0
The influence of line length on
productivity of serial production
systems: An empirical study.
University of Groningen Faculty of Economics and Business
Final Master Thesis Technology Management
July, 2012
Author: J. G. Juurlink 1st supervisor: N. Ziengs, MSc
Student number: 1733567 2nd supervisor: Dr. J. Riezebos
Address: Baanstraat 1
9717 GT Groningen Telephone: 050 280 290 7
2
The influence of line length on
productivity of serial production
systems: An empirical study.
Abstract
Purpose – This paper challenges the widely accepted believe that longer production lines will
result in loss of productivity. We use behavioral and traditional operations theory to show the effect of longer lines on the productivity of a line.
Design/methodology/approach – This study employed a between subjects laboratory
experiment to empirically test whether there is a tradeoff between line length and productivity of a production line. We used a 2x3 full factorial design and tested 97 participants.
Findings – We found mixed results with respect to the relationship between line length and the
productivity of a line. Throughput seems to drop initially but when line length further increases, the productivity will rise again.
Practical implications – Managers can choose to set up longer production lines, without the loss
of capacity that is historically associated with longer production lines.
Originality/value – As far as the authors know, this is the first empirical study on the effect of
longer production lines on productivity. Previous studies used simulation, mathematical models, or used assumptions without empirically validating their models.
4
Contents
1. Introduction ... 6
2. Background ... 8
2.1 Motivation gains of workers ... 9
2.1.1 Social comparison and social indispensability ... 9
2.1.2 Moderators ... 10
2.2 Motivation losses of workers ... 12
2.3 Material flow policies ... 12
2.4 Line length of serial production systems... 13
2.4.1 Operations Management on line length ... 13
2.4.2 Behavioral insights on line length ... 14
2.4.3 Integration of insights ... 15
3. Methodology ... 15
3.1 Participant characteristics and selection ... 16
3.2 Research design ... 16
3.3 Experimental instrument ... 17
3.4 Procedure ... 19
3.5 Validity ... 20
4. Results ... 21
4.1 Performance of production lines with different line lengths ... 23
4.2 Performance of production lines with different workinprocess restrictions ... 26
4.3 Performance of slowest and fastest worker in a production line ... 27
5. Discussion ... 29
6. Conclusion ... 31
Bibliography ... 34
Appendices ... 38
Appendix I – Checklist for planning experiment ... 38
Appendix II – Internal consistency of questionnaire on motivation ... 44
Appendix III – Ttest for Hypothesis 3 ... 45
Appendix IV – Ttest for Hypothesis 4... 46
6
1.
Introduction
Over the past few decades, serial production lines have been the subject of many studies (e.g. Gershwin 1987; Conway 1988; Doerr et al 1996; Schultz et al 1998). An important aspect in most of these studies is how to determine the length of the production line and the optimal buffer size (and place) between workstation in that line. From an organizational behavior view, increasing the length of a production line can have positive (e.g. upward social comparison) and negative (e.g. social indispensability) effects on productivity. However, from an operations management point of view, increasing the length of a production line will lead to larger idle time due to increased variability of the line. However, increased variability can only lead to larger idle time when buffer sizes are limited. In this paper, we explore the net effects of increasing the number of workstations in a production line on the productivity of the line and see whether there is a tradeoff between line length and productivity.
Increasing the number of workstations in a serial production line can have a number of advantages. For example, complex tasks can be divided into several smaller tasks that are performed over multiple workstations. The division of tasks can lead to higher specialization and therefore a reduction of time it takes to complete those tasks. This will lead to higher output of the system. Moreover, unbalanced lines can become more balanced because longer tasks can be divided into several shorter tasks that have the same cycle time (Caruso, 1965). Secondly, it can reduce the need to buffer between different sections of the line. Production lines that are cut into several shorter lines need extra buffer capacity between the different sections. These buffers do not necessary have to be larger, but the products have to be moved from and to inventory.
7 operations research in order to understand the underlying mechanisms which influence performance of production lines with different line lengths (i.e. number of workstations) and see if there is a tradeoff between line length and the performance of a line.
Empirical evidence of the effect of line length on the performance of a production line is lacking. Previous research used simulation models to determine the effect of line length on performance and did not take into account any behavioral aspects of the workers involved (e.g. Hunt 1956; Anderson & Moodie 1969; Gershwin 1987; Conway 1988; Hendricks 1992; Hendicks & McClain 1993). All studies showed that longer lines decrease throughput compared to shorter lines. However, experimental evidence suggests that productionline workers adjust their work rates in certain situations to prevent idle time (Schultz et al. 1998; Doerr et al. 1996). This behavior is known as statedependent behavior, in contrast to the stateindependent behavior of machines. The independence assumption states that the processing time of a machine is independent of past events and of the current state of the system, i.e. a worker’s processing time is not affected by changes in the size of the buffer, the amount of inventory in the buffer, or the other workers in the line. While the independence assumption appears valid in high inventory situations, in lowinventory systems work speed is not independent of the speed of coworkers or the inventory state of the system (Schultz, et al., 1998). Powell and Schultz (2004) construct a theoretical model on what is known empirically about statedependent behavior and then implement this model in Markov models for short lines and simulation models for longer lines. They show that statedependent behavior makes serial lines more efficient and reduces the detrimental effect that longer line lengths have on throughput. However, these findings are not empirically tested. Rather than empirically test their claims, the theoretical model they produced contains a number of assumptions about worker behavior. They assume that workers behave in a certain way and then simulate what happens if workers actually behave in this manner. In fact, they can be subjected to the same criticism they are trying to debate in their paper.
8 decoding numbers into characters. We used a 2 x 3 full factorial design, with the treatment factors “line length” and “buffer size”.
The structure of this paper is as follows. In the next section we review relevant literature on operations management and organizational behavior. We start with covering the operations management literature on serial production lines and material flow policies. Lastly, section 2 will cover social theories on behavior. The methodology is described in Section 3. Section 4 presents the results of the experiment. The results are discussed in Section 5 and Section 6 concludes.
2.
Background
As the number of workstations in a line increase, the probability of all stations being “up” (i.e. operational) decreases (Askin & Standridge, 1993). There are several causes for a workstation to be nonoperational but the most important are machine and/or line failures and variation in processing times between and/or within workstations. Buffers provide a means for decreasing the threat of a production line being down. Intuitively, managers will increase the buffer size when line length increases because longer lines are associated with a higher chance of workstations being down. On the other hand, large buffers will undermine a sense of responsibility of output because feedback on performance will be limited. When performance improvements are perceived as possible, inferior group members often try to match or even exceed these standards (e.g. Collins 2000; Munkes and Diehl 2003; Seta 1982). Furthermore, large buffers will increase the perceived dispensability of individual contributions to the group outcome and will result in a decrease of efforts (e.g. Kerr 1983; Kerr and Bruun 1983). Therefore, we argue that buffers sizes should not increase with line length because large buffers will decrease the sense of indispensability and the possibility of social comparison. These two sources of motivation gains of inferior group members will be discussed in more detail in the next sections.
9
2.1 Motivation gains of workers
Workers in production lines might have different work speeds. Slower workers, or inferior group members (IGM), are more likely to cause starving at upstream workstations or blocking at downstream workstations compared to other workers. These workers might be motivated to work harder because of the presence of more capable group members. The motivation gains of inferior group members seem to be based on different processes than the motivation gains of more capable, superior group members (Stroebe et al., 1996). This section describes two potential mechanisms triggering motivation gains of inferior group members, i.e. social comparison and social indispensability (Hertel et al. 2000; Weber and Hertel, 2007).
2.1.1 Social comparison and social indispensability
People usually adjust their performance to standards provided by their social environment, particularly in new or unknown tasks (Festinger, 1954). This adjustment is based on social comparison processes. The performance standards of superior group members may act as high and specific goals in the sense of goal setting theory and lead to higher performance of inferior group members (Locke & Latham, 1990). A number of studies have explicitly demonstrated upward comparison as an underlying process in the motivation gains of inferior group members (see Weber and Hertel 2007 for a Meta review). Their results suggest that upward comparison can trigger motivation gains of inferior group members regardless of the relative importance of their individual contribution to the group outcome (i.e. task structure). Compared to working alone, inferior group members increase their effort when working together with a superior other person under conjunctive task demands (Weber & Hertel, 2007).
10
2.1.2 Moderators
The relative impact of the two general mechanisms of motivation gains in groups is moderated by various context factors (i.e. moderators). Weber and Hertel (2007) identify some moderators that might influence the practical use of social comparison and social indispensability, which are: task structure, availability of partnerrelated performance information, physical presence, and gender. The first three moderators are of most significance for our study.
Task structure
The structure of a group task determines the relative importance of each member’s contribution to the group outcome and therefore has a strong impact on the effort of individual group members (Steiner, 1972). The production line is unpaced and each workstation acts independently. The task structure of the pull production line has a conjunctive structure, which means that the group performance is disproportionately dependent on the weakest member. The least competent (i.e. weakest) member must possess the resources required to meet the criterion of success. The potential productivity of the group is established by the resources of the least competent member; if he does not function at the level permitted by his resources, actual productivity will fall below potential productivity (Steiner, 1972). It should be noted that only the low inventory line has a conjunctive task structure. The productivity in the high inventory line is not disproportionately dependent on the productivity of the weakest member and therefore has no conjunctive task structure. Weber and Hertel (2007) hypothesized that compared to working alone, IGMs increase their effort when working together with a superior other person under conjunctive task demands, and these motivation gains should be higher than motivation gains under additive or coactive task demands. Their study confirms the hypothesis and this motivation gain is contributed to additional influences of social indispensability (Weber & Hertel, 2007).
Availability of partner-related performance information
11 and/or competition processes will automatically be facilitated (e.g., Seta 1982; Stroebe et al 1996).
Physical presence
The third moderator comprises the physical presence of coworkers. Nowadays, groups can be globally dispersed so group members are not physically in the same working area. Recent studies have demonstrated that working face to face leads to significantly higher effort increases in IGMs compared to working with physically absent partners (Lount, Park, Seok, Messé, & Kerr, 2007). The physical presence of coworkers enhances the social consequences of being the inferior person because evaluative feedback is more likely and social sanctions would be experienced immediately (Carron, Burke, & Prpavessis 2004; Goffman 1963). These effects are relevant both for social comparison and for social indispensability processes. Weber and Hertel (2007) hypothesized that the motivation gains of IGMs are higher when coworkers are physically present compared to conditions where coworkers are physically absent. They confirmed their hypothesis for group tasks with a conjunctive task structure. The tendency to show higher motivation gains due to indispensability of personal contributions for the group is higher when participants work face to face (Weber and Hertel, 2007).
The work pace of slowest worker will be influenced by other workers in the line. The magnitude of the influence will be determined by the presence of social comparison processes and social indispensability. We expect that social indispensability will decrease when lines become longer because workers might be less able to identify their contribution to the finished product. On the other hand, we expect social comparison to be of greater influence when lines are longer. We expect this because in longer lines, more coworkers will be present to give feedback on the slowest worker. Also, if longer lines lead to more blocking and starving, the slowest worker in the line will be more motivated to avoid being idle. This leads to the third hypothesis, which is also stated as a null hypothesis due to statistical reasons.
12
2.2 Motivation losses of workers
Although group work is considered as a remedy that might motivate individuals to work especially hard (Allen & Hecht, 2004), group work per se does not guarantee high effort and top performance (Weber & Hertel, 2007). The best know example of the latter is social loafing (Latané, et al., 1979), which describes motivation losses when individual contributions to the group product are not identifiable. Motivation losses occur when group members do not exert maximal effort in the group setting. Other sources of motivation losses of individuals when working in a group, compared to working alone, are the free rider effect (Kerr & Bruun, 1983) and the sucker effect (Kerr, 1983). Kerr and Bruun (1983) demonstrate in their experiments that under a conjunctive task structure, the superior group members felt more dispensable and exerted less effort.
We expect that the fastest worker in a line will reduce the effort that is exerted when production lines increase in length. The conjunctive task structure will cause the superior group member to feel more dispensable because in shorter lines, the superior group member is more likely to improve the performance of the weaker group members. Although we expect the work speed of the fastest worker to decrease, we stated the hypothesis as a null hypothesis because of statistical considerations.
H2: In shorter lines, the average work speed of the fastest worker will be the same as in longer lines.
2.3 Material flow policies
In a mass production environment, flow lines are the most dominant method of organizing work (Billesbach, 1991). Material flow policies dictate the method by which work is passed along a production line. Two widely used material flow policies are push and pull strategy. A pull system is a material control system that controls the release and dispatching of work by explicitly limiting the amount of work in process that can be in the system (Hopp & Spearman, 2004). By default, this implies that a push production system is one that has no explicit limit on the amount of work in process that can be in the system (i.e. the production line).
13 throughput rate in pull policies. On the other hand, behavioral operations theory suggests that workers may adjust their work rate when they receive feedback from the state of the system (Schultz et al 1998; Schultz et al 1999). In high inventory production lines, the feedback that workers receive from the system is limited. Therefore, we expect that workers will be more motivated to adjust their work pace to other workers in low inventory production lines compared to high inventory lines. Feedback on the state of the system might enhance both social comparison processes as well as social indispensability processes. The hypothesis is stated below. We stated the hypothesis as a null hypothesis for statistical considerations.
H3: If the work in process level is explicitly limited (i.e. pull strategy), then throughput of the production line will be lower compared to a production line that does not explicitly limit the amount of work in process.
2.4 Line length of serial production systems
Operations Management (OM) research typically uses mathematical models and searches for optimum solutions to specific service and production situations (e.g. Hunt 1956; Anderson and Moodie 1969; Gershwin 1987; Conway et al 1988; Hendricks 1992; and Hendricks and McClain 1993). On the other hand, behavioral and human resource management (HRM) models often focus on conceptual relationships and search for enhanced descriptions and predictions of employee work behaviors and the effects of practices that affect those behaviors. Boudreau et al. (2003) propose that OM researchers consider refining their mathematical models using behavioral elements. Human responses to OM systems often explain variations that would otherwise be treated as randomness or error variance in traditional operations research models. The following paragraphs will cover the OM research on line length of serial production lines and some behavioral insights that lead to our main hypothesis.
2.4.1 Operations Management on line length
14 to N workstations by Basu (1977). Anderson and Moodie (1969) used simulation runs to extent Hunt’s formula to N workstations with normal processing times. They all noted the effect that longer production lines decrease the utilization due to more idle time. This effect has also been noted by other authors (e.g. Hillier and Boling, 1966; Koenigsberg, 1959).The decrease in utilization levels is contributed to several factors. Hunt (1956) states that blocking in each stage will have an effect on the preceding stages and the maximum possible utilization will drop. In longer production lines, the chance of a station being blocked is higher because there are simply more workstations. Gershwin (1987) used approximation and simulation to identify the influence of line length on production rate and found that production rates decrease as lines become longer. A conservative design rule for unbuffered workstations in a production line states that the loss in capacity occurs mainly in the first 5 machines; additional machines cause little additional loss (Conway et al, 1988). Hendricks (1992) studied the effect of line length on variance amplification. Variance amplification is a measure to study the variability of the output process and is defined as the ratio of the variance of the interdeparture distribution to the variance of the processing time distribution. Simulation was used to analyze the output processes of serial production lines with exponential processing time distributions and finite buffers. They showed that variance amplification is increased when lines become longer. Similar results were found with simulations with general processing time distributions (Hendricks & McClain, 1993).
2.4.2 Behavioral insights on line length
15
2.4.3 Integration of insights
The most important aspect of the behavioral literature on line lengths is the rejection of state independent behavior. Although simulation models suggest that there is a tradeoff between line length and performance of a line, empirical evidence on the behavior of workers in a production line leads to the suggestion that workers might be motivated to work harder when line length increases. We claim that increasing the number of workstations in a production line does not influence the productivity of the line. Stateindependent behavior would lead to the suggestion that blocking and starving will occur more frequent in longer lines but we expect that workers will adjust their work pace to the state of the system (i.e. the production line). Moreover, we expect workers to work harder because workers compare their work pace with other workers and do not want to be the weakest link in the chain. This leads to the first of our two main hypotheses that is stated below.
H4A: If the number of workstations in a production line is increased, then the throughput per workstation will be the same.
We explained that we expect workers to adjust their work pace to other workers in the line to avoid blocking and/or starving. This leads to the suggestion that an increase of line length will not result in an increase of the variance of processing times. In fact, one would expect the variance to decrease when lines become longer because workers will adjust their workpace to other workers in the production line.
H4B: If the number of workstations in a production line is increased, then the variance of the processing times per workstation will be the same.
3.
Methodology
16 the procedure of the experiment. At last, Section 3.5 elaborates on the threats to validity and how we tried to reduce these threats.
3.1 Participant characteristics and selection
The participants are male (n = 70) and female (n = 27) undergraduate students from the University of Groningen. They range in age from 19 to 27 years old. The majority of the participants (n = 91) participated in the experiment to earn course credits. The remaining participants (n = 6) were volunteers to whom we raffled some book vouchers. We tested the participants in group sizes of two, three, and four members.
The data is collected at a practical room at the University of Groningen. We had different time slots, and at each time slot we tested multiple groups of different sizes. Participants could enroll for a timeslot based on selfselection. In order to eliminate the selection bias that might occur, we randomly assigned participants that enrolled for one time slot into different groups. We used MS Excel to create a random number generator.
3.2 Research design
17 Workinprocess restriction Line length 2 3 4 restricted 11 12 13 unrestricted 21 22 23
Table 1 Treatment combinations and coding
We used a between subjects design. One downside of this design is that we needed more participants because every participant only received one single treatment at one level. On the other hand, this design eliminated the carryover effect that might arise due to fatigue or learning. Furthermore, the length of successive trails would be excessive for the participants. After the experimental task, we tested for motivation using a survey.
The assignment of participants to treatment combinations is completely at random. Awareness of a certain source of error in a statistical test of significance can only be taken if it has been left entirely to chance which treatment will benefit from this source in any single comparison of the treatments. This is an important reason for randomizing error variations with reference to treatments in each replication (Lindquist, 1953).
The serial production line we consider is an unpaced (i.e. each workstation acts independently and workers are able to dictate their own pace) transfer line with similar tasks at all workstations. The production line produces two products that only deviate in task length. Tasks are shortcyclic and repetitive with a cycle time of approximately 45 seconds. Sources of variation in processing times between different workstations in a line can be task differences, cycle time of machinery, and the work pace of the worker. We are interested in the variation of processing time caused by the work pace of the workers so we limit ourselves to the cases where workers are the primary determinant of processing times.
3.3 Experimental instrument
18 (multi user access to databases) for the server side. The interface of the application shows the production line at the top of the screen, see Figure 2. In order to further improve the experience of working in a production line, we used paper cards to simulate work orders. These cards have to be physically transferred from one workstation to the other by the participants which ensured that participants were aware of the dependencies between the workstations.
Participants log in at the workstation using an account that has been prepared in advance. This account entails all the information about the characteristics of the particular workstation. These characteristics comprise of the place in the line, the line number, and the workinprocess level that is allowed at that particular line. The first screen is an information screen that will give some information about the tasks that have to be executed. After the information screen the participant can start with the tasks of the experiment. The data that is entered during the tasks is transferred to a database that is accessible to all participants in a particular line.
The task consists of decoding numbers into letters. The participant has to fill in the correct characters in a dialog box below the series of numbers, see Figure 1. Once all the characters are filled in, the participant can confirm the order by clicking on a button. After the order is confirmed, a new order will arrive. In the pull system, a new order will only arrive when there is an order available in the upstream buffer and available storage space in the downstream buffer. The tasks are designed to have a mean processing time of approximately thirty to sixty seconds. The setup of the tasks is developed to ensure that learning effects are minimized. Before the participants can start the experiment, we let them do a test run of ten orders.
19
Figure 2 Interface of the production line (three workstations with buffer size of two orders)
3.4 Procedure
Participants were randomly assigned to a computer terminal where they received a personal login name so that all participants can be traced back to their group (i.e. production line) and their position in the group (i.e. production line). Each treatment combination was tested by a group of participants and we aimed for ten experimental runs per treatment combination. Due to the small number of participants we did not reach ten runs per treatment combination. In order to match a normal production line setting, participants were seated next to each other as they would do when working in a production line. This setting enabled feedback on both group and individual performance.
At the start of the experiment, participants received instructions about the experiment. The instructor orally informed the participants about their role and the participants received an information sheet. Participants did not receive information about the goal of the experiment and the hypotheses but they were asked to form their own goals in terms of number of orders per time unit.
Orders will transfer between successive workstations by means of cards. These cards contain an order number that has to be entered in a dialog box on the computer terminal. This triggers the task to appear on the screen. In the push system (i.e. no workinprocess restriction) there are unlimited intermediate buffers, i.e. the amount of cards in the buffer was chosen to prohibit blocking or starving from occurring during the experiment. The WIP level in the pull system (i.e. restriction on workinprocess) was set at two orders based on the pilot experiment (See Appendix I). In the pull treatment, the WIP is restricted and authorization to start working on a work order is only possible when there is an order available in the upstream buffer and the workstation receives an authorization card to start producing.
20
3.5 Validity
Behavioral experiments are subject to several threats on validity. In this section we explain how we tried to eliminate these threats. Three major sources of internal validity threats are demand characteristics, participantpredisposition effects, and the experimenterexpectancy effect. The next paragraph will elaborate on these threats.
The first threat to internal validity is the presence of demand characteristics. Researchers are often concerned with the presence of demand characteristics, cues that make participants aware of what the experimenter expects to find or how participants are expected to behave, and researchers typically use methods for reducing the demand (Nichols & Maner, 2008). Orne (1962) states that demand characteristics cannot be eliminated from experiments; all experiments will have demand characteristics, and these will always have some effect. Therefore, instead of eliminating the demand characteristics, they should be taken into account and manipulated if necessary. We tried to reduce the influence of demand characteristics by limiting the amount of interaction between the participant and the experimenter. Participants were allowed to ask questions related to the task itself (e.g. problems they encountered) but not about the experiment. Furthermore, the participants did receive some questions about topics that were not related to the goal of the experiment. For example, we did ask the participants to formulate goals in terms of products per hour they expected to produce. Furthermore, the experimental task did differ in length (the task consisted of 8 or 12 numbers that had to be decoded). This may have led participants believe that this was one of the treatments. After the participants finished the experimental task, we asked them to write down what they thought was the goal of the experiment. Out of the 97 participants only six related the goal of the experiment to study the effect of buffer sizes and two participants thought the experiment was set up to study the effect of line length.
21 data. We tried to eliminate these effects by stressing out that the participants were not judged in any form about their performance.
Finally, we tried to eliminate the experimenterexpectancy effects. This threat is associated with the expectancy the experimenter has on the response of the subjects. These expectations are likely to affect the choice of the experimental design and procedure in such a way as to increase the likelihood that his expectation or hypotheses will be supported (Rosenthal, 1976). We used scripted protocols and we limited the interaction between the experimenter and the participants during the second experimental task.
4.
Results
In this section we will address and discuss the outcomes of the experiment. First, we will describe some of the baseline information of the participants in the experiment. Section 4.1 will consider the effect of the different treatments on the performance of a production line. This section will cover Hypotheses 4A and 4B, dealing with respectively throughput and variance of processing times. We will also discuss idle time of low inventory lines in the first section. Section 4.2 will deal with the performance of a production line under different workinprocess restriction strategies, and test Hypothesis 3. The strategies we consider are, as mentioned in section 2.2, a pull strategy and a push strategy. Finally, section 4.3 is concerned with the slowest (inferior group member) and the fastest (superior group member) worker in a production line. In this section we will discuss the results Hypotheses 1 and 2.
22 Line length
WIP restriction 2 3 4 Total
Restricted 16 (8) 21 (7) 12 (3) 49
Unrestricted 2 (1) 42 (14) 4 (1) 48
Total 18 63 16 97
Table 2 Assignment of participants to treatment combinations. (The first number shows the number of participant per treatment combination. The number between brackets shows the number of groups.)
Treatment fidelity was achieved in two ways. First, treatment differentiation was achieved by ensuring that the high inventory treatment subjects received a large pile of cards (to represent the workinprocess level) while the low inventory treatment subjects only received 2 cards. This has been strictly adhered to with all experimental runs. Secondly, treatment integrity was achieved by following the protocol for every experimental run.
At the end of each experimental run we took a survey to test for motivation (5 point Likert scale). The survey was tested for internal consistency (Cronbach’s α = 0,84, see Appendix II for tests and questionnaire). Motivation has been tested for correlation with mean throughput for all treatment combinations (0,246). The mean and standard deviation of throughput for the two levels of motivation are shown in Table 3. We used median split (2.75) to group the subjects into either the “low motivation group” (average score < 2.75) or the “high motivation group” (average score > 2.75). Table 3 shows that for line length of two and three workstations, highly motivated subjects do have a higher throughput mean than low motivated subjects. However, a line length of four workstations does show the opposite effect. This might be due to the fact that for a line length of four workstations with unrestricted WIP we have tested only one group.
Motivation
Low High Low High
Line Length WIP Mean Mean St.Dev. St.Dev.
2 Restricted 75.400 81.161 8.375 5.005 2 Unrestricted 74.436 15.950 3 Restricted 62.083 69.799 10.368 12.160 3 Unrestricted 68.181 73.836 8.513 11.801 4 Restricted 67.320 66.758 5.937 6.278 4 Unrestricted 81.818 77.397 5.530
23
4.1 Performance of production lines with different line lengths
We use throughput, variance of processing times, and idle time as measures of performance. The first hypothesis (H4A) is concerned with the throughput of a production line when line
length increases. We hypothesized that increasing the number of workstations in a line does not affect the throughput of the line. We did not obtain a large enough dataset to perform a statistical analysis on this hypothesis. Instead we calculated the mean throughput for all treatment combinations. Table 4 shows the trend of the mean throughput. As we noted in the first part of this section, there is only little data for line length 2 and 4 of the high inventory level treatment combinations. These data points are most likely to be flawed because they can be heavily influenced by the performance of only one participant. For the low inventory treatment we have more data. As we can clearly see in Figure 3, there is a clear decay in throughput between two and three workstations. Earlier research supports the decay of throughput when line length increases (e.g. Gershwin 1987; Conway et al. 1988). The throughput in a production line of four workstations is expected to be lower than the throughput of a production line of three workstations. Most interestingly, there is no further decay in throughput. Moreover, throughput is increasing compared to a line length of three workstations. The increase in throughput might be attributed to the increasing availability of feedback and social comparison processes.
Line length
Buffer size 2 3 4
restricted 76.60 64.29 66.67
unrestricted 73.47 69.23 78.26
24 Figure 3 Throughput means for all treatment combinations
The second measure of performance we want to study is the variance of the processing times per workstations. We hypothesize that when the number of workstations in a production line is increased, the variance of the processing times per workstation will be the same. Variance in processing times can lead to blocking and starving of a line, which are both sources of idle time. Buffers can protect a production line from these sources of idle time. In fact, a large difference in variances of adjacent workstations in a pull production system is more problematic than in a push production system.
In our production line, workstations are occupied by human workers that can change their work pace triggered by influences around them. We expect workers to work harder when they are about to cause blocking and slow down when they are about to be starved. This is why we expect a pull production system to balance itself more than a push production system. Consequently, the variance of processing times in a low inventory line will be lower than the variance of processing times in a high inventory line. Our findings support these assumptions, as can be seen in Figure 4 and Table 5. Again, some data points have only a few records but when we compare both low and high inventory at a line length of three (these measures contain sufficient data) we see that the low inventory line has a clearly lower variance of processing times.
The high inventory line shows a decay of the variance as lines grow in length. Again, we keep in mind that we have limited data for this treatment. The decrease in variance at the high inventory line might be attributed to the influence of social comparison processes. At the production line of three workstations only the middle workstation can be blocked and starved. The first workstation can never be starved and the last workstation can never be blocked, so only the second workstation in the line is dependent on two other workstations. In the
25 production line of four workstations, there are two workstations that are dependent on two workstations, namely the two in the middle. This increase in interdependence might enhance the social comparison processes and cause workers to work more at one work pace.
Line lenght
Buffer size 2 3 4
restricted 0.91 1.34 0.79
unrestricted 10.61 6.38 2.94
Table 5 Variance of processing times for all treatment combinations
Figure 4 Variance of processing times for all treatment combinations
Idle time
Although we did not formulate any hypotheses regarding idle time, we can draw some conclusions from the data we gathered. In general, we distinguish two different sources of idle time. The first is called balance delay. Balance delay is defined as idle time due to differences in mean processing times between workstations in a line (Schultz et al. 1998). The second source of idle time is variance delay. Variance delay is concerned with the variance of processing times of individual workstations. Both sources of idle time can lead to blocking and starving of a workstation. In this study, we are interested in idle time due to variance delay. The amount of idle time can be calculated by dividing the difference between the average processing time of the slowest worker and the fastest worker by the average processing time of the slowest worker. The formula for determining the amount of idle time is shown below.
26 Wherein:
WS = Average processing time of slowest worker;
WM(1) = Average processing time of second slowest worker;
WM(2) = Average processing time of third slowest worker;
WF = Average processing time of fastest worker.
The idle time measures of low inventory lines can be found in Table 6. Idle time for a production line with three workstations is calculated to be around 5.9 percent. In their experiment from 1998, Schultz et al. show that a production line with three workstations has an idle time of 9.7 percent. Although our result is a little bit lower, it is still substantially lower than the predicted idle time of 16.1 percent (based on simulation models). These findings support the claim by Schultz et al. (1998) that production line models that use the independence assumption can significantly overestimate the amount of idle time in lowinventory systems.
Line length
WIP 2 3 4 Avg. Idle Time
2 2.69% 5.87% 6.06% 4.87%
Table 6 Idle time of low inventory lines
4.2 Performance of production lines with different workinprocess
restrictions
27 This would explain an increase in throughput. However, we expected that social comparison processes would be enhanced under a pull strategy, due to an increased availability of partner related feedback. These influences would counterbalance for the fact of absence of blocking and starving in a push strategy.
4.3 Performance of slowest and fastest worker in a production line
Hypothesis 1 states that in short production lines the average work speed of the slowest worker will be the same as in longer production lines. The slowest worker in a transfer line with restricted workinprocess is the bottleneck of the line. In the long run, the throughput of the production line will be equal to the throughput of the slowest worker. Table 7 summarizes the average processing times of the slowest (IGM) and fastest (SGM) workers in the line. The processing times seem to increase as line length increases. Figure 5 shows the average processing times of the slowest worker under both pull and push strategy. We see a different pattern for the push strategy. Instead of an increase in processing times, processing times actually decrease as lines become larger.
28 Figure 5 - Average processing times of the slowest workers in a production line
IGM SGM
line length restricted unrestricted restricted unrestricted
2 47.38 57.00 46.13 42.00
3 57.57 58.00 55.00 45.71
4 54.67 49.00 53.00 43.00
Table 7 - Average processing time of IGMs and SGMs
Hypothesis 2 is concerned with the performance of the fastest worker in the line. We state that in short production lines, the average work speed of the fastest worker will be the same as in longer production lines. As we noted in the case of the slowest worker in a production line, patterns seem to change under a push and pull strategy. We note the same trend as for the slowest worker (see Figure 7). Under a push strategy, the processing times will initially be higher. However, when lines become larger, the fastest worker will speed up under a push strategy and slow down under a pull strategy. We expect that this behavior is caused by blocking and starving. Like we mentioned earlier, workers are living entities that react on their environment. In a high inventory environment, a worker does not receive signals to slow down because he is working too hard. This is not the case for a low inventory environment. The worker will immediately be signaled to slow down because he or she is about to be starved or blocked. This explains why the fastest workers under a pull strategy will slow down. Furthermore, Table 7 shows that the superior group member will work harder when workin process is unrestricted compared to when workinprocess is restricted. This effect is the opposite of the slowest worker in the line. Again, we attribute the lower productivity of the fastest worker in the restricted line to the feedback that the worker receives from the system. In
29 an unrestricted line, the worker will earlier detect that he is working too fast for other people in the production line.
Figure 6 - Average processing times of the fastest workers in a production line
5.
Discussion
In the previous section we discussed the results related to our hypotheses. In this section, we want to place these results into context. First we will discuss similarities and differences between our results and results of others. Secondly, we discuss the generalizability and the practical significance of our findings. Lastly, we discuss the limitations of our research and we describe our ideal instrument for future research on this topic.
We showed mixed results with respect to line length and performance of a line. Initially, productivity seems to drop (from two to three station line) but with a further increase in line length productivity seems to rise again. The drop in productivity is what we might expect from the traditional operations management literature (Hunt 1956; Basu 1977; Gerswhin 1987; Conway et al 1988; Hendricks 1992; Hendricks and McClain 1993). The variance of processing times difference between the high and low inventory line is also shown by other authors (Schultz, 1998; Doerr, 1996). Our research suggests that the variance in processing times is lower in the low inventory line. The average idle time in our low inventory lines is around 5.9 percent. This is lower than the 9.7 percent idle time measured in by Schultz (1998). They already showed that simulation models could significantly overestimate the amount of idle time in low
30 inventory lines (simulation of three workstation low inventory line results in 16.1 percent idle time) but our research shows even smaller idle time.
Limitations
Our research has several limitations. The primary limitation is the small data set that disabled us to statistically test our hypotheses. Although we could partially test Hypothesis 2, we were unable to statistically test the other hypotheses. Another limitation that arose from the small number of participants is an unevenly distribution of participants over the treatment combinations. We tested the two and four station line with unrestricted WIP at only one group. The length of the experimental task is a second limitation of our research. The second round of the experiment lasted thirty minutes. All participants were able to finish at least thirty tasks but we believe that longer runs will lead to more accurate results.
We used a digital environment to imitate a production line setting. In order to more accurately match this production line setting, we used several tools. Physical tools we used are the cards that resemble the tasks, the line setting in which participants were seated, and the buffer spaces next to the participants. Apart from the physical tools, we used a visual tool. The visual tool is the interface of the software application that showed the production line and the buffer spaces on the computer screen. Furthermore, participants were allowed to communicate with each other. All these tools were developed to enhance the feeling of working in an actual production line setting, but this cannot conceal the fact that it is still a laboratory study.
Undergraduate university students were used as participants because they were most convenient available. This might jeopardize the generalizability of the findings because our participants will most likely not match the average production line worker.
Future research
31 As explained earlier, this research has some flaws that can be avoided in future research on this topic. We will outline our ‘ideal’ experiment that can be used as a guideline for researchers interested in this topic. We recommend to use high school students rather than university students to participate in the experiment because they will more likely match the population of production line workers. Next, the longest production line that will be tested should be determined. We noted that earlier research in operations management (Conway et al. 1988) suggests that lines longer than five workstations will show little additional loss in productivity. Our research suggests that behavioral theory might be more influential so longer lines will be more productive. We advise to use a pilot to test the treatment levels for line length. We suggest using a line length of three workstations as the shortest production line. In this line blocking and starving can occur at one and the same workstation. At a two station line, the first workstation can only become blocked and the second workstation can only become starved.
Participants have to be evenly distributed over the treatment combinations. This requires special attention at forehand. Finally, the second run in our experiment had a duration of thirty minutes. We feel that this is relatively short and that a longer duration might give more accurate results. In order to speed up the time till the production line is fully operational, the experimenter can choose to fill the buffer spaces with half the permitted amount of orders.
6.
Conclusion
Many authors have proposed that line length and the performance of a production line are inversely related. However, the studies lack an empirical foundation. In our study, we tried to empirically test the relationship between line length and the performance of a production line. We want to show whether there is a tradeoff between these two variables and perhaps challenge the widely accepted believe that longer production lines will result in a loss of productivity. Our results provide some interesting cues for for practitioners and researchers. In the next paragraph, we will concisely answer our research questions.
The last two hypotheses (H4A and H4B) were concerned with the number of
32 interesting results. Initially, throughput seems to decrease when lines become longer, for both the high and low inventory treatment. In the four station line, throughput increases where we would have expected a further decrease in throughput. With respect to the variance of processing times, our results are mixed. In the high inventory treatment, the variance of processing times decreases with increasing line length. In the low inventory treatment, the variance of processing times does not seem to be affected by increasing the length of the line. We further found that for a production line with three workstations, the mean variance in processing times in the high inventory treatment was significantly higher than in the low inventory treatment (t = 3.56 with p = 0.002).
Hypothesis H3 stated that the throughput of a production line that explicitly limits the workinprocess will be equal to a production line that does not explicitly limit the workin process. Again, we do not have sufficient data to statistically test this hypothesis for all cases. However, we can statistically test the three station production line for restricted workin process (N = 21) and unrestricted workinprocess (N = 42). Based on this data, we reject H2 (t = 1.596 with p = 0.116).
The last two hypotheses are concerned with the slowest (H1) and the fastests (H2) workers in a production line. Hypothesis 1 states that in short production lines, the average work speed of the slowest worker will be the same as in longer lines. We could not statistically test our data but the data showed that after an initial increase in mean processing times, the processing times of the slowest worker decrease when production lines become longer. We found similar results for the fastest worker in a production line. Again, after an initial increase in processing times, the processing times decreases in the four station line. These results are found both at the restricted and unrestricted production line. Furthermore, the slowest worker will work faster in a production line that does not restrict the workinprocess levels. We found contrary results for the fastest worker. The fastest worker will work faster in a line that does restrict the workinprocess levels.
Managerial implications
33 productive than short lines. This might motivate managers to reevaluate current production line systems. Furthermore, we showed that idle time in low inventory production line systems is significantly overestimated by simulation models. Based on our research, idle times of around 5 percent are expected in low inventory production line systems. The additional idle time that is incurred due to longer lines is small. Furthermore, our data suggests that longer lines will not result in higher variabililty of the individual workstations in that line. In fact, in high inventory lines the variability seems to drop when lines are longer.
34
Bibliography
� Allen, J. & Hecht, T., 2004. The "romance of teams": Toward an understanding of its psychological underpinnings and implications. Journal of Occupational and Organizational Psychology, pp. 439461.
� Anderson, D. R. & Moodie, C. L., 1969. Optimal buffers storage in production line systems. International Journal of Production Research, 7(3), pp. 233241.
� Askin, R. G. & Standridge, C. R., 1993. Modeling and Analysis of Manufacturing Systems. 1st ed. New York: John Wiley & Sons.
� Baker, K., Powell, S. & Pyke, D., 1990. Buffered and unbuffered assembly systems with variable processing times.. Journal of Manufacturing and Operations Management, 3(3), pp. 200223.
� Basu, R., 1977. The interstage buffer storage capacity of nonpowered assembly lines: a simple mathematical approach.. International Journal of Production Research, 15(4), pp. 365382.
� Bendoly, E., Donohue, K. & Schultz, K. L., 2006. Behavior in operations management: Assessing recent findings and revisiting old assumptions. Journal of Operations Management, 24(6), pp. 737752.
� Billesbach, T., 1991. A study of the implementation of justintime in the United States. Production and inventory Management, pp. 14.
� Boudreau, J., Hopp, W., McClain, J. O. & Thomas, J. L., 2003. On the interface between operations and human resources management. Manufacturing and Service Operations Management, 5(3), pp. 179202.
� Carron, A. V., Burke, S. M., & Prapavessis, H., 1991. Selfpresentation and group influence. Journal of Applied Sport Psychology, Vol. 16, pp 4158
� Caruso, F R, Assembly Line Balancing for Improved Profits, Automation, Vol. 12, (January 1965), pp. 485
� Collins, R. L., 2000. Among the better ones: Upward assimilation in social comparison. In J. Suls & L. Wheeler (Eds.), Handbook of social comparison: Theory and research (pp. 159172) New York: Plenum Press.
35 � Doerr, K. H., Brown, K. A., Mitchell, T. R. & Klastorin, T. D., 1996. Impact of Material Flow
Policies and Goals on Job Outcomes. Journal of Applied Psychology, 81(2), pp. 142152. � Doerr, K. H. et al., 2004. Work Flow Policy and WithinWorker and BetweenWorker
Variability in Performance. Journal of Applied Psychology, 89(5), pp. 911921.
� Festinger, L., 1954. A theory of social comparison processes. Human Relations, pp. 117 140.
� Gershwin, S. B., 1987. An Efficient Decomposition Method for the Approximate Evaluation of Tandem Queues with Finite Storage Space and Blocking. Operations Research, 35(2), pp. 291 305.
� Goffman, E., 1963. Stigma: Notes on the management of spoiled identity. New York: Simon & Schuster.
� Hendricks, K. B., 1992. The output processes of serial production lines of exponential machines with finite buffers. Operations Research, 40(6), pp. 11391147.
� Hendricks, K. B. & McClain, J. O., 1993. The output processes of serial production lines of general machines with finite buffers. Management Science, 39(10), pp. 11941201. � Hertel, G., Kerr, N. L., & Messé, L. A., 2000. Motivation gains in performance groups:
Paradigmatic and theoretical developments on the Köhler effect. Journal of Personality and Social Psychology, 79, pp. 580601.
� Hertel, G., Niemeyer, G. & Clauss, A., 2008. Social Indispensability or Social Comparison: The Why and When of Motivation Gains of Inferior Group Members. Journal of Applied Social Psychology, 38(5), pp. 13291363.
� Hillier, F. & Boling, R., 1966. The effect of some design factors on the efficiency of production lines with variable operation times. Journal of Industrial Engineering, 17(12), pp. 651658.
� Hopp, W. J. & Spearman, M. L., 2004. To pull or not to pull: what is the question?. Manufacturing & Service Operations Management, 6(2), pp. 133148.
� Hunt, G. C., 1956. Sequential arrays of waiting lines. Operations Research, 4(6), pp. 674 683.
� Jaffee, D., 2001. Organizational Theory: Tension and Change. New York: McGrawHill Companies, Inc..
36 � Kerr, N. L. & Bruun, S. E., 1983. The dispensability of member effort and group motivation losses: Free riders effects. journal of Personality and Social Psychology, 44(1), pp. 7894.
� Koenigsberg, E., 1959. Cyclic queues. Operations Research Quarterly, 9(2).
� Latané, B., Williams, K. & Harkins, S., 1979. Many hands make light the work: The causes and consequences of social loafing. Journal of Personality and Social Psychology, pp. 822832.
� Lindquist, E. F., 1953. Design and Analysis of Experiments in Psychology and Education. Boston: Houghton Mifflin.
� Locke, E. A. & Latham, G. P., 1990. A theory of goal setting and task performance.. Englewood Cliffs, NJ: Prentice Hall.
� Munkes, J., & Diehl, M., 2003. Matching or competition? Performance comparison processes in an idea generation task. Group Processes and Intergroup Relations, 6, 305 320
� Nichols, A. L. & Maner, J. K., 2008. The GoodSubject Effect: Investigating Participant Demand Characteristics. The Journal of General Psychology, 135(2), pp. 151165.
� Orne, M. T., 1962. On the Social Psychology of the Psychological Experiment: With Particular Reference to Demand Characteristics and their Implications. American Psychologist, 17(11), pp. 776783.
� Powell, S. G. & Schultz, K. L., 2004. Throughput in Serial Lines with StateDependent Behavior. Management Science, 50(8), pp. 10951105.
� Rosenberg, M. J., 1965. When dissonance fails: on eliminating evaluation apprehension from attitude measurement. Journal of Personality and Social Psychology, 1(1), pp. 28 42.
� Schultz, K. L., Juran, D. C. & Boudreau, J. W., 1999. The Effects of Low Inventory on the Developement of Productivity Norms. Management Science, pp. 16641678.
� Schultz, K. L. et al., 1998. Modeling and Worker Motivation in JIT Production Systems. Management Science, 44(12), pp. 15951607.
� Seta, J. J., 1982. The impact of comparison processes on coactors' task performance. Journal of Personality and Social Psychology, 42(2), pp. 281291.
37 � Stroebe, W., Diehl, M. & Abakoumkin, G., 1996. Social compensation and the Köhler effect: Toward a theoretical explanation of motivation gains in group productivity. . In: E. Witte & J. Davis, eds. Understanding group behavior: Vol. 2. Small group processes and interpersonal relations. Mahwah, NJ: Erlbaum, pp. 3765.
38
Appendices
Appendix I – Checklist for planning experiment
Objectives of the experiment
The objectives of the experiment are elaborated in the background section of the report. In this chapter we will only highlight the hypotheses. The first hypothesis is concerned with the processing times of individual workstations when production lines increase in length (i.e. number of workstations in one line). We claim that increasing the number of workstations in a line does not change the throughput and the variance. The hypotheses are stated as null hypotheses.
H4A: If the number of workstations in a production line is increased, then the throughput per
workstation will be the same.
H4B: If the number of workstations in a production line is increased, then the variance of the
processing times per workstation will be the same.
The second hypothesis is concerned with the workinprocess restriction of the line. We argue that workers do react on other workers and signals from their environment. These signals can be provided by means of the amount of orders in an upstream or downstream buffer. In a pull strategy, the amount of workinprocess is explicitly limited. This means that workers receive feedback about the state of the system because a station can become blocked or starved. In a push strategy on the other hand, the amount of workinprocess is not limited and therefore large buffers are allowed. We claim that the throughput of a production line that uses a pull strategy is equal to the throughput of a production line that uses a push strategy for order release. This leads to the following hypothesis:
39 Workers in a production line do have different work speeds. The slowest worker in a line (i.e. inferior group member) is the bottleneck when all buffer sizes are equal. Social comparison processes and social indispensability processes might motivate inferior group members to work harder. We want to study the impact of line length on the perceived indispensability and the possibility for social upward comparison. Therefore, the third hypothesis is stated as:
H1: In shorter lines, the average work speed of the slowest worker will be the same as in longer lines.
The fastest workers might also adjust their work speed. For an elaborated background, we refer to chapter 2. For the fastest worker, we formulated the following hypothesis:
H2: In shorter lines, the average work speed of the fastest worker will be the same as in longer lines.
Sources of variation
A source of variation is anything that could cause an observation to have a different numerical value from another observation (Dean & Voss, 1999). As major sources of variation we distinguish two types: treatment factors and nuisance factors. Treatment factors are of particular interest of the experimenter. The term treatment factor is used to mean any substance or item whose effect on the data is to be studied.
Treatment factors and their levels
The experiment employs a 2 x 3 factorial design. The treatment factors are “workinprocess restriction” and “line length”. The first treatment factor, workinprocess restriction, has two levels, “restricted” and “unrestricted”. A pilot was run to determine a suitable level for the “restricted” workinprocess restriction. The level for “restricted” WIP is set at 2 work orders. The levels for the line length treatment are set at 2, 3, and 4 workstations.
Experimental units
40 level is groups of students in a line. Data at this level consists of throughput and throughput time of the line. The second level is individual students. Data at this level consists of idle time, processing time, and throughput.
Blocking factors, noise factors, and covariates No blocking factors, complete randomized design.
Rule for assigning the experimental units to the treatments
We do not use any block designs or stratification so the assignment of experimental units to treatments will be completely at random. Randomization is achieved by a simple Microsoft Office Excel sheet in which we create two columns. The first column contains the number 1 till i, wherein i is the number of participants. The second column contains randomized numbers (random 0 till 100) and is sorted from low to high. Now the first column that contains the workstation numbers is randomized.
Pilot experiment
A pilot experiment can provide an opportunity to practice the experimental technique and to identify unsuspected problems in the data collection. We also had the opportunity to test the web application for a large number of participants. This section will describe the pilot experiment and elaborate on problems that we encountered during the pilot experiment.
The pilot
Date: February, 29 Time: 9:00 A.M.
Place: Computer room at University of Groningen
We (the experimenter and the assistant) arrived at the computer room at 8:00 in the morning to prepare the workstations for the experiment. Preparations consisted of:
1. turning on the personal computer;
2. opening Internet Explorer (IE) browser and enter URL for online questionnaire1;
1
41 3. opening Mozilla Firefox browser, set homepage to web application address, and
download Real Kiosk2 addon;
4. supply every workstation with an instruction sheet and an upstream and downstream buffer for the individual task of the experiment.
Fifteen students (from a second year undergraduate course Business) participated in the pilot experiment. The experimenter started with a brief introduction about the characteristics of the experiment. All participants received a consent form but the experimenter emphasized that it was completely voluntary to sign this form.
After the brief introduction, the participants were randomly assigned to a workstation. Randomization was done by using a Microsoft Office Excel sheet in which the numbers 1 till 15 were randomly sequenced and assigned to the participants.
The first task is an individual assignment intended to introduce the task to the participants. Instructions for the task were given threefold. The experimenter read the instructions, the instructions appeared on the screens of the workstations, and the participants had instructions on paper. All participants had to complete 15 tasks. The task involved decoding numbers into letters. For an elaborated description of the task, see chapter #.##.
After all participants finished their individual task, the experimenter explained the second task and handed out a form on which the participants could formulate their goal in terms of productivity. Participants were asked to formulate two goals: an individual goal and a group goal. After the participants filled in their goal form, the buffer spaces were stacked to one half of their limits.
The second task was running for about 3 minutes when a software error appeared. The pilot had to be stopped and the participants left the room.
Next, we outline the all phases of the pilot experiment.
2
