values that maximize the likelihood function.

(Equations 2 and 3) in order to better understand the incremental impact of each control variable. In Column 2 of Table 1 calendar time is introduced to capture the possibility that technical change associated with the passage of time rather than learning associated with organizational experience was responsible for decreases in unit production costs. The positive coefficient for the time variable in Column 2 of Table 1 indicates that time is not a viable alternative explanation for the decrease observed in unit production cost. Rather the coefficient on the time variable indicates that the cost of pizza production increased with the passage of time, perhaps reflecting increases in food and labor costs over the one and a half year period of study. Current pizza count and the square of current pizza count are also included in Column 2 to capture the possible effects of economies of scale on cost per unit. The negative coefficient for current pizza count and the positive coefficient for the square of current pizza count in Column 2 of Table 1 indicate that significant scale effects are present. Cost per unit first decreased and then increased with increases in the current volume of production. The decrease in cost per unit as volume rises from relatively low output levels is quite natural since some labor and operating costs must be borne merely to keep a store open, and those costs are spread over more units as volume increases. Increasing cost per unit at higher volumes results from increased coordination costs. Coordination becomes difficult for high-volume production, especially since less experienced part-time employees are used to supplement regular employees during peak loads. A comparison between Column 1 and Column 2 of Table 1 reveals that the impact of transfer between differently owned stores is no longer significant, whereas the effects of store-specific learning and transfer between commonly owned stores on the unit cost of production remain unchanged with the addition of calendar time, current pizza count, and the square of current count. This illustrates the importance of controlling for scale economies since other variables may pick up their effects if scale variables are excluded. We conducted a specification test (Hausman 1978) to assess whether there might be simultaneity in the determination of cost per pizza and current pizza count.

MANAGEMENT SCIENCE/VOl.

41, No. 11, November 1995

1757

DARR, ARGOTE, AND EPPLE Productivity in Franichises

Our model allows for current pizza count to affect costs through economies of scale. It is conceivable that there is also an effect in the reverse direction if stores with lower cost per pizza charge lower prices and thereby generate a larger sales volume. Such simultaneity or endogeneity would lead to biased coefficient estimates. To test for the possibility of endogeneity of current count, the model was estimated with a two-stage least squares procedure that Fair (1970) developed. This instrumental variables procedure provides consistent estimates of models with endogenous variables and serially correlated errors. The coefficient estimates using instrumental variables were compared to the original coefficient estimates. The two sets of coefficient estimates are very similar. A test statistic of 8.16 was calculated for Hausman's specification test. The test statistic is distributed as X 2, df = 44, with a critical value of 60.48 at the 0.05 level. Thus, there is no evidence of endogeneity of current pizza count in the model. In Column 3 of Table 1 the square of the knowledge variable was introduced into the model to allow for possible changes in the learning rate. The coefficient for this variable was insignificant, indicating that there was no change in the rate of learning over the length of the study. The proportion of total pizza production accounted for by pan pizza was also introduced into the model at this stage to control for product mix. The estimate of pan pizza effects in Column 3 was insignificant, indicating that product mix does not affect the unit cost of production. A comparison between Column 2 and Column 3 reveals that the learning, time and scale effects are unchanged by the inclusion of these additional variables. The possibility that the results were driven by a few newly opened stores in the sample was investigated, by removing the four new stores from the sample. The results from the reduced sample were almost identical to the results shown in Table 1. 3.2. Depreciation Results Results on the depreciation of learning are presented in Columns 4 and 5 of Table 1. Column 4 is a linear model that does not include entire production histories, whereas Column 5 is a nonlinear model including entire production histories. The maximum likelihood estimate

of X for the model shown in Column 4 is 0.80. The hypothesis of no depreciation (X = 1.0) is very strongly rejected. Learning curve analysis has traditionally proceeded from the beginning of production in an organization. The majority of stores in this sample were in operation several years prior to the beginning of data collection. Through further data collection, we obtained complete production histories for 18 of the 36 stores in our sample. The impact of including entire production histories for each store on estimated learning effects was investigated using a nonlinear model in which pizza production prior to the beginning of our sample was added to the store, intrafranchise and interfranchise aggregates. For each of the eighteen stores in which we were unable to obtain complete data, production history was treated as an unknown coefficient. The results of these analyses are shown in Column 5 of Table 1. Including complete production histories does not change our results concerning store-specific learning, time, scale effects, and product mix. The effect of transfer between commonly owned stores remained significant (p < 0.05) but became somewhat less significant than in the previous analysis. The maximum likelihood estimate of X, the depreciation parameter, remained significantly less than one. The results indicate a very rapid rate of depreciation. A value of X = 0.83 implies that roughly one half (0.834) of the stock of knowledge 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 negligible amount (0.83 52) would remain one year later. In fact, without continuing production to replenish the stock of knowledge, virtually all production knowledge would be lost by mid-year.

3.3. Simulations of Effects of Learning, Depreciation, and Transfer on Unit Costs The following calculations provide an indication of the magnitude of the effects of learning, depreciation, transfer, and scale effects on costs. A typical store has a rate of production per store on the order of 1000 pizzas per week. Consider a store producing uniformly at this rate from the date of opening onward. Based on the results in Column 5 of Table 1, learning effects would

1758

MANAGEMENT SCIENCE/VOl. 41, No. 11, November

1995

DARR, ARGOTE, AND EPPLE Productivity in Franichises

lead such a store to have costs at the end of one month of operation that would be 20% lower than at the end of the first week. At the end of the second month, costs would have fallen an additional 8% below the level at the end of the first month. At the end of the fourth month, costs would have fallen an additional 4% below the level at the end of the second month. Beyond the fourth month, loss of knowledge through depreciation would offset the contributions to knowledge from ongoing production. If the store continued at the same rate of production in the future, learning would not contribute any further cost reductions. Next consider two stores opening at the same date each producing 1000 units per week. Suppose one is a single-store franchise while the other is a member of a five-store franchise, and, for simplicity, suppose all stores in the latter franchise open at the same date and produce 1,000 pizzas per week. The results in Column 5 of Table 1 imply that, at every date, a store in the five-store franchise would have costs 14% lower than the single-store franchise. Similarly, at each date, a store in a ten-store franchise would have costs 20% below those of a one-store franchise and 6.5% below those of a five-store franchise. To compare intrastore learning to intrafranchise learning, it is convenient to hold constant static scale effects (i.e., effects captured by the current count and square of current count variables). Consider a store producing 3,000 pizzas per week compared to a franchise with three stores each producing 1,000 pizzas per week. At each date, the store producing 3,000 units per week would have costs 11 % below the costs of a store producing 1,000 units per week in a franchise with 3 identical stores. The above comparison illustrates the scale economies associated with learning in multiple-store franchises. Conventional static scale effects, measured by inclusion of current count and the square of current count, are also present. These static scale economies are most easily illustrated by comparisons holding learning effects constant. A store operating at a weekly output that minimizes average costs (2,222 pizzas per week) would have a 6% cost advantage over a store operating at the average output rate observed in our sample (1,119), a 26% cost advantage over a store operating at the smallest weekly rate observed in the sample (140), and a

17% advantage over a store operating at the highest weekly rate observed in our sample (3,429).3 In sum, the results in Table 1 uniformly support the hypothesis that store-specific learning significantly decreases unit production costs. The results also support the hypothesis that transfer between commonly owned stores significantly decreases unit production costs. Little evidence of transfer across stores that are owned by different franchisees was found. We hypothesized that knowledge transfer between commonly owned stores would be greater than transfer between differently owned stores because of greater use of transfer mechanisms between -commonly owned stores. Table 2 presents the results concerning the frequency of transfer mechanism use. As can be seen, the frequencies of phone calls, personal acquaintances, and meetings were significantly greater between commonly owned stores than between differently owned stores.

4. Discussion
The stores in our sample evidenced firm-specific learning: as they gained experience, the unit cost of production decreased at a decreasing rate. The results on firmspecific learning are robust: firm-specific learning effects contributed to reductions in production cost independent of calendar time, scale effects and product mix. Additionally, store-specific learning was evident when we added complete production histories and allowed for knowledge depreciation.

The various comparisons of learning and scale effects presented in the text are calculated from the following series of relationships using the results of Column (5) of Table 1. For a store producing at a constant rate q per week, store-specific knowledge is K, = 0.83. K,_1+ q. For a single store, franchise knowledge and store knowledge are the same. Holding other things constant, costs for such a store at date t compared . Costs to costs j weeks earlier are given by (K,/K, j)(KtlKt-j) at date t for a franchise with ii stores each producing q per period compared to costs at that date for a single-store franchise are given by n1-0094. Abstracting from static scale economies, costs for a single store producing ni q units per period relative to a franchise with ni Abidentical stores each producing q per period are given by nm-0106. stracting from learning effects, static scale economies imply that costs for a store producing at rate q1 per week relative to costs for a store producing at rate q2per week are e[-ooo36(q-q2)+o91x1o7(q-q2)1
3

MANAGEMENT SCIENCE/VOl.

41, No. 11, November 1995

1759

DARR, ARGOTE, AND EPPLE Productivity in Franchises

Table 2

TransferMechanismUse WithinFranchises versus Between Franchises Phone

Callsa

Personal
Acquaintancesb Meetingsa

transfer Intrafranchise transfer lnterfranchise X2

2.5 0.4 16.57

2.6 0.8 9.3

1.7 0.0 13.33

p = 0.002

p = 0.05

p =0.004
=

a Mechanismuse was measuredon a scale from 0 = never to 5 than once a week. I Mechanismuse was measuredon a scale from 0 = none to 5 than 10.

more more

This is one of the first studies to focus on learning in service organizations. Although the modal progress ratio in the food franchises we studied was less than the modal figure found in manufacturing, learning effects were significant contributors to the productivity of the stores. Further research is needed to determine whether the slower rate of learning is characteristic of most service organizations and if so, why the rate is slower than that typically observed in manufacturing. To accomplish this, we believe that it will be more fruitful to move beyond the diffuse characterization of "service" versus "manufacturing"organizations and focus on the specific variables that differentiate the two production environments. Factors that are likely to contribute to the differences observed include prevailing characteristicsof individual employees (e.g., their skill levels and lengths of service), characteristics of the organizations (e.g., opportunities for specialization and standardization), and the nature of the demand function for the product. For example, when we have interviewed managers in manufacturing organizations about factors responsible for organizational learning, they emphasize matching tasks to the expertise and interests of individual workers (Argote 1993). There was much less opportunity to do this in the pizza stores we studied than in the manufacturing plants. Thus, an important source of productivity gains was not available to these stores. Similarly, many manufacturing organizations are able to sequence their products in a way that maximizes productivity, while

this option may not be available to service organizations who produce on demand. We also found that knowledge acquired through learning by doing transferred across stores owned by the same franchisee. Knowledge did not appear to transfer, however, across stores owned by different franchisees. The frequency of phone calls, personal acquaintances, and meetings was significantly greater between stores owned by the same franchisee than between stores owned by different franchisees. These results on transfer of learning extend our current understanding of the conditions under which transfer occurs. They indicate that knowledge transfer between affiliated organizations is greater than transfer between independent organizations. It is interesting to compare the transfer results in this paper with previous transfer of learning results (Argote et al. 1990, Epple et al. 1991). The previous investigations found that intraplant transfer was greater than transfer between geographically separated production facilities. Intuitively, groups within a single plant are less independent than geographically separated groups. The previous results concerning transfer of learning are, therefore, consistent with the results presented here. Results indicate that rapid depreciation of knowledge occurs within the pizza stores sampled. This is not surprising given that the typical turnover rate of employees in these stores is approximately 300% per year. Managerial turnover is approximately 50% per year. A great deal of production experience may be lost through such rapid personnel turnover. Store-specific learning and transfer from stores owned by the same franchisee contributed significantly to unit cost reductions, however, despite the rapid knowledge depreciation. Several events that occurred in the food franchises after we completed data collection provide some validation for our findings. Three of the stores in the sample changed owners following completion of data collection for the study. These stores maintained only minimal operation or closed completely for a short time. Consistent with our results on depreciation, unit costs were considerably higher when the stores reopened than unit costs had been when the stores closed. Of special import is the fact that the three stores that closed or changed owners were all single-store fran-

1760

MANAGEMENT SCIENCE/VOl.

41, No.

1, November

1995

DARR, ARGOTE, AND EPPLE Productivity in Franchises

chises. These stores were not able to benefit from production experience at other stores. Consistent with our results, they were less productive than their counterparts in multiple store franchises. Future research is needed to understand more fully the conditions under which knowledge transfers and to determine the effectiveness of these and other transfer mechanisms. Does the effectiveness of a particular mechanism vary as a function of the type of knowledge being transferred? More generally, a greater understanding of the micro processes underlying the transfer of knowledge is needed. Research is also needed to determine the conditions under which knowledge depreciates and variables affecting the rate of depreciation. Bailey (1989) performed an interesting laboratory study that analyzed certain factors affecting individual forgetting. Bailey (1989) raised the interesting issue as to whether the rate of individual forgetting is constant across a spectrum of tasks. Our research has focused on organizational "forgetting" or depreciation rather than on individual forgetting. Comparing the estimated rate of depreciation of organizational knowledge across several studies suggests that the rate is indeed not constant across production environments. The fastest rate of depreciation has been found in the current study's analysis of fast food franchises. The next most rapid rate of depreciation was found in the study of shipbuilding (Argote et al. 1990). Our studies of automotive production have revealed the slowest rate of depreciation (Epple et al. 1991, Epple et al. in press). While further research is needed to explain the variation observed in depreciation rates, the research to date suggests that the technological sophistication of the production process may be a key factor. The fastest rates of depreciation were observed in organizations low in technological sophistication while the slowest rates were observed in technologically sophisticated production processes. These latter organizations have more opportunity to embed knowledge in the technology through changes in tooling and programming. and the like. Knowledge embedded in the technology may be more resistant to depreciation than knowledge embedded in individual workers or in other aspects of the organization. Knowledge embedded in

technology may also be a mechanism through which knowledge transfers across organizations in an industry (cf. Bahk and Gort 1993). In summary, our results demonstrate that service organizations also evidence learning: as stores gain experience in production the unit cost of production declines significantly. Knowledge acquired through learning by doing in these service organizations depreciates quite rapidly. We also observe that knowledge transfers across stores owned by the same franchisee but not across stores owned by different franchisees. Stores benefit from production experience acquired in other stores in the same franchise.4
We gratefully acknowledge the Carnegie Bosch Institute at Carnegie Mellon University and its support of the empirical component of this research. We also thank the Decision, Risk and Management Sciences program of the National Science Foundation (Grant Number SES9009930) for its support of the development of the methods used here. Portions of this work were presented at the 1991 meetings of ORSA/TIMS in Anaheim, California, the 1991 conference on Current Issues in Productivity at Rutgers University, the 1992 meetings of the Academy of Management in Las Vegas, Carnegie Mellon University, Cornell University, and Northwestern University. The authors wish to thank participants in these forums and the reviewers for their very helpful comments.
4

References
Argote, L., "Group and Organizational Learning Curves: Individual, System and Environmental Components," British J. Social Psychology, 32 (1993), 31-51. and D. Epple, "Learning Curves in Manufacturing," Science, 23 (1990), 920-924. , S. Beckman and D. Epple, "The Persistence and Transfer of Learningin IndustrialSettings," Management Sci., 36 (1990), 140154. Bahk, B. and M. Gort, "Decomposing Learning by Doing in New Plants," J. Political Economy, 101 (1993), 561-582. Bailey, C. D., "Forgetting and the Learning Curve: A Laboratory Study," ManagementSci., 35 (1989), 340-352. Baloff, N., "Startup Management," IEEETransactions,EM-17 (1970), 132-141. Dutton, J. M. and W. H. Starbuck, "Diffusion of an Intellectual Techand Control in nology," in K. Krippendorff (Ed.), Communiicationi Society, Gordon and Breach, New York, 1978. and A. Thomas, "Treating Progress Functions as a Managerial Opportunity," Academy of Management Review, 9 (1984), 235247. Epple, D., L. Argote, and R. Devadas, "Organizational Learning

MANAGEMENT SCIENCE/VOl.

41, No. 11, November 1995

1761

DARR, ARGOTE, AND EPPLE Productivity in Franchises

Curves: A Method for Investigating Intra-plant Transfer of Knowledge Acquired Through Learning by Doing," Organization Sci., 2 (1991), 58-70. , and K. Murphy, "An Empirical Investigation of the Micro , Structure of Knowledge Acquisition and Transfer Through Learning by Doing," Oper. Res. (in press). Fair, R. C., "The Estimation of Simultaneous Equation Models with Lagged Endogenous Variables and First Order Serially Correlated 38 (1970), 507-516. Errors,"Economnetrica, Galaskiewicz, J. and S. Wasserman, "Mimetic Processes Within an Interorganizational Field: An Empirical Test," AdministrativeSci. Quarterly, 28 (1989), 22-39. Ghoshal, S. and C. A. Bartlett, "Creation, Adoption and Diffusion of Innovations by Subsidiaries of Multinational Corporations," J. Initernationial Business Studies, Fall (1988), 365-388. Gruenspecht, H., "Dumping and Dynamic Competition," J. InternationialEconiomics,12 (1988), 225-248. Hausman, T. A., "Specification Tests in Econometrics," ECTRA, 46 (1978), 1251-1271. Hirsch, W. Z., "Manufacturing Progress Functions," Review of Economics and Statistics, 34 (1952), 143-155. Huber, G. P., "Organizational Learning: The Contributing Processes and the Literatures," OrganizationSci., 2 (1991), 88-115. Huberman, A. M., "Improving Social Practice Through the Utilization of University-based Knowledge," Higher Education, 12 (1983), 257-272. Joskow, P. L. and N. L. Rose, "The Effects of Technological Change, Experience, and Environmental Regulation on the Construction Cost of Coal-burning Generating Units," Rand J. Economics, 16 (1985), 1-27. Kelsey, S. F., S. M. Mullin, K. M. Detre, H. Mitchell, M. J. Cowley, A. R. Gruentzig, and K. M. Kent, "Effectof Investigator Experience on Percutaneous Transluminal Coronary Angioplasty," American J. Cardiology, 53 (1984), 56C-64C.

Kogut, B., "A Study of the Life Cycles of Joint Ventures," Management InternationalReview, 28 (1988), 39-52. Levitt, B. and J. G. March, "Organizational Learning," AnnlualReview of Sociology, 14 (1988), 319-340. Liebenz, M. L., Transferof Technology:U. S. Multinationals and Eastern Europe, Praeger Publishers, New York, 1982. Martilla, J. A., "Word-of-mouth Communication in the Industrial Adoption Process," J. MarketingRes., 8 (1971), 173-178. Mody, A., "Firm Strategies for Costly Engineering Learning," Managenent Sci., 35 (1989), 496-512. Rothwell, R., "Some Problems of Technology Transfer into Industry: Examples from the Textile Machinery Sector," IEEETrans. Engineering Management, 25 (1978), 15-20. Spence, A. M., "The Learning Curve and Competition," Bell J. of Economics, 12 (1981), 49-70. Tichy, N. M., M. L. Tushman, and C. Frombrun, "Social Network Analysis for Organizations," Academy of Management Review, 4 (1979), 507-519. Tushman, M. L., "Communication Across Organizational Boundaries: Special Boundary Roles in the Innovation Process," Administrative Sci. Quarterly, 22 (1977), 587-605. Wall Street Journal, "Lockheed Loses Hope-TriStar Program Will Show Profit but Sees Improvement," May 14, 1980a. , "Lockheed Plans to End Output of L-1011 Jet," June 10, 1980b. , "Lockheed to Cut L-101 1 Production by Fall, Fueling Speculation on Plane's Survival," July 1, 1981a. , "Lockheed Plans to End Output," December 8, 1981b. , "Delayed Takeoff: Stalled Jetliner Makers May Not Rise Steeply Even if the Airlines Do," December 9, 1981c. Yelle, L. E., "The Learning Curve: Historical Review and Comprehensive Survey," Decision Sciences, 10 (1979), 302-328. Zimmerman, M. B., "Learning Effects and the Commercialization of New Energy Technologies: The Case of Nuclear Power," Bell J. of Economics, 13 (1982), 297-310.

Accepted by RichardM. Burtoni;received May 15, 1992. This paper has beeno with the authors 7 moniths for 2 revisions.

1762

MANAGEMENT SCIENCE/VOL

41, No. 11, November 1995