Académique Documents
Professionnel Documents
Culture Documents
in BPM
Aarhus School of Business, Aarhus University Department of Marketing and Statistics August 2010
Foreword
The present thesis represents the final project of the Master of Science programme at the Aarhus School of Business, Aarhus University, during the spring and summer 2010. In connection with the creation of the thesis, the authors would like to thank Karen Tybjerg from the former Arla Foods Global Ingredients Supply Chain Development for agreeing to collaborate with the authors with respect to the OEE initiative in Arla Foods. Furthermore, appreciations are sent to OEE Project Manager Per Hjort Petersen as well as the Production and Technology Development managers at the Global Ingredients production sites for the helpful sparring and assistance with the creation of the OEE study. Special thanks are given to Trine Obel Gregersen and the employees at HOCO for the kindness shown to the authors at their visits to the production site. The authors are both accountable for the entire content of the thesis.
II
Abstract
The present thesis consists of an assessment of the measurement of Overall Equipment Effectiveness (OEE), as well as an evaluation of the current and future OEE implementation in Arla Foods. The analysis is therefore divided into two sections: a theoretical part determining the components and validating the structure and relationships of OEE by means of structural equation modeling, as well as a practical analysis assessing the current effectiveness of the spray towers together with the applicability of OEE as a key performance indicator in Arla Foods Global Ingredients (GIN), and recommending the future initiatives, which the company must consider in order to enhance the success of the future OEE roll-out. Before the theoretical and practical objectives of the thesis are pursued, several preliminary exercises are performed to prepare the grounds for the comprehensive theoretical and practical analyses. First of all, the preliminary analysis begins with a review of the existing literature on performance measurement, management and key performance indicators, which reveals the importance of identifying, measuring, and managing the indicators central to organizational strategy. Furthermore, the empirical knowledge indicated the importance of aligning the measures at all organizational levels, so strategy can be implemented throughout the organization, and the entire company strives towards the same goal. Secondly, a theoretical and empirical assessment of the key performance indicator, OEE, is carried out. OEE is an aggregated measure of availability, performance, and quality of individual equipment, in which the various types of effectiveness losses are subtracted from the scheduled operating time. The OEE measure is subject to several points of critique ranging from the exclusion of planned downtime in the OEE calculation to the question of whether OEE is in fact the multiplication of availability, performance, and quality with equal weights. In the light of the review of key performance indicators and the SMART criteria, it is evident that OEE, despite the critique points, can be considered a key performance indicator, but the measure is modified into OEE* to accommodate the criticism and comply with the OEE definition in Arla Foods. Furthermore, an OEE study is designed in collaboration with Arla Foods, which is an international production company already in the process of implementing OEE with the
III
purpose of identifying bottlenecks, visualizing the utilization of the high capital cost equipment, and benchmarking the equipment effectiveness over time and across production sites. Arla Foods has decided to begin the OEE implementation on the spray towers at the Arla Foods GIN production sites, as the spray towers are assumed to be the bottlenecks in the production of ingredients, and at an OEE workshop in March 2010, it was therefore decided to collect data for the availability, performance, and quality of the spray towers in the fourth quarter of 2009. Finally, the preliminary analysis is concluded by an econometric analysis of the data to discover the tendencies and effects in the data set, and assess the applicability of structural equation modeling. It is revealed that there are fixed time-constant effects for the individual spray towers and production sites, as well as lags and time effects of the respective weeks, which entail that the error terms correlate over time and across units. The correlation can be remedied by means of fixed effects estimation and inclusion of lags and dummies. The theoretical part of the thesis is based on a structural equation modeling approach to OEE to investigate the components and structure of OEE, as well as assess whether a relationship between OEE and financial performance exists. The outer specification of the Partial Least Squares model is underidentified, and several of the manifest indicators load poorly on the associated latent construct. The inner model indicates that performance and quality loss, as expected, relate negatively to OEE, whereas availability loss relates positively to OEE, and the established relationship between OEE and financial performance proves to be weak. Even though not statistically significant, the path coefficients for the relations between the components and OEE reveal that validation of the multiplication of availability, performance, and quality with equal weights is not feasible. The data basis has deemed it difficult to draw conclusions based on the findings of the structural equation model, and it is therefore only possible to indicate the tendencies of the model. The practical part of the thesis comprises the computation of the current OEEs of the spray towers, and benchmarking of the OEE results across production sites by means of Repeated Measures ANOVAs in an attempt to discover the best practices. It is evident that the OEE results fluctuate significantly during the 12 week period, and that the defined threshold of 56% for satisfactory effectiveness is not always achieved. The
IV
Repeated Measures ANOVA for OEE reveals that differences in the OEE levels exist between the spray towers, and the reasons behind these differences are discussed in the light of the development in the variables for the OEE components. FIMIX-PLS segmentation is performed to discover whether the spray towers belong to latent groups, and even though the analysis reveals that a grouping into three segments is feasible, this segmentation cannot be substantiated by a pattern, which is intuitively reasonable, and it is therefore disregarded. The practical part of the thesis is concluded by recommendations for the future initiatives, which Arla Foods should consider in order to enhance the success of the complete OEE roll-out. It is essential that the production sites are harmonized and the data basis is improved, so benchmarking can be attempted and best practices communicated throughout the organization. Furthermore, it is necessary to create buy-in and a sense of ownership of the OEE project, from the front workers to the top management, as well as incorporate a cost perspective in the OEE consideration, so decisions concerning OEE are well informed. Finally, Arla Foods is in the process of implementing several strategic initiatives, and the company must therefore identify the must win battles, so the efforts and resources can be directed towards these particular initiatives.
Table of Contents
1 Introduction .............................................................................................................. 1 1.1 Problem Statement .............................................................................................. 2 1.1.1 Structure of Thesis ..................................................................................... 3 1.2 1.3 2 Delimitation ......................................................................................................... 4 Research Methodology ........................................................................................ 5
Performance Management ...................................................................................... 7 2.1 2.2 2.3 Performance Measurement Systems ................................................................... 7 From Measurement to Management.................................................................... 8 Key Performance Indicators ................................................................................ 8
Overall Equipment Effectiveness .......................................................................... 12 3.1 3.2 3.3 3.4 3.5 Total Productive Maintenance........................................................................... 12 Overall Equipment Effectiveness ...................................................................... 13 Critique of Original OEE .................................................................................. 14 OEE as KPI ....................................................................................................... 16 Modification of OEE into OEE* ....................................................................... 19 3.5.1 Availability ............................................................................................... 21 3.5.2 Performance ............................................................................................. 22 3.5.3 Quality...................................................................................................... 23
OEE* Initiative in Arla Foods ............................................................................... 24 4.1 Profile of Arla Foods ......................................................................................... 24 4.1.1 OEE Initiative .......................................................................................... 25 4.2 Optimal Measurement of OEE* in Arla Foods GIN ......................................... 26 4.2.1 Availability in Arla Foods GIN ................................................................ 27 4.2.2 Performance in Arla Foods GIN.............................................................. 27 4.2.3 Quality in Arla Foods GIN ...................................................................... 28
Design of OEE* Study ............................................................................................ 29 5.1 Data Collection for OEE* ................................................................................. 29 5.1.1 Background Variables ............................................................................. 30 5.2 Productivity and Financial Performance ........................................................... 31 5.2.1 Proxies for Productivity ........................................................................... 31 5.2.2 Proxies for Financial Performance ......................................................... 32 5.3 Validity and Reliability ..................................................................................... 33
VI
Econometric Analysis of Longitudinal Data ........................................................ 34 6.1 6.2 6.3 Sample Size and Missing Data ...................................................................... 35 Descriptive Statistics ......................................................................................... 35 Regression Analyses.......................................................................................... 36 6.3.1 Model 0: Panel LS Estimation ................................................................. 36 6.3.2 Model 1: Cross-Sectional FE Estimation ................................................ 37 6.3.3 Model 2: Cross-Sectional FE Estimation with Week Dummies............... 40 6.4 Discussion of the Final Models ......................................................................... 42
Structural Equation Model.................................................................................... 44 7.1 Design of the Structural Equation Model .......................................................... 45 7.1.1 Formulation of Hypotheses I, II, and III .................................................. 45 7.1.2 Approach to Structural Equation Modeling ............................................ 45 7.1.3 Estimation Technique............................................................................... 48 7.2 Estimation of the Structural Equation Model .................................................... 52 7.2.1 Theoretical Re-Estimation ....................................................................... 53 7.2.2 Practical Re-Estimation ........................................................................... 54 7.3 Validation of the Structural Equation Model .................................................... 56 7.3.1 Theoretical Evaluation of Fit Criteria ..................................................... 56 7.3.2 Practical Evaluation of Fit Criteria ........................................................ 57 7.4 Interpretation of the Structural Equation Model ............................................... 59 7.4.1 Relationships and Scores ......................................................................... 59 7.4.2 Components and Structure....................................................................... 62 7.4.3 Validation of Hypotheses I, II, and III ..................................................... 64
OEE* Results in Arla Foods GIN ......................................................................... 66 8.1 Effectiveness of Spray Towers in Arla Foods GIN ........................................... 66 8.1.1 Analysis of OEE* at Factory Level .......................................................... 67 8.2 Benchmarking across Spray Towers in Arla Foods GIN .................................. 69 8.2.1 Formulation of Hypotheses IV and V....................................................... 69 8.2.2 Repeated Measures ANOVA .................................................................... 69 8.2.3 FIMIX-PLS Segmentation ........................................................................ 75 8.2.4 Validation of Hypotheses IV and V .......................................................... 77 8.3 Applicability of OEE* in Arla Foods ................................................................ 78 8.3.1 Transparency of Utilization of Equipment ............................................... 78
VII
8.3.2 Best Practice ............................................................................................ 78 8.3.3 Concerns of OEE* Implementation ......................................................... 79 8.4 Recommendations for Further Roll-out of OEE* in Arla Foods ...................... 80 8.4.1 Harmonization ......................................................................................... 81 8.4.2 Buy-In and Ownership ............................................................................. 82 8.4.3 Bottlenecks and Interdependence of Production Sites ............................. 82 8.4.4 Data Basis ................................................................................................ 83 8.4.5 Cost Perspective....................................................................................... 83 8.4.6 Must Win Battles.................................................................................. 84 9 Further Studies ....................................................................................................... 86
List of Appendices
VIII
List of Figures
Figure 1 Structure of Thesis .......................................................................................... 3 Figure 2 OEE and the Six Big Losses ......................................................................... 14 Figure 3 OEE* ............................................................................................................. 21 Figure 4 Cross-Sectional FE Estimation ..................................................................... 38 Figure 5 Cross-Sectional FE Estimation with Week Dummies and Whites CrossSection Standard Errors .................................................................................................. 42 Figure 6 Approach to SEM.......................................................................................... 46 Figure 7 Path Diagram ................................................................................................. 47 Figure 8 PLS Algorithm .............................................................................................. 49 Figure 9 Model 0: PLS Algorithm ............................................................................... 52 Figure 10 Model 0: Bootstrapping .............................................................................. 53 Figure 11 Model 2: PLS Algorithm............................................................................. 55 Figure 12 Model 2: Bootstrapping .............................................................................. 56 Figure 13 Validation of Model 2 ................................................................................. 58 Figure 14 Path Coefficients from PLS Estimation ...................................................... 60 Figure 15 OEE* Results .............................................................................................. 67 Figure 16 Benchmarking across Spray Towers ........................................................... 70 Figure 17 OEE* Mean Differences ............................................................................. 74 Figure 18 Fit Criteria of FIMIX-PLS Segmentation ................................................... 76
1 Introduction
The current global business environment has led to an increased need for performance management systems, as continuous attention to and improvement of key strategic areas are essential to align organizational activities with strategic objectives and sustain competitiveness. Performance management is not only a method for measuring performance, but a comprehensive framework permeating the underlying culture of the organization. There is a tendency that organizations, which up until today have not focused on performance management, are initiating the implementation of key strategic activities leading to the development of a comprehensive performance management framework. In manufacturing companies, the implementation of performance management is particularly directed towards the optimization of the production and utilization of the production facilities. In these types of organizations, competitiveness is highly dependent on the availability and productivity of the equipment (Muchiri and Pintelon 2008), however, the effectiveness of equipment is difficult to assess, as many external and uncontrollable factors affect the individual machine. In spite of the abstract nature of the effectiveness concept, comprehensive research in the area has led to the development of several effectiveness measures, and many organizations have chosen to employ the measure overall equipment effectiveness (OEE) to increase the focus on utilization and productivity of individual equipment. However, empirical research has revealed that the method for measuring and computing OEE is ambiguous with certain limitations, and a critical analysis of the OEE measure is therefore interesting to conduct. It is desired to evaluate OEE from both an academic and practical viewpoint, and it is therefore ideal to collaborate with a large manufacturing company already in the process of implementing OEE, so the present thesis can contribute to a successful implementation, and the company can serve as the basis for the study. Arla Foods is an international production company, which has acknowledged the benefits of implementing performance management in the organization, and is currently in the process of launching OEE among several other value-creating initiatives. Collaboration with Arla Foods has been established in order to conduct the present
study, which can assist in the introduction of OEE, as well as in the complete roll-out of OEE in the entire production of all the business units.
The thesis will aim at answering the problem statement by means of analysis and discussion of empirical research, as well as knowledge gained at Arla Foods. For the
questions preparing the ground for statistical inference, i.e. questions ii), iii), iv), and vii), hypotheses will furthermore be formulated, tested and validated in the relevant chapters. 1.1.1 Structure of Thesis
When striving to answer the problem statement, it is reasonable to structure the thesis according to the outline illustrated in Figure 1. Figure 1 Structure of Thesis
The thesis commences with a general introduction to performance management and key performance indicators, and continues with a presentation and critique of OEE as an indicator of equipment effectiveness. Hereafter, the OEE initiative in Arla Foods is introduced, a research study is designed, and a preliminary analysis of the data set collected at Arla Foods is performed, before the twofold objective of a theoretical and practical OEE analysis is fulfilled. Finally, the thesis is completed by ideas for further studies, which consist of the issues surfacing in the theoretical and practical analysis that are interesting to investigate further, but is considered outside the scope of the present thesis.
1.2 Delimitation
Since a significant part of the thesis relies on collection of data for OEE, it has been deemed necessary to set a terminal date for accepting alterations in terms of structural changes in Arla Foods, as well as further changes in the data basis, and it is therefore decided not to incorporate information obtained after April 30th 2010. Eventually, it is desired to implement OEE across the entire production of Arla Foods, however, as it is unrealistic to successfully implement such a comprehensive initiative at once, it has been decided by top management in Arla Foods to begin with the production sites of the business group Global Ingredients (Arla Foods GIN). In the production of milk powder and whey protein, the spray towers are considered the bottlenecks, and the present thesis is therefore purposely limited to the spray towers on the Arla Foods GIN production sites. In addition, since it is already decided to continue with a complete roll-out of OEE, the objective of the thesis is not to make recommendations as to whether or not OEE should be implemented throughout the entire production in Arla Foods. Instead the thesis is limited to recommendations and suggestions as to how the full potential of OEE can be exploited, when attempting a successful implementation of the measure in the rest of the production. At the OEE Workshop in March 20101, it was furthermore decided, in collaboration with the OEE Project Manager and the Production and Technology Development (PTD) managers at the respective production sites, that the study is limited to the fourth quarter of 2009, and that all data therefore is collected retroactively, thereby accepting the constraints this decision may impose on the analysis. In addition, it was decided to refrain from collecting data for all spray towers, and the PTD managers at the respective plants therefore made a selection of spray towers, for which the data is collected. In this connection, it must be mentioned that the production site Visby is not represented in the selected sample, since it has not been possible to retrieve any data from this plant. Arla Foods GIN Supply Chain Development has developed a dashboard consisting of selected key performance indicators, and even though it is relevant to evaluate OEE in
the light of the other indicators in the dashboard, it is considered outside the scope of this thesis, and only OEE will be evaluated. In addition, it must be noticed that the econometric analysis performed in the thesis is not an objective in itself, but simply a preliminary exercise to discover the patterns and tendencies in the data set and assess the applicability of a structural equation model on this data. A comprehensive econometric analysis of the issues that are not relevant for the application of the structural equation model, is therefore purposely omitted. It must be mentioned that when discussing statistical significance, the significance level is set at = 0.05, if nothing else is stated in the thesis.
objective reality, and in order to approximate this reality, a triangulation of theoretical and empirical research is performed. The study is initiated by a deductive approach seeking to verify or falsify the well-founded theory, however, it is desired to create a basis for an inductive approach of generalizing the findings. The research process for the present study is inspired by the validity network schema, stating that research must include some set of concepts, methods, and focal problems of interest, which entail a combination of the conceptual, methodological and substantive domains. Depending on the order of the application of the domains, three different approaches to the research process are generated, i.e. the experimental, theoretical, and empirical paths (Brinberg and McGrath 1989). The theoretical path is chosen in this study, as it is desired to build a set of hypotheses based on elements from the conceptual and substantive domains and test them by evaluating them with the appropriate techniques from the methodological domain. This approach is chosen, since the main focus is to formulate and test hypotheses regarding the concept of OEE and its application in Arla Foods. The focal point of the thesis is derived from the initial research, which has revealed OEEs potential insufficient predictive ability of the equipment effectiveness, as well as the possible inappropriateness of OEE as a performance indicator in Arla Foods. The present thesis is characterised by both an academic and practical orientation, as it is the objective to test the theoretical basis for OEE, as well as to evaluate the OEE implementation in Arla Foods. According to Brinberg (1986), an academic orientation refers to conducting research, focusing on concepts and the relations between those concepts, whereas a practical orientation refers to conducting research focusing on a system, organization, or set of events in the real world. The difference between the orientations is the purpose guiding the study, and the objectives the researcher attempts to maximize. Due to the nature of the present thesis and the triangulation of the academic and practitioner approach, the authors expect to face decisions requiring certain compromises in order to satisfy both orientations in terms of theoretical tests of OEE, as well as problem oriented research necessary to conduct Arla Foods OEE study in a satisfactory and understandable manner.
2 Performance Management
The drive for quality management in the last couple of decades has enhanced the interest for performance measurement and management systems, as well as for the implementation of key performance measures, which is considered an important part of managing strategy (Eckerson 2009). There are several different approaches to and studies about performance measurement and management systems, but common for these approaches is the recognition of key performance indicators (KPIs) as an essential tool for alignment of strategy and performance.
implementing performance measurement and management systems, but the challenge is to implement and measure the right metrics. Most KPIs measure events in the past, and it is important to remember that information based on the past is only interesting, when it helps understand the potential for success in the future (Lebas 1995). As stated by Amaratunga and Baldry (2002), measurement provides the basis for an organization to assess how well it is progressing towards its predetermined objectives, helps to identify areas of strengths and weaknesses, and decides on future initiatives, with the goal of improving performance. In empirical literature, there is strong consensus of the value of measuring performance, however, it is important to determine whether all measures are KPIs or merely other relevant measures of performance. Parmenter (2010) suggests four categories of performance measures; Key Result Indicators (KRIs) state how a person, unit, or company has done in a perspective or critical success factor, Result Indicators (RIs) state what they have done, Performance Indicators (PIs) state what to do, and Key Performance Indicators (KPIs) state what to do to increase performance dramatically. Organizations must be aware of this distinction in order to be able to analyse and identify the core representative measures driving organizational objectives and goals. Furthermore, it is important to recognize that irrespective of the chosen indicator, a performance measure can at best pinpoint where to improve performance, but fails to provide information on the actions necessary for the improvement, and a qualitative analysis is therefore always suitable in combination with performance measurement. The method for measuring performance has many different names, such as KPIs, PIs, performance measures, performance metrics, etc., but there is consensus of the importance of creating and monitoring the right measures. The challenge is to design measurement metrics, which can act as a tool for the management to monitor, control, and improve the organization. The metrics are indicators that measure progress towards and achievements of certain goals. The measurements described will from here on be referred to as KPIs. Parmenter (2010) defines KPIs based on seven characteristics. The metric must be nonfinancial, be measured frequently, be acted on by CEO or senior management, clearly indicate the actions that are required by staff, tie responsibility to a team, have significant impact, and encourage appropriate action. A KPI therefore focuses on those
10
aspects of organizational performance that are most critical for the current and future success of the organization. Eckerson (2009) describes six important elements of a KPI, i.e. strategy, targets, ranges, encodings, time, and benchmarks. A KPI is characterized by the fact that it is measureable, encompasses a strategic objective, and measures performance towards a goal. The goal associated with a KPI is known as a target and is specified by a measurable outcome. This target may be a specific point or a range of values, within which the metric should lie. If the KPI falls outside this range and crosses a specific threshold, the visual display in the dashboard would change, and indicate a failure or success in reaching the target. Furthermore, it is common to specify certain timeframes as to when the KPI should be measured, and as to when the target outcome should be achieved. Finally, KPIs become really interesting, when benchmarking is possible, as this is a direct point for improving performance. Many authors have supported the concept of integrating the SMART (Specific, Measurable, Attainable, Realistic and Time-sensitive) criteria when implementing KPIs and pursuing goals in an organization. This concept ensures that each indicator is aligned with the strategy and appropriate for further actions and analysis (Shahin and Mahbod 2007). KPIs can be divided into two fundamental types: output KPIs, which measure past activity, and driver KPIs, which measure activities with significant impact on output KPIs (Eckerson 2009). It can be challenging to define accurate KPIs that drive future performance, and it is important to differentiate between KPIs and operational measures, which are relevant to measure, but not necessarily central to the strategy. In a large organization with a complex structure, issues are raised as to how to utilize, align, and integrate these performance measures in the various parts of the organization. In this connection, it would be beneficial to compute different dashboards, representing the different levels in the organization, and recognize that these various levels need different information and amount of details. The strategic dashboard enables the senior executives to execute strategy, manage performance, and drive new or optimal behaviour across the enterprise. Managers can optimize the performance of the people and processes under their supervision by using the tactical dashboard, whereas the operational dashboard enables frontline workers to monitor and control core processes
11
on a daily or weekly basis. Therefore, in order to potentially achieve the full, successful implementation of performance systems, the goals must be completely aligned throughout the organization. If each level of the organization implements performance measures, useful for fulfilment of the organization goals, the overall measurements of KPIs are optimized at top level, and ownership of the initiative exists from frontline workers to top management (Eckerson 2009). Implementation of KPIs in performance management and measurement systems is of great value to an organization, provided that the right measures and goals are chosen. Empirical literature suggests that goal setting, especially specific challenging goals, results in improved performance and productivity (Shahin and Mahbod 2007). The identification of KPIs helps the entire organization to work in the same direction, to reach the same objectives, as well as to create transparency and clarity throughout the business. Implementation and monitoring of goals and measures must be adapted to the multiple layers of organization in order to create the necessary ownership of the project, as well as an understanding of the alignment of measurement and strategy.
12
13
(1)
(2) (3)
Performance =
Quality =
(4)
The starting point for the OEE calculation is the planned operating time, which is the scheduled downtime such as holidays and scheduled maintenance subtracted from the total time (24 hours a day). The availability rate of the equipment is the fraction of the actual operating time, which is the planned operating time minus the unforeseen downtime, and the planned operating time, as seen in equation (2). Equation (3) shows that the equipments performance rate is the fraction of the smallest production time theoretically possible to the actual production time. Finally, the quality component is calculated as the rate of conforming output to the total output produced, as seen in equation (4).
14
The availability component captures the breakdown losses and setup and adjustment time, whereas the performance component includes idling, minor stoppages and speed loss, and the quality component encapsulates the yield and quality loss. Figure 2 OEE and the Six Big Losses
Eliminating all effectiveness losses, and achieving the zero loss objective, will yield an OEE of 100%. However, as in many cases this is not a realistic target, Nakajima (1988) has defined the optimal values for the OEE components. Ideally, the availability rate exceeds 90%, the performance rate is above 95% and the quality rate surpasses 99%, which results in world class performance and an OEE of 85%.
15
in the time base, and availability is therefore not measured against 24 hours a day. From one viewpoint this seems appropriate, as it is misleading to consider any planned downtime as a loss in equipment efficiency, when the equipment is not scheduled to operate, and therefore only effectiveness losses caused by the equipment should be accounted for by the equipment (De Ron and Rooda 2006). However, when aiming at visualizing and maximizing the utilization of the equipment, it is favourable to include the scheduled downtime in the time base (Ljungberg 1998). It is particularly important to modify the definition of OEE when operating in capital-intensive environments with high-capital cost equipment, as the utilization of the equipment is even more important under these circumstances (Jeong and Phillips 2001). It is suggested by several authors that the time base is extended to 24 hours a day, as it is theoretically feasible to constantly utilize the equipment (Muchiri and Pintelon 2008). It is important to notice that the size of the OEE measure depends on the time base, and the larger the time base, the lower the OEE. It is therefore not possible to compare OEE results with different time bases. Another point of debate is the difference between integrated and stand-alone equipment. OEE was originally intended to estimate the effectiveness of the individual, stand-alone equipment, but as the metric is influenced by various external factors such as the operator, input material etc., OEE actually measures the effectiveness of equipment integrated in a production line, i.e. the effectiveness of the equipment, including the effects from the equipment before and after the given equipment as well as other external effects. It can be argued that it is only appropriate to attribute effectiveness losses that are actually caused by the equipment to OEE. If the equipment is idle due to lack of feed or lack of operator, this is per se not attributable to the equipment, even though it decreases the utilization of the equipment (De Ron and Rooda 2006). Muchiri and Pintelon (2008) substantiate the belief that OEE is unsuitable, because it is limited to individual equipment and no machine is in fact isolated. On the other hand, it can be argued that there is a general risk that productivity measures are influenced by external factors such as skills of operator, type of product etc., and OEE is therefore not particularly disadvantaged. Furthermore, the OEE target levels have been heavily debated, and empirical literature does not agree with Nakajima on the target levels defined in 1988. Some authors argue
16
that an OEE target of 50% is more realistic, whereas others believe that OEE figures are acceptable between 60% and 75% or even between 30% and 60% (Dal, Tugwell and Greatbanks 2000). It must be kept in mind that the appropriate target level is of course dependent on the time base, so if OEE is modified and the time base extended, a lower OEE target is acceptable. Another interesting point of critique is the lack of cost perspective, which would be relevant to incorporate, since a tradeoff between performance and quality is likely to occur, and it is therefore necessary to be able to estimate the price of this tradeoff. E.g. if the equipment operates at a higher speed, which increases performance, there is a risk that the quality is reduced. Moreover, it could be interesting to consider the tradeoff between maintenance and breakdowns, as it must be assessed whether the costs related to preventive maintenance are less than the costs related to breakdowns. This critique of OEE is supported by the empirical literature, in which there is consensus that performance measures, such as OEE, must be based on a thorough understanding of cost relationship and behaviour (Keegan, Eiler and Jones 1989). Due to the before mentioned points of critique, it is essential for successful use of OEE that it is not applied in isolation, but in connection with other performance and operational measures, so an overview of the context can be obtained and the reality approximated. Finally, it is assumed that the three components of availability, performance and quality are equally important to OEE, however, this may not be the case (Pophaley 2010). It is questioned whether the multiplication of the three components with equal weights is the appropriate choice, and it is suggested that the three components carry different weights (Raouf 1994). In this thesis, a study of the components and the weights will be performed, as it is essential that the weights are correctly specified to avoid a wrongful and misguiding OEE measure, and research question iv) is therefore constructed to shed light on this issue.
17
performance, and quality as individual KPIs may be an exaggeration, however, OEE is a concise indicator because it includes more aspects in the calculation of productivity or effectiveness of equipment. There are several benefits of implementing OEE as a KPI. First of all, it is important to acknowledge that what is not measured, cannot be improved. No process can be improved, nor any goal achieved, if the performance is not measured and monitored. Furthermore, monitoring the effectiveness of equipment gives a good indication of the utilization of the equipment, and this visualization helps detect potential bottlenecks in a production site, as the different equipment may be dependent on the completion of the products in earlier parts of the production. In determining status quo of the effectiveness, the OEE measure creates a potential breeding ground for internal benchmarking across the production site or against previous years performance. The monitoring and benchmarking of performance create an incentive to perform better, and target and goal setting is known to have a positive effect on productivity (Shahin and Mahbod 2007). It is relevant to evaluate OEE as a KPI in the light of the different theories defining a KPI, and thereby determine whether OEE serves as a well-defined and useful KPI. When evaluating the use of OEE as a KPI in a strategic context, the SMART criteria are used in order to gain an understanding of the overall aspects of OEE as a metric used to measure progress towards the achievement of a specific goal. However, when discussing a more practical application of OEE, it seems relevant to incorporate various KPI strategies and definitions, and thereby provide a more practical view on how to define a KPI. When establishing whether OEE is more than an operational metric, focus must be held on the strategic aspect of OEE. A KPI must have a strategic objective and measure performance against a goal, whereas an operational measure merely refers to the measurement of business activity (Eckerson 2009). Even though the components and variables, which constitute OEE, are identified and measured at an operational level, and activities to improve production are initiated at a tactical level, the aggregated OEE measure serves as a part of the foundation for top management decision making, and it is therefore considered a strategic KPI.
18
The SMART criteria take strategic and organizational goals as its starting point, which is why this is not a specific criterion in the framework. The goal of improving productivity through OEE must be specific and detailed, in order not to lose track of the measurement and to be able to hold someone responsible for the execution and continuous improvement of OEE. As OEE is a measure used to monitor and increase the effectiveness of the equipment, which is derived from the objective of increasing productivity, OEE is a specific measure, which supports the goals of the organization. A key criterion is that the KPI is measurable, either in a qualitative or quantitative form, against a specific target or standard of expectation, which is fulfilled by the OEE measure, even though it is questioned whether OEE measures the effectiveness of standalone equipment, as it is intended to. It can be hard to establish precisely where to measure OEE, as it can be complicated to determine the exact starting and finishing point for various products moving through a chain of production processes. Any goal should be attainable and reasonable, but it should also be challenging and demand an effort in order to be successful. It is important for a KPI to measure performance against a target, so the success can easily be established. Originally, Nakajima (1988) stated that a reasonable target for an OEE calculation is 85%, as seen in section 3.2. However, any organization should determine an OEE target internally, as the OEE measure can be dependent on different factors not visible in the OEE metric, e.g. number of product shifts, age of the equipment, time base of the calculation, etc. Furthermore, it is important that the targets are set, so that they are demanding, but still realistic in the given organization, and so the targets are achievable provided that the equipment operates effectively. Finally, it is important that any goal has a time frame for completion. It is important to establish the time frame, so it can be determined whether the targets have been met in time. OEE can be calculated on a daily basis, so an OEE measure could possibly have a daily target. However, this makes little sense, since certain parts of the production activities, e.g. cleaning and maintenance, are related to the production over a longer period, and should not be attributed to only a few days. It would therefore be more reasonable to aggregate the OEE measures on a weekly or monthly basis. When evaluating OEE in terms of the SMART goal setting framework, it makes sense to consider OEE a KPI. There are, however, certain pitfalls for using OEE as a KPI and
19
thereby being a measure for strategic goals in the organization. These pitfalls will probably not occur until the implementation, and they are therefore more practical considerations associated with using OEE as a KPI. One potential problem is the point stated by Parmenter (2010), namely that a KPI should be sufficiently embedded in an organization to be tied to a single team. The OEE measure concerns and engages staff from many departments, because the measure is built upon three different components, which may have to be measured in different parts of the organization. Eckerson (2009) identifies a potential issue in terms of benchmarking. As previously stated, it can be risky to compare targets of OEE across industries and organizations, as OEE is not an isolated measure unaffected by environmental factors. Benchmarking therefore primarily becomes relevant in terms of internal benchmarking for the same equipment across a period of time. Furthermore, De Ron and Rooda (2006) criticizes OEE as a diagnostic measure explaining the causes behind the effectiveness losses, because OEE only indicates how effective the equipment is, but not why it is more or less effective. This conflicts with the initial definition of a KPI by Parmenter (2010), as a KPI should give directions as to how to increase performance dramatically. Even though, OEE may not explain direct causes, it still visualizes which parts of the production that are problematic and identifies potential bottlenecks. Therefore, the indirect cause of the effectiveness of the equipment can be identified through a more transparent production. Different authors have different views on the definition of a KPI, which criteria it must fulfil to be useful, and how it should be implemented and integrated in the organization. Referring to the discussion above, OEE fulfils the majority of the criteria for a useful KPI, which can help an organization achieve its objective of increased productivity and effectiveness of the equipment. However, since OEE can depend on various both internal and external factors, such as production structure, equipment diversity across the organization, etc., it is important to compare this evaluation of OEE as a KPI with the overall critique of OEE, in order to conclude whether OEE is a relevant KPI to implement in the organization in question.
20
collection at Arla Foods GIN, an attempt is made to create a modified OEE measure. As an important objective of OEE is to visualize the utilization of the equipment and eliminate all effectiveness losses, the original time base is extended to include scheduled downtime, as recommended by e.g. Muchiri and Pintelon (2008) and Dal, Tugweel & Greatbanks (2000). Equipment availability is calculated based on 24 hours a day, as this gives the full picture of the utilization of the equipment. Furthermore, the inclusion of planned downtime in the time base hinders the manipulation of the OEE measure in terms of unrealistic downtime by placing an emphasis on preventive maintenance as a means of minimizing unforeseen downtime (Dal, Tugwell and Greatbanks 2000). Thus, effectiveness losses unrelated to the equipment itself are also included in the modified OEE calculation, and it is therefore only reasonable to consider other external factors affecting the effectiveness of the equipment too. In continuation hereof, it is therefore decided not to consider the individual equipment in complete isolation, but to determine individual equipment effectiveness while accepting the equipment as an integrated part of the production site. The modification contradicts the original idea of a metric measuring the effectiveness of individual, isolated equipment, however, the modified OEE metric ensures a more useful and beneficial measure, as it enhances transparency and better explains the reasons for losses in effectiveness. Inspired by Jeong and Phillips (2001), it is furthermore decided to make alterations as to the categorization of the six big losses in OEE. Nakajima (1988) defined quality loss as quality and yield loss, however, it is chosen to disregard the yield loss in this component and instead include the yield loss in the performance rate, as the yield loss is in essence the loss in output resulting from start-up of the equipment. In addition, idling and minor stoppages are not included in the performance rate, but represented by the availability loss, as it is considered unplanned downtime. The modified version of OEE is termed OEE*, as seen in Figure 3, and forms the basis of the analysis performed in the present thesis. This modification is substantiated by the construction of the equipment-level productivity measure Total Equipment
Effectiveness Performance, which includes planned downtime in the total time horizon, so the time base is extended to 24 hours a day (Muchiri and Pintelon 2008).
21
Figure 3 OEE*
Source: Based on Nakajima (1988) and modified in collaboration with Arla Foods
The three overall components of OEE* remain the same as for OEE, however, the underlying reasons for effectiveness losses are scrutinized to ensure that all effectiveness losses are covered sufficiently. Also, it is attempted to break OEE* down to as many types of effectiveness losses as possible, in order to increase the transparency and make OEE* a better diagnostic measure. Based on the modification of OEE into OEE*, the following sections aim at answering research question i), by determining the variables which best depict the effectiveness losses when measuring OEE*. 3.5.1 Availability
The empirical literature suggests that availability losses can be broken down to a greater extent than proposed in the original definition of OEE. Availability losses encompass both non-production time, which is scheduled, and unforeseen downtime, and it is relevant to distinguish between these two categories (Smith and Hinchcliffe 2006). This tendency is substantiated by Turkcan (1999), who distinguishes between deterministic and stochastic reasons for equipment unavailability. Deterministic unavailability is downtime anticipated on beforehand, whereas the stochastic unavailability is the non-productive time not known a priori. Availability of equipment is constrained by several factors, but in relation to equipment effectiveness it is
22
necessary to distinguish between factors, which cause the equipment to be nonproductive, and factors making the equipment unavailable for a particular product, as the latter do not constitute a loss in effectiveness. As discussed by Turkcan (1999), facility unavailability, lack of operators, and preventive maintenance are the availability constraints necessary to consider in relation to the above criteria. However, as it is assumed that the facility is available 24 hours a day as seen in Figure 3, the unavailability of the facility becomes obsolete. In addition to preventive maintenance and lack of operator, it is furthermore relevant to measure the time for setup/adjustment and the changeover time when shifting from one product to another, separately (Muchiri and Pintelon 2008). In addition, standstill, idling, cleaning, and lack of feed constrain availability, and moreover, it is possible to specify breakdowns in detail by measuring people failure, machine failure, and line failure. By including the people failure and line failure, it is evident that the equipment is integrated and not considered in isolation. Thus, the availability component in OEE* is the total time (24 hours a day) minus the time lost seen in Figure 3. The nature of these availability constraints is likely to vary depending on the production plant in question, however, it must be recognized that some of the time losses are deterministic, whereas others are stochastic. 3.5.2 Performance
The performance rate expresses the actual performance of the equipment relative to its optimum performance. The actual performance can be measured in terms of output per time unit, whereas the theoretical optimum performance is harder to determine. The best estimation of the ideal performance will vary depending on the company implementing OEE*. Possible measures of maximum capacity are the maximum output produced in the last x number of batches, maximum output produced in the last month, the output guaranteed by the supplier of the equipment etc., however, the product type must be considered when comparing actual output against maximum capacity. Regardless of the decision, it is essential for the performance rate that the optimum performance is a good proxy for the maximum output possible to produce, as the performance rate otherwise risks exceeding 100%. If actual capacity exceeds maximum capacity, it is evident that the proxy for maximum capacity is misspecified, since the actual output produced can only be tangent to the real maximum capacity and never surpass it.
23
3.5.3
Quality
The quality component is defined as the ratio of conforming products to the total output produced. There are various viewpoints on quality, and depending on the product, the production site, and the company, there are different requirements for conformance. In some cases, there is a range in which the product is in conformance with the product standards, however, according to Taguchi and Clausing (1990), all deviations of a product from the specific target constitute non-conformance. When turning to the non-conformance procedure of ISO 9000, which is a family of standards for quality management systems, there are four kinds of non-conforming products, i.e. rework, scrap, products used for alternative use, and products accepted by customers as is or with concession, as seen in Figure 3.
24
25
producing and selling milk powder and whey protein. These milk-based ingredients can be applied in connection with fermented dairy products, ice cream, meat, bakery and infant, health and performance nutrition.2 The business group Global Ingredients (Arla Foods GIN) can furthermore be split into four business units, i.e. BU Milk Powder and BU Whey Protein, handling marketing, management and retail of milk powder and whey protein, respectively, BU Trading selling excess ingredients to industrial buyers, and Supply Chain, producing whey protein and milk powder and handling logistics. The ingredients are produced by six Arla Foods sites (HOCO, AKAFA, ARINCO, Denmark Protein (DP), Vimmerby and Visby) in Denmark and Sweden, and by a number of joint ventures and partnerships. The vision of the supply chain of Arla Foods GIN is to be a world-class supply chain of dairy based ingredients, and this entails the best logistics and supply chain structure, the best spray drying and ingredients technology competencies, the highest quality standards, cost efficiency in all aspects of the supply chain and setting the technology level for value-added ingredients. The strategic ambition for the business unit is to become the global leader for whey production, achieve a leading position in milk based nutrition products and establish a leading cost position for milk based bulk products in Europe. In order to meet this objective, a number of initiatives have been formulated to be implemented in 2010 along with a dashboard of performance measures indicating how the strategy implementation is progressing.3 4.1.1 OEE Initiative
One of the initiatives in the 2010 dashboard is the implementation of the equipment effectiveness measure, OEE, in the production of whey protein and milk powder. The aim of this project it to identify the bottlenecks in the production, visualize the utilization of the high capital cost equipment, and benchmark the effectiveness of the equipment over time and across production sites. The OEE activities are initiated on the spray towers, as Arla Foods GIN suspects this equipment to be the bottlenecks, and an increase in the effectiveness of the spray towers is therefore expected to result in substantial gain.
The authors are familiar with the organizational structure implemented in Arla Foods per 1st of May 2010, however, due to the terminal date set in the present thesis, discussion hereof is omitted. 3 This section is based on information from www.arla.com, annual reports and information provided by the OEE project manager regarding organization structure and business units.
26
The OEE initiative is managed by the OEE Project Manager, and the roll-out and continuous data collection are performed in close cooperation with the PTD managers at the respective production sites. Since Arla Foods GIN wishes to visualize the utilization of the equipment 24 hours a day, and determine the effectiveness of the equipment as an integrated part of the production line, the OEE initiative is based on the modified OEE* measure presented in section 3.5.
The sprays at Arla Foods GIN production sites are used to extract the milk powder or whey protein from the milk or whey. Liquid products are sent through spray towers under extreme pressure and high temperature, so only the powder and protein are left after the water has evaporated.
27
generally by means of the measures in bold, whereas the measures specific for the sprays in Arla Foods GIN are the sub-measures written in italics, as seen in Appendix A. 4.2.1 Availability in Arla Foods GIN
The measures defined as being ideal for OEE* in section 3, can be applied directly to OEE* in Arla Foods without further explanation or modification. However, there are some effectiveness losses, which require alterations to be suitable for the sprays in Arla Foods GIN. When considering the planned downtime, it must be noticed that the time, when an operator is not present, is not measured on the sprays in Arla Foods GIN, as the sprays are the bottlenecks and operators therefore are assumed to be available at all times. The downtime due to lack of feed, which for the sprays entail no milk or whey, is split into scheduled and unforeseen lack of feed, as both these scenarios are plausible. Standstill is included as an effectiveness loss, however, standstill will often be caused by lack of feed and therefore included under this measure. Preventive maintenance is only recorded as a loss in effectiveness, if maintenance is performed when the equipment could be operating. As preventive maintenance on the sprays is typically carried out, when the spray is unavailable all the same, maintenance rarely causes losses in effectiveness. It is chosen to distinguish between cleaning in place (CIP) of the tower and CIP of the miscellaneous parts of the spray, as these are not necessarily performed at the same time. In order to make the unplanned availability loss of the OEE* measure as diagnostic as possible, it is favourable to distinguish between the different types of people failures, machine failures, and line failures. In relation to people failures, it is relevant to distinguish between a breakdown caused by the operator and planning errors resulting in a filled silo or sudden lack of feed. The machine errors can be analyzed in terms of which part of the spray is causing the breakdown, and the line failures will represent a power blackout, problems with the steam or a miscellaneous breakdown. Repair time is not included as a measure, as the time necessary to fix a breakdown is included in the breakdown. 4.2.2 Performance in Arla Foods GIN
In relation to the performance loss, it is decided to measure the actual output produced per minute against the maximum output produced per minute in the same quarter the previous year, taking the product type into consideration. In order to reflect the maximum output possible to produce, it is necessary to consider the seasonal
28
fluctuations, as temperature and humidity have significant impact on the productivity of the equipment, and it is therefore chosen to apply the maximum output produced from the same quarter the previous year. Furthermore, it is relevant to take the product type into account, as the productivity is dependent on the type of whey protein or milk powder. 4.2.3 Quality in Arla Foods GIN
The quality control of the output produced by the spray must be performed by measuring the quality of the milk and whey before entering the spray, as well as of the ingredients when leaving the spray, as this will provide the best estimate of the quality loss attributable to the spray as stand-alone equipment. Furthermore, it is essential to ensure quality control before the products enter the spray, as it is very ineffective to risk wasting the resources of the spray on milk affected by bacteria, when the spray is the bottleneck in the production of ingredients. The quality loss is the rate of nonconforming products to the total output produced, and ideally the non-conforming products are divided into scrap, rework, and products sold for alternative use or/and at concession.
29
30
The performance component is defined as the rate of the actual performance to the maximum performance possible. Since data is only collected retroactively for the fourth quarter 2009, it is decided to apply the maximum performance for each product on each spray as the measure against which the actual performance is rated. Going forward, the maximum performance in the same quarter, but previous year, will be applied. The quality loss caused by the spray as stand-alone equipment is not possible to determine, as the quality is currently not controlled until the product is bagged. The quality loss actually measured does not represent the quality loss from the spray as stand-alone equipment, but the quality loss in the entire production of whey protein and milk powder. Quality problems attributable to the spray mostly arise when the water content is higher than acceptable, however, when the quality loss is not isolated to the spray, the loss will therefore capture e.g. bacterial problems not caused by the spray, and the quality loss will be greater than if only measured on the spray. The underlying assumptions for the classification of the data retrieved from the Arla Foods GIN production plants can be found in Appendix C. These assumptions have been necessary to make in order to obtain a similar data basis for each spray tower, which is detailed enough to figure in the desired statistical analysis. If these assumptions were not made, the data bases would be very different, and the number of variables possible to obtain, only based on the least common denominator, would be insufficient. The assumptions entail that the results from the analyses are less valid, however, since they are based on expert assessments from Arla Foods, they are considered the best estimates. 5.1.1 Background Variables
As previously discussed, successful use of OEE* as a KPI requires that OEE* is considered in connection with other operational measures or performance indicators, as OEE* should not be viewed in isolation. It is therefore chosen to collect data that can enhance the diagnostic power of OEE* and contribute to the explanation of the reasons behind the size and development of OEE*. The choice of background variables has been verified by PTD managers as well as the OEE Project Manager at Arla Foods GIN. For the purpose of background variables, data is retrieved on the number of tower CIPs and product shifts performed in the fourth quarter of 2009.
31
Empirical literature suggests that productivity is the relation between output and one or more associated input (National Research Council 1979), and it is measured in terms of the rate of output to input (Stehrer and Woerz 2009). It is therefore decided to apply the daily quantity of finished product and the daily weighing-in of milk at each spray, as this is an indication of the amount of raw material processed per day. The daily finished product produced on a given spray is easily obtained, however, the quantity of raw milk weighed-in is measured for the entire site. The best estimate for the amount of raw material weighed-in for each spray is the total raw material weighed-in on the site multiplied by the fraction of the finished product produced per spray over the finished product produced in total on the site. The calculations can be found in Appendix DD. On four out of five production sites, the raw material is raw milk, however, on Denmark Protein (DP) whey is weighed in. It is not reasonable to compare productivity of the respective sites, when the weighing-in of raw material used as a productivity measure is not identical. In order to correct for this issue, it is decided to find the best estimate of a factor converting the whey to raw milk. As the whey on DP differs, it is first of all essential to convert the various types of whey to thin whey equivalents. Even though it
32
is critical to convert the thin whey equivalents to raw milk, it is necessary to assume that the conversion factor is 1:16, as this is the best estimate and necessary to achieve a comparable dataset. 5.2.2 Proxies for Financial Performance
Daily data on the financial result for the respective Arla Foods GIN sites was not possible to retrieve, however, the financial performance of each site can be illustrated in terms of three different monthly indicators, i.e. monthly salary costs, monthly cost of raw material, and monthly capacity costs, the latter including energy consumption, maintenance, cleaning, water, and other operational costs. Naturally, these variables do not provide the full picture of the financial performance, as they only reflect the different types of costs on the production sites and neglect to illustrate the turnover, however, these variables are the best indicators possible to obtain. When estimating the relationships between the components of OEE*, productivity, and the various types of costs, it is crucial to keep in mind that high losses in effectiveness, i.e. a low OEE*, are likely to result in low variable costs and low material costs, as the equipment is not fully utilized. However, when the equipment is not utilized, no ingredients are produced and no turnover is incurred, which means that there is a cost of lost turnover, which is not reflected in the above cost measures. Since the monthly cost measures only result in three observations during the fourth quarter of 2009, it is decided to convert the monthly data to daily data and afterwards compute weekly averages of the daily data, so 12 weekly observations, corresponding to week 41-52, are obtained. The decision to convert to weekly data applies to the data on OEE* and productivity as well, and it is chosen, since monthly data leaves too few observations and daily data implies meaningless OEE* values, as some days are e.g. characterized by 24 hours of CIP or standstill. During a week, the extreme observations with zero productive or zero non-productive time will be less outspoken and provide values of OEE*, which are more reasonable. As the data for the financial performance and productivity measures are obtained for the entire Arla Foods GIN sites and not for the individual sprays, it is necessary to allocate the right amounts and quantities to the respective sprays on the site, so relationships between the effectiveness of the equipment, productivity, and financial performance can
6
The best estimate for a conversion factor is based on advice from Arla Foods contacts.
33
be established. The allocation is based on factors derived from the ratio of the respective sprays output in the fourth quarter of 2009 to the total output produced on the plant in the fourth quarter of 2009, seen in Appendix DD. It is assumed that the distribution among the sprays in terms of output produced corresponds to the distribution of the raw material weighed-in and the costs incurred. Considering the empirical knowledge of the effect of productivity on profitability, as well as the practical constraints caused by the proxies for financial performance, it is expected that a positive relationship between productivity and financial performance can be established, but that the relation is weak.
34
35
36
with the structural equation model in section 7, it is noted that the spray towers within a given series appear to follow the same pattern, which is characterised by the fact that monthly data has been recalculated on a daily basis and aggregated in weekly averages.
The initial model is constructed based on multiple regression of the 10 exogenous variables on productivity (IO), as seen in equation (5).
= + + + + + + + + + + +
where = 1, 2, , and = 1, 2, ,
(5)
The resulting model, seen in Appendix H, indicates that there are four explanatory variables, which have a positive effect (CIP tower (CT), setup (SE), breakdown (BD), idling (ID)) on productivity (IO), whereas the remaining six variables have a negative effect. However, as all the variables except from product shifts (PS), number of CIP (NC) and CIP tower (CT) are insignificant at 5% significance level, interpretation should be handled with care. The regression yields an R-Square of 0.318, indicating that the variation in the exogenous variables explains 31.8% of the variation in productivity (IO). Furthermore, the quality of the model specification is illustrated by a significant F-statistic indicating overall significance of the model.
37
Since LS is sensitive to autocorrelation, it is relevant to consider the Durbin-Watson statistic provided in the output in Appendix H, which tests for first order autocorrelation (Durbin and Watson 1950). With a value of 0.585, there is evidence of first-order autocorrelation, and actions must be taken to improve the model specifications and remedy the correlated error terms. The Durbin-Watson statistic only tests for first order autocorrelation, so it unreasonable to draw conclusions on the correlation of the error terms entirely based on this statistic, however, it is a good indication that there is an unobserved effect or omitted variables bias, which must be remedied. However, before alternative models are specified, the data is screened for outliers. 6.3.1.1 Outliers Outliers are observations in the dataset that are particularly influential on statistical tests. LS estimates are particularly susceptible to outliers, as the estimates are based on minimization of the sum of squared residuals. It can therefore be argued that outliers ought to be removed from the sample or winsorized, so they are no longer outlying observations. The data set is analysed for potential outliers by means of box plots of the residuals resulting from the LS estimation. As seen in the box plot in Appendix H, there is no evidence of outliers in the initial model specification, so the analysis is continued with a balanced panel of 120 observations. 6.3.2 Model 1: Cross-Sectional FE Estimation
The panel structure of the data set makes it interesting to test, whether there are unobserved effects over time and across units, which cause the error terms of the initial model to correlate. It is therefore tested, if there is an unobserved cross-sectional effect affecting all the observations related to a specific spray. This can be done by first differencing (FD) or fixed effects (FE) estimation. The FD and FE estimator are identical, when T 2, however, as T = 12 the estimators may not produce the same results. It is generally advised to apply FD estimation, when there are more time periods than cross sections (Wooldridge 2006), but as there may be insufficient variation in some of the exogenous variables, FD may be subject to substantial bias, and it is therefore decided to apply FE estimation. This decision is substantiated by the fact that the differenced error terms correlate as seen in Appendix I, which speaks in favour of FE.
38
where = 1, 2, , and = 1, 2, ,
(6)
As seen in equation (6), a potential unobserved time-constant effect, ai, which may correlate with the exogenous variables, is extracted from the composite error term, so ai and the idiosyncratic error ui are separated. This is done by correcting for fixed crosssectional effects in the LS estimation. Figure 4 Cross-Sectional FE Estimation
Source: EViews
It can be seen in the output above that the re-estimated model, including corrections for fixed cross-sectional effects, yields a significant increase in R-square and lower information criteria, indicating better model fit. However, it is important to be aware that fixed effects models tend to have very high R-squares, so the interpretation of Rsquare, as well as any re-specifications based on this quality measure, must be handled with care. It is furthermore important to notice that the Durbin-Watson statistic has
39
increased and no longer signals a problem of first order autocorrelation, which indicates that an unobserved time constant effect has in fact been extracted from the error term. It must be noticed that N = 10 degrees of freedom are lost, since for all cross-sectional observation i, one degree of freedom is lost because of the time demeaning. It is evident that the sign and size of the coefficients of the regressors have changed as a result of the correction for fixed effects. Unexpectedly, only the coefficients for CIP tower (CT), setup (SE), idling (ID) and non-conforming products (NP) are negative, however, it must be noticed that all the relationships are insignificant, which makes any interpretation difficult. These changes can be explained by the correction for the unobserved effect, which reduces the omitted variables bias and re-specifies the model. 6.3.2.1 Cross-Sectional RE Estimation In the above, it is assumed that the unobserved effect correlates with the exogenous variables, however, if this is not the case, the FE estimator is inefficient and a random effects (RE) model is more appropriate. However, an RE model assumes that there are more units than parameters to estimate, and therefore an RE model cannot be constructed for the total model. Since the correlation matrix, as seen in Appendix J, indicates multicollinearity between the number of CIP (NC) and the CIP tower (CT), one should be excluded from the analysis. It is chosen to retain CIP tower (CT) in the model, as it is part of the OEE* calculation and is more detailed, as it specifies the time spent on cleaning as opposed to the number of CIPs. This decision is substantiated by the fact that CIP tower is the variable defined at the OEE workshop, and therefore assumed to be a better indicator of the effectiveness loss. It must be noticed that number of CIP (NC) is excluded from the model specifications from this point on in the econometric analysis, however, it will be included in the structural equation model in section 7. The degree of explanation deteriorates for the RE model, as seen in Appendix K, and when the Hausman test is performed to test the correlation of the unobserved effect with exogenous variables, it is rejected, indicating that the key RE assumption is false and that FE estimates should be used (Wooldridge 2006). 6.3.2.2 LS Estimation with Factory Dummies Since it is evident that there is a significant fixed cross-sectional effect for the individual sprays, it makes sense to investigate whether there is an unobserved effect for each
40
production site, implying that the sprays are segmented according to production site. Since it is not possible to correct for effects of sprays and production sites simultaneously, it is necessary to perform LS estimation without correction for fixed cross-sectional effects on the sprays, as seen in equation (7).
= + + + + + + + + + + + + + + +
where = 1, 2, , and = 1, 2, ,
(7)
Four dummies are therefore computed, as seen in Appendix D, with the production site DP as reference, since it is evident in the initial research that this particular production site distinguishes itself from the rest of the plants due to differences in input, production plans, and product types. It is seen, in Appendix L, that the factories HOCO and ARINCO are significantly different from DP. As previously discussed, this difference may be partly explained by the fact that the input registered for DP differs from the other production sites, however, this will be discussed in section 8. The output indicates that the R-Square has slightly deteriorated, and it is therefore decided to exclude the production site dummies, and instead proceed with the crosssectional FE estimation and attempt to improve this model further. This decision is substantiated by the initial research, which implies that there are substantial differences not only between production sites but also between the individual sprays, and since it is no option to consider both effects, FE estimation is preferred, as it is likely to capture more of the omitted variables bias. The assumptions for FE estimation are therefore discussed in depth in Appendix J, and they will be continuously monitored in the following sections while experimenting with model improvements. 6.3.3 Model 2: Cross-Sectional FE Estimation with Week Dummies
The FE estimation, in Figure 4, reveals that there is a significant unobserved timeconstant effect, but it is also relevant to investigate the existence of an unobserved effect, which is constant across sprays for a given time period. The assessment of the normality assumption revealed that four outliers have been necessary to remove, so the analysis continues with an unbalanced panel of 116 observations. In order to investigate
41
a potential time effect, 12 dummy variables are constructed, as seen in Appendix D, corresponding to every week in the fourth quarter of 2009, and the effect of introducing these variables in the model is presented in Appendix M. The first week in the quarter is used as the week of reference, and the dummies estimated in the model are therefore week 2 to week 12, as seen in equation (8).
= + + + + + + + + + + + 2 + 3 + 4 + 5 + 6 + 7 + 8 + 9 + 10 + 11 + 12 + +
where = 1, 2, , and = 1, 2, ,
(8)
The output, seen in Appendix M, reveals that the dummies for week 5, 6, 7, 8, 9, 10 and 11 are statistically significant, which indicates that there is an unobserved time effect across the cross-sections in the mentioned weeks, implying that there is an unknown factor affecting all the sprays in a respective week. The dummies that are not independently significant may prove jointly significant, and a Wald Test for Joint Significance is therefore performed as seen in Appendix M, which reveals that the dummies are also jointly insignificant. However, as it is not sought to reduce the model until only significant variables remain, the insignificant dummies are maintained in the model. Even though 12 degrees of freedom are lost when including the dummies, and the Akaike and Schwarz Information Criteria have not improved, indicating that the benefit of including the dummies is outweighed by the penalty of estimating more parameters, it is decided to keep the week dummies in the model, as the degree of explanation increases and the correlation of the error terms decreases. Moreover, when assessing the correlogram, seen in Appendix M, it is evident that despite reduction in the correlation of the error terms, correlation still exists at almost all lags. The longer the time horizon, the more severe is the issue of correlation and the more appropriate is the use of robust standard errors (Kzdi 2003). This can be remedied by applying the White cross-section standard errors, which are robust to
42
contemporaneous serial correlation, i.e. the time-varying correlation across sprays which is not captured by the inclusion of dummy variables. Figure 5 Cross-Sectional FE Estimation with Week Dummies and Whites CrossSection Standard Errors
Source: EViews
Even though the error terms are still correlated, the inclusion of Whites cross-section standard errors allows for an interpretation of the output above.
43
in the above with the inclusion of week dummies and Whites robust standard errors, and yields an R-square of 98.42% with particular influence of the miscellaneous CIP (CM) and various week effects on productivity (IO). However, despite the FE estimation and the week dummies, this model specification still violates the condition of uncorrelated error terms, and this indicates that there may be more suitable models, despite the risk that these models may contradict the intuition behind the study. Even though inclusion of lags is difficult to justifiy in intuition, it is attempted to remedy the correlation of the error terms by including a lagged endogenous variable in the regression. It is seen in Appendix N that the correlation is improved, but since it is still present, a second lag is introduced in the FE estimation, and ultimately it is deemed necessary to discard the FE transformation and apply LS estimation with two lagged endogenous variables to accommodate the issue of correlation completely. These actions prove that the correlated error terms can be corrected for if necessary, but it is a tradeoff between a model, which is intuitively reasonable, and a model, which fulfils all required conditions but renders meaningless. Finally, it has been attempted to improve the model specifications and challenge the composition of the OEE structure by including squared and interaction variables in the regressions, wherever such effects could be justified in intuition. For instance, experiments with squared variables of the availability component have been performed, as the effect of breakdown, cleaning etc. may have a decreasing marginal effect. However, none of the attempts contributed to any further insight in the dataset retrieved at Arla Foods, as seen in Appendix O, and the idea of including interaction and squared variables is rejected. Even though the models presented in the above are all inadequate and do not provide complete insight in the data, which is to be applied in the structural equation model, there is generally consensus that unobserved time-constant as well as week effects exist, which cause the error terms to correlate. Due to the violated assumptions and the insignificance of the exogenous variables, it is not possible to draw conclusions based on the models, but it is evident that there are unobserved effects, which must be kept in mind when performing the SEM analysis.
44
45
The overall objective of the following statistical analyses is to determine the relations between the components and OEE, to discuss to which extend a relationship exists between OEE and financial performance, as well as to assess the multiplication of availability, performance, and quality with equal weights. For this purpose, the following hypotheses are formulated: Hypotheses Hypothesis I: The effectiveness losses relate negatively to OEE Hypothesis II: OEE relates positively to financial performance Hypothesis III: OEE is the multiplication of availability, performance and quality with equal weights As OEE* is not measured exogenously, it is necessary to apply a proxy, which can be measured exogenously, when testing the relationships. Since OEE* can be considered a measure of productivity for individual equipment, productivity is selected as the proxy for OEE*. 7.1.2 Approach to Structural Equation Modeling
In order to build the SEM model, the researcher must draw upon theory, prior experience, and the research objective, in order to specify which independent variables are likely to predict each dependent variable in the structural model. No model should be designed for SEM without some underlying theory to support the hypotheses and structure of the model specifications. In applying the SEM to test the theory of OEE, a confirmatory modelling strategy is chosen, where the objective is to test a prespecified relationship. It is important to notice that the resulting model, though having a good fit, is not necessarily the best possible model, but only one of several acceptable models.
46
A theoretical approach to SEM could be the Six Stage structure described by Hair (2005), which can be found in Appendix P. This structure indicates that the constructs and their corresponding manifest variables should be defined, specified and validated before the structural model can be specified. However, as the primary focus of the thesis is to test the theory behind OEE by means of OEE*, and its effect on financial performance, many of the inner relations are predefined based on thoroughly tested theories, and the approach to SEM is therefore customized to this specific scenario. Figure 6 Approach to SEM
The approach illustrated in Figure 6 differs from the Six Stage Process in the sense that the measurement and structural model are simultaneously specified, estimated, reestimated, and validated, as opposed to concluding a valid measurement model before introducing the structural relations. If the final SEM model is invalid, it has a significant effect on the depth of the conclusions possible to draw, as well as the interpretation of the model, and instead the structural model should be used as an indication of the relationship between the constructs. Re-estimation and validation are constrained by the limited data basis, as well as the specifications imposed by the theoretical background of OEE. The detailed approach to re-estimation and validation will be discussed in section 7.2 and 7.3.
47
The measurement model is formulated based on theoretical and empirical literature, as well as the conditions, under which the production sites in Arla Foods operate. As seen in section 4.2, the measurement theory prescribes that the constructs are represented by a number of manifest variables, among which it has been possible to retrieve data on some of the indicators. In the case of availability loss, the indicators are CIP tower (CT), CIP misc. (CM), setup (SE), standstill (ST), idling (ID) and breakdown (BD). Similarly, the performance loss and quality loss constructs are represented by capacity (CA) and non-conforming products (NP) respectively, whereas background loss is indicated by number of CIP (NC) and product shifts (PS). Furthermore, the construct of productivity is represented by input/output (IO), finished product (FP) and raw material (P), whereas financial performance is indicated by capacity costs (CC), material costs (MC) and salary costs (SC). In order to have the same expected valence of the loadings of the manifest indicators on the latent constructs, the data behind the three manifest cost variables representing financial performance is multiplied by -1. By doing this, it is expected that all indicators have a positive loading on their respective constructs. The measurement model is reflective, as the indicators do not fully explain the latent construct, but each manifest variable only reflects its latent variable. Likewise, the structural model is built on theoretical research as well as the application of OEE* in Arla Foods. Theory suggests that the structural model consists of a series of dependence relationships as illustrated in the path diagram below. Figure 7 Path Diagram
Source: Own creation based on the article by Fornell and Cha (Bagozzi 1994)
48
It is expected that productivity and financial performance depend negatively on availability, performance, quality and background loss, whereas it is expected that there is a positive effect of productivity on financial performance. Empirical literature suggests that the effect of productivity on profitability is delayed and therefore particularly evident in the long run, however, since the length of the lag is unknown, it is therefore attempted to establish a relationship between productivity and financial performance in the same time period. A list of the symbols applied in the path diagram can be found in Appendix Q. 7.1.3 Estimation Technique
When seeking to estimate a set of simultaneous structural equations and confirm an existing theory, the prevalent estimation technique is LISREL, which is a covariance structure analysis based on maximum likelihood estimation. This technique has, however, often been criticized for its inability to meet the requirements for maximizing the true scores of the estimates, as it depends heavily on fulfilments of strict assumptions (Fornell and Bookstein 1982). A violation of these assumptions leads to an inability to provide a solution, and in a study where the data basis is known to be inadequate, assumptions are expected to be violated, and a solution nevertheless is desired, LISREL may not prove suitable. It is therefore decided to estimate the structural equation model by means of partial least squares (PLS), as the econometric analysis indicated lack of assumption fulfilment. As opposed to LISREL, PLS is less sensitive to small samples and is not dependent on distributional assumptions (Bagozzi 1994), and it is preferred under conditions of statistical underidentification (Hair 2005). The greatest advantage of PLS is the robustness, which entails that it always provides a solution, despite the fact that problems exist. Even though PLS estimation is able to deal with non-normal data, it is relevant to state that normality exists, as this has consequences for the size of the sample. It is generally seen that a larger sample size provides better predictions and more stable results, however, the sample size depends on factors such as the number of indicators and constructs, the level of communalities, the deviations from normality, and the extent of missing data. The missing data only amounts to 5% non-random missing data, which is remedied by an EM algorithm, as seen in section 6.1, which is generally believed to outperform other missing data techniques. It must be noticed that there is an issue of underidentification, since there is general consensus that at least four indicators should
49
reflect each construct (Hair 2005), and in this study half of the constructs are only reflected by one or two indicators. Even though, in theory these issues cause a problem in relation to the size of the sample, the practical purpose of the thesis allows for imperfect data, since the problem of missing data is remedied and normality exists. 7.1.3.1 PLS Algorithm The PLS estimation is essentially a sequence of regressions, where one part of the model is held fixed, while the other is estimated. The basic idea of PLS is to minimize the variance in the residuals in the predictive relations, similarly to OLS, which seeks parameters that minimize residual variance (Bagozzi 1994). The variance must be minimized in both the outward directed blocks, caused by the reflective nature of the model, and in the endogenous latent variables. Figure 8 PLS Algorithm
In order to estimate the PLS model, the algorithm developed by Wold, who was the frontrunner of path modelling, is used. This algorithm is a three step estimation of the latent variables in the PLS model. However, before estimating the algorithm, arbitrary outer weights must be chosen, which are standardised by dividing by a normalized factor to obtain unit variance (Tenenhaus et al. 2005). The initial weights are often chosen to be 1, and this value is often predefined in PLS software.
50
The first step in the algorithm is the outer estimation of the standardized latent variables, and it is performed by linear combination of the manifest indicators related to the particular construct. The relationships between the latent constructs and the manifest indicators in the reflective model are illustrated in equation (9) and (10) (Bagozzi 1994), and an explanation of the symbols can be found in Appendix R. = + = + (9) (10)
The second step is the internal estimation of the standardized latent variables performed by means of the centroid weighting scheme, which was Wolds original algorithm for estimation of the inner relation. The centroid method is based on the sign of the correlations between the family latent constructs (Henseler, Ringle and Sinkovics 1994), and it is generally preferred over the factorial and path weighting schemes, as it is simple to carry out, and proves quite efficient. There is a risk of exaggerating the impact of certain correlation coefficients, if these are close to zero, as the method is based on the valence of the correlations only. From this viewpoint, it can be argued to apply the factorial weighting scheme (Bagozzi 1994), however, it is chosen to apply the centroid method as it is robust, and there is consensus that small fluctuations around zero are not a problem in practice (Tenenhaus et al. 2005). Equation (11) shows the relationship among the latent constructs according to Fornell and Cha (Bagozzi 1994), and the symbols are described in Appendix R. = + + (11)
The final step in the PLS estimation is the recalculation of the outer weights, by computing the covariance between the manifest indicators and the latent variables provided by the centroid method. There are two ways of estimating the weights, mode A and mode B, for the reflective and formative models respectively. It is chosen to apply mode A, as the PLS model, in this case, is of reflective nature (Tenenhaus et al. 2005). According to Fornell and Cha, the latent variables are estimated by the weights seen in equation (12) and (13) (Bagozzi 1994). = = (12) (13)
51
The three steps of the PLS estimation are iterated until convergence, after which the best estimates of the latent variables are determined. 7.1.3.2 Bootstrapping In order to determine the statistical significance of the relationships indicated by the PLS algorithm, it is chosen to apply a bootstrap technique. Bootstrapping is an approach to statistical inference, based on resampling with replacement from the original sample, which provides significance levels for all parameter estimates in PLS path modeling. The number of resamples should exceed 200, as this leads to more reasonable standard error estimates (Tenenhaus et al. 2004). The bootstrap technique yields estimates of the mean values and standard errors for each path model coefficient, which allows for a t-test of the significance of the structural relations. The t-distribution is essential for performing the t-test, as it provides the basis for the p-value (Henseler, Ringle and Sinkovics 1994). Therefore, even though PLS modeling is generally viewed as being free of assumptions in estimation, distributional assumptions are required in testing (Bagozzi 1994). Since it is given in the econometric analysis that the residuals are normally distributed, and it is known that the t-distribution approximates the normal distribution, when the sample size is sufficiently large, the condition is assumed to be fulfilled. There is a potential risk of increased standard errors and correspondingly low tstatistics, if arbitrary sign changes occur in outer weight estimation in the PLS algorithm and are not remedied. Any sign changes can be compensated for by choosing the relevant sign change procedure in the software used. The standard options are no sign changes, where no compensation is made, individual sign changes, where the signs in each resample are made consistent with the signs in the original sample, and the construct level changes, where the vector of loadings for each latent variable in each resample is compared to the corresponding vector of loadings in the original sample (Tenenhaus et al. 2004). Immediately, it seems ideal to apply the construct level changes as this ensures global coherence, however, as equal signs cannot be ensured prior to bootstrapping and model validation, it is preferred to apply the individual sign changes, where, if a path coefficient estimation for a bootstrap subsample shows a different sign compared with the original path model estimation, the procedure reverses
52
the sign of that path coefficient in the bootstrap subsample (Henseler, Ringle and Sinkovics 1994). 7.1.3.3 SmartPLS In order to construct the PLS model, it is chosen to apply the software SmartPLS, which is an acknowledged programme to estimate the measurement and structural model and to visualize the path structure. The PLS model is built by means of simple tools and estimated through a PLS algorithm, as described in the previous section. The PLS algorithm has a shortcoming, though, as it neglects to unstandardize the inner relations in the three step procedure, which entails that the t-statistics provided by the bootstrapping technique are imprecise. Since the prediction error is assumed to be present only in the third decimal, this problem can be disregarded and is considered outside the scope of this thesis. However, this shortcoming of SmartPLS could be remedied by transferring the standardized latent variable scores to another statistical software such as SPSS, and unstandardizing these scores before proceeding with the PLS model estimation.
Source: SmartPLS
53
Source: SmartPLS
In order to investigate whether the model can be improved, the approach in Figure 6 is applied, starting with the re-estimation of the PLS model. Since there may be differences between the optimal approach to re-estimation, and what is practically feasible with the restrictions of the econometric analysis and the objective of the thesis in mind, a theoretical approach is reviewed before the practical re-estimation is carried out. 7.2.1 Theoretical Re-Estimation
As previously introduced, the model consists of an inner and outer structure, which must be assessed in order to ensure that the model is correctly specified. The measurement model is evaluated to ensure equal valence, since it is important to make sure that the manifest variables all influence the construct in the same, expected direction (Tenenhaus et al. 2005). The outer weights are assessed to ensure that the manifest indicators affect the construct as anticipated, so the indicators that load with different sign than expected can be identified and potentially removed from the model. In addition, the manifest variables with weights exceeding 0.7 indicate sufficient impact on the construct on which they load, whereas the manifest indicators with weights smaller than 0.7 should be removed, as they do not reflect the construct adequately. In addition, it is necessary to assess whether multidimensionality exists, as it is essential to ensure that only one dimension is present in each construct, and unidimensionality
54
can be analysed by means of principal component factor analysis. The present thesis is built on a PLS model, which in theory is exploratory by nature, however, since the study has a confirmatory objective, a principal components analysis is omitted, even though multiple dimensions are suspected to exist in the availability construct. Despite potential multidimensionality, it is not desired to split the latent multidimensional construct into unidimensional constructs, due to the confirmatory modeling strategy of validating the OEE theory. Furthermore, it should be ensured that the model is based on internally consistent measures, which means that only one underlying construct exists for the given indicator. Internal consistency can be assessed by evaluating the individual cross-loadings, which should exceed 0.5 for the construct, on which the indicator is expected to load, and be below 0.5 for the remaining constructs (Hair 2005). Moreover, internal consistency can be confirmed by Cronbachs Alpha, where a value exceeding 0.7 indicates internal consistency (Tenenhaus et al. 2005). However, according to Netemeyer (2003), Cronbachs Alpha should only be used to assess internal consistency after unidimensionality has been established, and in the practical re-estimation the measure is therefore not used to confirm internal consistency. The structural model is assessed by looking at the relationships in between the latent constructs. The significance of the loadings estimated by the PLS algorithm is determined by performing a bootstrapping re-sampling technique, which visualises the t-statistics. The model is reduced by backwards elimination, removing the least significant relationship first, until all remaining relationships are significant at the chosen significance level, which is typically set at = 0.05, corresponding to a tstatistic of 1.96. 7.2.2 Practical Re-Estimation
In the practical re-estimation, it is important to keep in mind that the objective of the structural equation model is to evaluate the effect of the three OEE* components on productivity, as well as the influence of productivity and potentially the OEE* components on financial performance. The model is therefore not reduced beyond the relationships essential for the OEE* theory. The first step in the re-estimation is the removal of the manifest indicators, which load on the constructs with different signs than expected, and thereby entails that the
55
indicators, belonging to a given block, load on the latent construct with opposite valence. This implies that the indicators CIP tower (CT), setup (SE) and standstill (ST), which load negatively on availability loss, are eliminated from the model as positive valence is expected. Due to the existing problem of underidentification of the latent constructs, it is decided to accept all indicators with the expected valence, even though the loadings do not exceed 0.7. The measurement model is not further reduced and Model 1, based on the improvements discussed above, is re-estimated and illustrated in Appendix T, and tables of the path coefficients and t-statistics are presented in Appendix S. The second step in the re-estimation is the improvement of the structural model, which is conducted by assessing the structural paths in Model 1 and eliminating the insignificant relationships. It can be argued that path coefficients that are insignificant at = 0.05 corresponding to t = 1.96 ought to be eliminated. As previously mentioned, the objective of the analysis is to discuss the relationships essential for the OEE* theory, and these path coefficients will therefore be maintained regardless of the t-statistics. As there is consensus that t-statistics exceeding 1.0 contribute positively to the explanatory ability of the model, all relationships with values larger than 1.0 will be maintained in the model. Figure 11 Model 2: PLS Algorithm
Source: SmartPLS
56
Source: SmartPLS
The insignificant relationships are removed by means of backwards elimination, which entails that the path between performance loss and quality loss is removed at first, as it is the least significant. The model is re-estimated, and the relation between quality loss and financial performance is eliminated. Since all remaining paths carry t-statistics higher than 1.0, no further reductions in the structural model are necessary, and Model 2 is estimated and validated in terms of inner, outer, and overall model fit.
The path model is validated at three levels; the quality of the measurement model, the structural model, and the overall model. 7.3.1.1 Measurement Model Fit The measurement model fit can be evaluated by the Average Variance Extracted (AVE), which indicates how much of the variation in the manifest variables is captured
57
by the latent constructs. Usually, the threshold is 0.7, which indicates that 70% of the variance is captured in relation to the amount of variance due to measurement error. In addition, AVE can be used to evaluate the discriminant validity, which indicates the extent to which a construct is truly distinct from other constructs (Bagozzi 1994). Furthermore, the communalities can be used to assess the quality of the measurement model, by determining the total amount of variance a manifest variable has in common with the construct on which it loads. The threshold for the communalities index is approximately 0.5, and values below this threshold indicate model misspecification. Finally, the composite reliability specifies, whether the manifest variables load on the correct constructs, and this can be confirmed by means of the cross-loadings between the manifest variables and the latent constructs. Each manifest variable should load more than 0.5 on the construct to which it is intended, and less than 0.5 on the all other latent constructs (Hair 2005). 7.3.1.2 Structural Model Fit The quality of each structural equation is assessed by redundancy, which can be defined as the sum of variance in an endogenous latent construct, which is explained by the related exogenous latent constructs. There is no specific threshold for the redundancy index, since it is dependent on the number of exogenous constructs connected to the endogenous construct. Another measure for the structural model fit is the R-square indicating the explanatory power of the structural model. The R-square must exceed 0.6 to ensure a reasonable fit. 7.3.1.3 Overall Model Fit The overall path model can be evaluated by means of Goodness-of-Fit (GoF), which is a multiplication of AVE and R-Square, reflecting a combined validation of the measurement and structural model. The GoF measure is a globally accepted criterion for overall model fit, however, the overall path model should only be validated, if the respective measurement and structural model are also valid independently. The overall model fit in terms of GoF is intended accommodate the critique point that PLS has no optimization point. 7.3.2 Practical Evaluation of Fit Criteria
When validating the final model proposed in section 7.2.2, certain issues are important to keep in mind. First of all, the underidentification, which is particularly outspoken for performance loss, quality loss, and background loss, makes interpretation of the
58
validation measures of the given latent constructs difficult, and occasionally meaningless. Secondly, the correlation of the error terms, which was revealed in the econometric analysis, is expected to affect the standard errors and the t-statistics in the bootstrapping, and thereby also influence the validation of the model. Figure 13 Validation of Model 2
The quality of the measurement model can be assessed by evaluating AVE, composite reliability, and communalities. As seen in Figure 13, the six latent constructs can be categorised into three groups according to the respective levels of fit. The two latent constructs, financial performance and productivity, which are represented by fairly similar indicators, are, as anticipated, of highest quality according to the given threshold, as stated in section 7.3.1. Availability loss and background loss constitute the second group, which has difficulties meeting the prescribed fit criteria, but can be rendered acceptable in most occasions. The last group consisting of the two OEE* components, performance loss and quality loss, is characterised by only being identified by one manifest indicator, and therefore yielding meaningless results in all measurement model fits. The quality of the structural model is evaluated by means of redundancy and R-square. Except for an assumed acceptable measure for financial performance, the redundancies are very low, indicating poor structural regression equations. The R-squares for financial performance and productivity are again substantially higher than the degree of explanation for the three OEE* components. None of them satisfy the suggested threshold, but the measure for financial performance is, though, viewed acceptable, since one third of the variance in the construct is explained by the estimated model.
59
The overall model fit does not make sense to evaluate, since neither the measurement nor the structural model satisfies the respective quality criteria. The constructs, performance loss and quality loss, will wrongly improve the validation of the inner model in the GoF measure due to the extreme underidentification of these constructs. As previously stated, both the outer and the inner model must be validated in order for the GoF to be relevant. The validation of the PLS model must be performed, while keeping in mind that even though cut-off values are indicated for the respective fits, no model fit thresholds distinguish good models from bad models. It is furthermore important to acknowledge that SEM is applied to theory testing, and not necessarily the pursuance of best model fits. Model improvements must therefore not be attempted at the expense of unacceptable model specifications or compromise of the theory testing (Hair 2005).
The final model, seen in Figure 11, is characterised by t-statistics of the structural model, which are larger than 1.0, indicating that all the paths in the model contribute to a better degree of explanation. Since not all the path coefficients are statistically significant at = 0.05, as seen in Appendix S, the risk of committing a type I error is unknown and attempts to conclude definitively on the structural model should be avoided. The relations shown in Figure 14, will thus be interpreted with this uncertainty in mind.
60
Source: SmartPLS
Originally, it was expected that all path coefficients, except for the path between productivity and financial performance, would have negative signs, indicating that the exogenous variables in the structural equations would have a negative effect on the related endogenous variables. When looking at Figure 14, it is evident that the relations between background loss and financial performance as well as performance loss are positive, just as it is the case with the relation between availability loss and productivity. Furthermore, the path coefficient between productivity and financial performance is unexpectedly negative and is discussed in detail in section 7.4.1.1. Model 2 therefore indicates that Hypothesis I, stating a negative relationship between the effectiveness losses and OEE may not be possible to reject, as seen in section 7.4.3. However, availability loss loads positively on productivity with 0.09, indicating that the OEE components may have ambiguous effects on OEE. The three largest path coefficients are the ones between availability loss and financial performance, background loss and productivity, as well as background loss and financial performance, which are all numerically larger than 0.3. The structural regression coefficients between the effectiveness losses and productivity, which proxies for OEE* in order to test Hypothesis III on the structure of OEE, indicate that the equally weighted multiplication of the three components is misguiding even though this is not possible to confirm with statistical significance, as seen in section 7.4.3. It can be seen that the path coefficients between performance loss and productivity, as well as between quality loss and productivity, are similar with values of -0.155 and -0.134 respectively, whereas the relation between availability loss and productivity is characterised by a smaller, positive coefficient of 0.090. The discussion of the multiplication of availability, performance, and quality with equal weights must be
61
considered in the light of the chosen time base and the optimal measures of the OEE* components, since the OEE* differs from the original OEE formulated by Nakajima (1988). Furthermore, it is evident that background loss, which is not an original component of the OEE* measure, has important effects on the endogenous variables, especially on availability loss, productivity, and financial performance with large negative effects. 7.4.1.1 The Effect on Financial Performance The relationship between productivity and financial performance is fairly weak and the smallest path coefficient in the model. The small effect is anticipated since many other factors influence the financial performance of the company (Tangen 2005), however, this thesis is delimited from investigating other effects than the effects of productivity. As previously mentioned, the construct financial performance is reflected by inadequate indicators, since they only display costs and not revenues or other indicators of financial performance. It must therefore be kept in mind that the construct is built, based on indicators only reflecting one side of the financial performance. However, when the latent construct is estimated, it is viewed strictly as a proxy for financial performance, and the consequences of the measurement model are disregarded. Hypothesis II states that it is expected that when productivity increases, financial performance improves as well, however, as seen in Figure 14, it is not possible to reject, as the relation between productivity and financial performance is statistically insignificant at = 0.05. Despite the insignificance, there is an indication that the relationship may not be positive due to a loading of -0.078, which can partly be explained by an indirect negative relationship between productivity and financial performance, since higher productivity is assumed to correspond with higher costs, as the equipment demands input in terms of labour, raw material, and capacity. However, it is assumed that the positive relation between productivity and financial performance outweighs the potential conflicting effect. It would have been advantageous to include revenue-oriented indicators, when estimating the latent construct, as this would provide a more holistic image of the reality. Furthermore, the PLS model does not consider the expected time lag associated with the delayed effect of productivity on financial performance. If the time lag is included, and both revenue and cost aspects are represented, making it possible to determine which effect is more dominant, a better fit in the PLS model is expected.
62
7.4.1.2 Degree of Explanation When looking at the R-squares in Figure 13, the three OEE* components are, as expected, very poorly explained, with degrees of explanations below 1% for the two underidentified constructs, performance loss and quality loss, whereas the R-square for availability loss indicates that 7.6% of the variation in the construct is explained by the model. It must be kept in mind that background loss is the only exogenous variable explaining the other effectiveness losses, and without this construct, the three OEE* components would be completely exogenous in the model. The degrees of explanation for productivity and financial performance are somewhat higher at 19.5% and 31.1% respectively, but must still be regarded as below the desired threshold. These R-square values reflect the issues of underidentification and a poor data basis as previously discussed, and it must be kept in mind that there is a substantial amount of variation explained by factors outside the model. 7.4.2 Components and Structure
The present thesis relies on a deductive approach, where hypotheses are formulated based on theoretical and empirical research, and tested to verify or falsify the existing theory. The objective is to test the OEE* structure, but it must be kept in mind that it is not possible to blindly generalize the findings from this analysis, due to the compromised external validity, since the sample is not representative for all organizations, as only Arla Foods GIN production sites are studied. However, it is still desired to validate the existing OEE theory, and induce the potential falsification of the OEE structure to future studies. SEM is therefore not applied in this thesis to discover the optimal components and structure of OEE, but to test the existing theory. The model has primarily been based on empirical research on underlying theory, but practical implications have constrained the specification of the model. In order to construct an OEE* structure to test the theory, all the important indicators known from theoretical research, should be included in the model. Due to the practical aspect of the thesis, the inadequate data basis is accepted, implying that no hypothesis on OEE theory can fully be confirmed or rejected, and the results from the model can therefore not serve as conclusions, but only as indications of the relationships.
63
As discussed in section 7.4.1, the path coefficients in the OEE* structure seems to differ from the original theory of multiplication with equal weights, as seen in equation (1). Particularly, the relationship between availability loss and productivity differs from the originally formulated equation, as it is positive and yields a coefficient of 0.090. This indicates that an increase in availability loss will imply a higher productivity, indicating that a decrease in production time available will provide a higher OEE* measure. Immediately, this dependence relationship conflicts with the intuition that the more available production time, the higher utilization and effectiveness of the equipment. The two other components of OEE* have similar values and follow the expected structure, where an increase in performance loss and quality loss lead to a decrease in productivity. This implies that when the capacity of the equipment is maximized, and the number of non-conforming products minimized, OEE* increases. If the coefficient of availability had approximated the coefficients of performance and quality, the model would have indicated a confirmation of the OEE* structure. Even though this coefficient differs to a large degree from the other two components, it is not possible to reject Hypothesis III, stating that OEE* is composed by an multiplication of availability, performance and quality with equal weights, as only the relationship between performance loss and productivity is statistically significant at = 0.05. It must be noticed that an improvement in the identification of the model, according to the theoretical background for optimally measuring availability, performance, and quality, can potentially alter the path coefficients and thereby change the weights of the different components of OEE* estimated in the PLS model. Particularly, performance and quality are inadequately indicated, and even though this was given by the practical constraints of this study, there may be other and/or better variables reflecting these components. From the empirical research on OEE*, as stated in section 3.5, it is furthermore evident that many important indicators are not identified in the model, and these are likely to cause an omitted variables bias, as they are expected to have an effect on OEE*. Apart from the poor data basis, there is a risk that the transformation of the original OEE measure, in terms of inclusion of planned downtime and adjustments in the variables, reflecting the components, contributes to the fact that the original OEE structure differs from the path coefficients seen in Model 2, Figure 12. Even though the
64
objective is not to develop the optimal model, it has been necessary to modify the OEE into OEE* in order to accommodate the objectives of Arla Foods. 7.4.2.1 Background Variables In addition to the transformation of OEE into OEE*, it is preferable to consider background variables in an attempt to be able to explain the effectiveness of the equipment in a better way. This is substantiated by the intuition that changes in the number of product shifts, as well as the amount of cleaning of the spray tower, can explain changes in the effectiveness of the equipment. In Arla Foods, it has only been possible to consider these two background variables; though other variables seemed relevant, data on these measures was not possible to retrieve. The choice of background variables is dependent on the organization in question, the specific production site, as well as the equipment for which OEE* is measured. The PLS model indicates that the background variables are very important, as the path coefficients related to background loss are among the highest in the model. Referring to the previous discussion in section 3.3, many companies use OEE as a measure for benchmarking across time, equipment and the entire organization. Since the PLS model indicates that the background variables have substantial influence on both productivity and the three components, it is even more crucial to investigate the background variables affecting the production, before any benchmark is performed. 7.4.3 Validation of Hypotheses I, II, and III
Based on the interpretation of the PLS model, a recap of the validations of the formulated hypotheses is provided below. Hypothesis I, stating that effectiveness losses relate negatively to OEE, is not possible to reject, as no positive statistically significant relation at = 0.05 exists between the effectiveness losses and productivity. Hypothesis II, stating that OEE relates positively to financial performance, is not possible to reject, as the relation is characterised by a statistically insignificant loading of -0.078 at = 0.05. Hypothesis III, stating that OEE is the multiplication of availability, performance, and quality with equal weights, is not possible to reject, as only the performance loss and productivity is statistically significant at = 0.05.
65
In spite of the fact that it is not possible to reject the formulated hypotheses, due to the statistical insignificance, the structural equation model suggests a weak relationship between OEE and financial performance, and provides an indication of a falsification of the existing OEE theory, already questioned in the empirical research of OEE.
66
67
Ideally, the threshold should be based on the targets defined by Nakajima (1988) and other theorists, and adapted by a conversion factor, which considers the extra availability loss due to planned downtime. This conversion factor could be calculated, based on the level of scheduled downtime in Arla Foods compared to the total availability loss, however, due to practical constraints, it has only been possible to retrieve data on certain variables, of which most of the availability variables represent planned downtime. This approach would therefore be misleading, and the thresholds for availability and OEE* are therefore assumed to be 60% and 56% respectively.7 In Figure 15, the OEE* values, meeting the target of 56%, are shown in green, whereas the OEE* values, failing to meet the threshold, are marked with red. Overall, there is evidence that slightly more than half of the observations meet the prescribed standard. When the spray towers are grouped according to factory, it is seen that all production sites have OEE* values, which exceed or almost meet the threshold determined for successful equipment effectiveness. However, several observations somewhat below the desired target are also found, but many of these are possible to explain and the causes will be discussed in the following sections. 8.1.1 Analysis of OEE* at Factory Level
It is desired by Arla Foods to acquire an overview of the effectiveness of each spray tower, and the OEE* results at each production site are therefore briefly presented
Based on inspiration from other OEE studies and the knowledge obtained in the analysis of Arla Foods, it is assumed that availability loss consists of 75% planned downtime and 25% unplanned downtime. Nakajimas target rate of 90% entails that there is 10% unscheduled downtime. If planned downtime constitutes 75% of total downtime, the availability target is 60%, implying an OEE target of 56%.
7
68
below, whereas a detailed analysis of the differences between the individual spray towers and production sites is discussed in section 8.2. As seen in Figure 15, the two sprays towers at HOCO have different OEE* patterns. HOCO 2 has satisfactory OEE* measures almost every week, except in week 41, where OEE* is 0% due to discard of all products in this period as they did not meet the quality standards. HOCO 3 has several lower observations, which can be considered a result of fluctuations in the availability, performance and quality measures. AKAFA distinguishes itself from the rest of the production sites, with very satisfactory effectiveness results, exceeding 73% in all weeks. This is caused by consistently high availability, performance and quality measures. Both Vimmerby spray towers show several poor OEE* values, mainly caused by availability losses in terms of standstill. The first three weeks at Vimmerby 2 yields OEE* values of 0%, caused by total standstill, and the fourth week shows an extremely low value of 3%, mainly due to a high quality loss. The level of effectiveness at DP is dependent on the spray tower in question, as DP 1 and DP 3 almost meet the target level in the entire fourth quarter of 2009, with the lowest OEE* values being 52%, as opposed to DP 2, which OEE* values are almost all below the threshold. In general, the differences observed for DP, compared to the other production sites, should be interpreted in the light of the nature of the raw material weighed in, as the weighing-in of whey, as opposed to raw milk, may influence the results, even though the thin whey equivalents and raw milk are assumed to be 1:1, as discussed in section 5.2. The OEE* values at ARINCO are high with only two fairly low observations at ARINCO 5 standing out. These high values are derived from few availability losses, primarily due to the fact that the equipment is never in standstill. Overall, the level of OEE* is considered satisfactory as the OEE* initiative is only in the starting phase of the implementation, and despite the fact that almost half of the observations falls below the threshold, the measures serve as a good starting point, if measured and improved on a continuous basis. Since it is generally not advised to compare specific company studies to a given threshold, not considering product, organization and industry, Arla Foods should be careful to rely on the thresholds to
69
indicate satisfactory effectiveness. Effectiveness measures as OEE* should not be subjected to indiscriminate use of benchmarking of the current level to general threshold, but instead be used to monitor the development of each equipment over time and serve as a starting point for establishing best practise in Arla Foods GIN.
Benchmarking of the effectiveness of the spray towers is one of the main objectives of the practical analysis of the OEE* implementation, and an attempt to answer research question vi) will be done by formulating and validating the following hypotheses. Hypotheses Hypothesis IV: The OEE means across spray towers in the fourth quarter of 2009 are equal Hypothesis V: No segmentation of the spray towers exists The above hypotheses will be tested by means of an analysis of variance (ANOVA) in SPSS and a test of potential segments in FIMIX-PLS, respectively. 8.2.2 Repeated Measures ANOVA
In order to attempt to benchmark across the respective spray towers and determine the statistical significance of the potential differences, an analysis of variance (ANOVA) is performed as a way of comparing the means of the series. The idea behind ANOVA is to compare the variation within samples with the variation between samples. Since the same spray towers are observed over the same 12 weeks, the observations are not independent, and it is therefore chosen to apply the intuition behind the Repeated Measures ANOVA, which is robust to related samples, and the resulting correlation
70
from this lack of independence (Coolican 1999). Since the focus is to determine the variation between spray towers and partial out any variation due to time effects, it is attempted to modify the usual structure of a Repeated Measures ANOVA by allowing the spray towers to function as conditions and by considering the 12 weeks as subjects. In a repeated measures design the variation in each condition is related to the variation in other conditions. Rather than thinking of the variation in every condition as separate samples of variation, part of the variation in each condition is predictable from knowing the variation in other conditions, as it results from the differences between the same subjects. The Repeated Measures ANOVA allows testing of the statistical significance of the difference in the means of the measured variables, see assumptions in Appendix X, and will be the basis of the comparison of the 10 spray towers in Arla Foods GIN. In the following section, the results of the Repeated Measures ANOVA are analysed, however, only the statistically significant differences in the means between the spray towers are discussed. This implies that if the spray towers perform statistically equally well within a given parameter, it will not be commented on, however, all comparisons can be found in Appendix Y. An overview of all the significant differences between the spray towers is presented in Figure 16. Figure 16 Benchmarking across Spray Towers
71
It is chosen to begin the benchmarking by analysing the differences in the means of the measured variables and identify the causes for these differences, in order to understand the overall differences between the OEE* values of respective spray towers. In the following analysis of the mean differences, only the spray towers that are the common denominator in more comparisons are highlighted and discussed. The detailed differences between all relevant spray towers are presented in Appendix W. The three measured variables in the availability component, containing significantly different means across the spray towers, are CIP misc., setup and standstill. In addition, availability, performance and OEE* are characterised by significant differences. 8.2.2.1 CIP Misc. When assessing the differences in CIP misc., in general HOCO 3, AKAFA 2 and ARINCO 4 have substantially more CIP misc., as seen in Appendix W. At HOCO 3, CIP misc. is performed after every batch produced, and in several weeks, especially in the beginning of the quarter, CIP misc. was registered almost the entire day. The large amount of CIP misc. may partly be caused by the fact that only one cleaning of the tower is performed in the fourth quarter of 2009, and CIP misc. can maybe compensate for the rare tower cleaning. The production at AKAFA 2 is characterised by the same pattern every day and by CIP misc. lasting at least one hour after every batch produced. It can be argued that the time spent on CIP misc. at ARINCO ought to be even higher, since the CIP misc. measured at ARINCO has been reduced by one hour in each end of CIP misc., which is dedicated to setup.8 Despite this assumption, CIP misc. at ARINCO 4 stands out, as there are long periods of CIP misc. after each production. A potential reason for the long periods of CIP is that ARINCO produces many special ingredients products, for which it can be assumed that extra CIP is needed to ensure that no remains of the previous production affects the special products. It is furthermore seen that HOCO 2 and Vimmerby 1 and 2 in general have registered less CIP misc. This tendency is likely to be caused by the fact that these three spray towers have been out of production for long time periods in the fourth quarter, and therefore CIP is redundant.
The assumption of setup at ARINCO has been proposed by the authors and approved by Arla Foods GIN, and can be found in Appendix C.
72
8.2.2.2 Setup It is seen in Appendix W that in general AKAFA 2 and Vimmerby 2 spend significantly less time on setup, whereas ARINCO 4 has substantially more setup. At AKAFA 2, setup is performed before and after each production similarly to the rest of the production sites, however, AKAFA 2 distinguishes itself as the setups are very short, which can be explained by a very stable production, which entails that the equipment is not subject to many unforeseen events, causing it to be in need of additional setup. The setup time at Vimmerby 2 is defined as two hours of setup (one before and one after the production) in connection with changeover of products, and as previously mentioned, the setup time at ARINCO 4 has been defined as one hour in each end of a CIP misc.9 Since the data related to setup at these two spray towers has been assumed and not directly measured at the site, it is unfair to draw conclusions based on this data. However, the statistically significant differences in setup between these spray towers, indicate that the respective assumptions for setup time at ARINCO and Vimmerby have been over- and understated and should be estimated differently if measured again. Nevertheless, despite the imprecise estimates for setup time, these measures are still relevant to include in the analysis, as they are the best estimates available. 8.2.2.3 Standstill When assessing the results from the Repeated Measures ANOVA seen in Appendix W, it is evident that Vimmerby 1 and 2 have significantly more standstill than almost every other spray tower. The long non-productive times are likely to be caused by lack of feed, causing the spray towers to be unable to produce. It is important to notice that the production sites in Arla Foods are interdependent, as the feed to one plant is provided by another plant and is not always available. This implies that standstill in a specific site is not necessarily controlled by the site in question, but is dependent on feed from another plant. Low availability and correspondingly low OEE* values as a result of substantial standstill should therefore be interpreted carefully, as the loss in effectiveness cannot necessarily be attributed to the specific site. Furthermore, it is evident in the ANOVA that ARINCO has significantly less standstill, which was anticipated, as no standstill was recorded at ARINCO. As previously mentioned, it has been attempted to create similar data set for all five production sites,
The assumption of setup at Vimmerby has been proposed by the authors and approved by Arla Foods GIN, and can be found in Appendix C.
9
73
however, as the PTD manager at ARINCO maintains that there has been no standstill in the fourth quarter 2009, no assumptions are therefore formulated to overcome this issue. In addition, AKAFA 2 has significantly less standstill, except in the comparison with ARINCO where AKAFA 2, though, has slightly more standstill. 8.2.2.4 Availability As anticipated, it is primarily Vimmerby, which distinguishes itself from the rest of the production sites in terms of the availability of the equipment as seen in Appendix W. Considering the discussions above and the fact that Vimmerby has less CIP misc. and setup, there is evidence that the large periods of standstill is the main reason for the availability loss at Vimmerby. In addition, it is seen that there is significant difference between the availability at AKAFA 2 and DP 2, which appears to be caused by the fact that AKAFA 2 has significantly less standstill than DP 2. Even though there is evidence of significant differences between the spray towers in three of the availability variables, it appears that the effect of some of these differences is neutralized, when looking at availability over all. 8.2.2.5 Performance When assessing the differences in the means of the performance component, DP 2 stands out by being significantly different from five of the nine other spray towers. In relation to the maximum capacity possible, when producing a specific product, DP 2 underperforms relative to the other towers. It appears that there is significant difference in the output per time unit produced by DP 2, indicating an unstable production. This would entail that the maximum capacity registered in the fourth quarter might be unrealistically high relative to the actual output produced. Furthermore, there is a difference in the performance component between HOCO 2 and HOCO 3. This can be explained by the extent of the product portfolio produced at the respective spray towers, and the measurement technique of performance, which is the amount of products actually produced relative to the maximum output possible to produce. If one product type is produced often during the fourth quarter 2009, there is a risk of higher variation of the performance within one product type. On the other hand, if a product is only produced once in the fourth quarter 2009, the performance rate will automatically be 100%. As previously discussed, this method for calculating
74
performance is therefore not ideal for all production sites; however, it is considered the best estimate. 8.2.2.6 OEE* Overall, the differences discussed in the above trigger the same patterns and tendencies in OEE* as observed for the individual variables. As seen in Figure 17, all production sites differ from at least one other plant in terms of OEE*, which substantiates the discussion that the production sites operate differently. This leads to a rejection of Hypothesis IV, stating that the spray towers have equal OEE means, already indicated in the econometric analysis through significant fixed effects and factory dummies. Figure 17 OEE* Mean Differences
When merely looking at the OEE* values in Figure 15, it is clear that AKAFA 2 is very effective with high values above 73%. In the benchmarking of the OEE* values, AKAFA 2 is statistically more effective than the spray towers Vimmerby 1 and 2 and HOCO 3, which is most likely caused by the high level of availability at AKAFA 2 compared to the availability loss due to standstill at Vimmerby and HOCO. It must be mentioned that one of the reasons for standstill at HOCO 3 is the fact that HOCO 2 is prioritized, when there is lack of feed.10 Furthermore, AKAFA 2 has a significantly higher OEE* value than DP 2 and 3 due to significantly better ability of operating at maximum capacity.
10
75
As anticipated, the two spray towers at Vimmerby have lower effectiveness due to the availability loss caused by standstill. These low OEE* values should not be given too much weight, cf. the discussion of lack of feed, as the standstill to a large degree is uncontrollable. The low effectiveness at Vimmerby should rather trigger an increased focus on the interdependence of the production plants at a higher corporate level. DP 2 is characterised by lower OEE* values than DP 1 and ARINCO 4 and 5, which is primarily caused by the performance loss, as discussed in the previous section. Furthermore, DP 2 shows several availability and quality values below the defined threshold, so even though these differences are not as outspoken as for performance, they still contribute to a lower OEE*. The effectiveness at ARINCO is relatively higher due to high availability, primarily as a result of no registered standstill, which outweighs the relatively poor performance and quality observations. As previously discussed, it is evident that benchmarking across sprays and production sites cannot be performed without careful considerations of the differences in the conditions, under which they operate. Since the production sites are subject to different constraints in terms of different products and different equipment, which cannot be altered in the short run, and the defined variables are not measured in the same way at every site, benchmarking must not be used indiscriminately. However, benchmarking, when not seen in isolation, is useful for breaking down the OEE* values in order to identify the underlying causes of effectiveness losses, and the findings will constitute the basis for the recommendations for further roll-out of OEE* in Arla Foods. 8.2.3 FIMIX-PLS Segmentation
As the 10 spray towers stem from five production sites, and the econometric analysis indicates that unobserved effects exist, there is reason to believe that there are latent classes in the PLS estimation, and FIMIX-PLS segmentation is therefore performed. The purpose of the segmentation is not to force the observations into arbitrary groups by only basing the groupings on the FIMIX-PLS segmentation and thereby ignoring intuition. Instead it is attempted to explore the tendencies previously identified in the FE estimation and Repeated Measures ANOVA, such as groupings based on week, production site, or OEE* level. FIMIX-PLS is a tool for modelling latent class structures by uncovering groups of data with similar characteristics (Ringle 2006).
76
FIMIX-PLS allows for capturing heterogeneity for the estimated variable scores in path models by segmenting observations, based on maximum likelihood estimation of the probabilities that certain observations belong to specific groups (Ringle, Wende and Will). Since there is evidence in the econometric analysis that there are unobserved effects for the respective production sites, FIMIX-PLS segmentation is initially performed for k = 5 groups. As seen in Appendix Z, the segmentation into five groups makes little sense as segment 4 and 5 contain very few observations. This is substantiated by the regression with factory dummies in the econometric analysis, which reveals that there are not significant differences between all production sites, as seen in Appendix L. It is therefore chosen to continue with FIMIX-PLS with k = 3 segments, which immediately appears more reasonable with all three groups, including several observations identified by initial partition in the SmartPLS output, as seen in Appendix Z. Nevertheless, FIMIX-PLS with k = 2 segments is additionally attempted, and since this yields two equally sized groups, this segmentation could also be interesting. In order to determine which segmentation to proceed with, the quality criterion Consistent Akaike Information Criterion (CAIC) presented in Figure 18 is evaluated. Figure 18 Fit Criteria of FIMIX-PLS Segmentation
The ideal number of segments can be identified by minimizing the quality criterion, which incorporates a penalty for including more segments. It is decided to continue with three segments, since this yield the overall lowest quality criterion of CAIC = 1452.96. The Entropy (EN) statistic, ranging from 0 to 1, indicates the degree of separation estimated in the individual class, and values above 0.5 indicate unambiguous membership probabilities. As seen in Figure 18, the 0.5 threshold is achieved for all FIMIX-PLS segmentations, indicating a high quality of separation. After having identified three seemingly relevant segments, it makes sense to compare these groups with the underlying intuition, as arbitrary groups unjustified by week,
77
production site, or OEE* level are invaluable in the attempt to benchmark in Arla Foods GIN. It is therefore attempted to find a pattern in the segmentation based on week, production site, or OEE* level, so the grouping of the observations is no longer arbitrary, but substantiated by an intuitive pattern. The individual observations are therefore inserted in a matrix attempting to identify a pattern across spray towers or weeks, but as seen in Appendix AA, no obvious pattern occurs and the segmentation appears to be arbitrary. Furthermore, it seems relevant to analyse whether the OEE* level is determinant for the segment, to which an observation belongs, as this can potentially affect the inner relations in the PLS model. In order to test whether the OEE* level is significantly different between the three segments, a One-Way ANOVA is evaluated as seen in Appendix CC, and the assumptions for One-Way ANOVA are discussed in Appendix BB. It is evident that there is no statistically significant difference between the OEE* levels in the respective groups, and it can therefore be concluded that a segmentation according to the OEE* level does not exist. The above analysis reveals that Hypothesis V must be rejected, as FIMIX-PLS suggests segmentation into three groups. However, it appears that the segmentation into three groups is arbitrary and does not comply with the underlying intuition of possible segments. It is therefore chosen to disregard any segmentation of observations based on FIMIX-PLS, and benchmarking should only be attempted based on the knowledge on the differences between the spray towers as well as production sites from the Repeated Measures ANOVA. 8.2.4 Validation of Hypotheses IV and V
Based on the above Repeated Measures ANOVA and FIMIX-PLS segmentation, a recap of the validation of Hypothesis IV and V is provided below. Hypothesis IV, stating that the OEE means across spray towers in the fourth quarter of 2009 are equal, is rejected as the Repeated Measures ANOVA shows several differences between spray towers, statistically significant at = 0.05.
78
Hypothesis V, stating that no segmentation of the spray towers exists, is rejected, as the FIMIX-PLS segmentation suggests that groupings of both two and three segments are theoretically feasible.
The primary gain is an increased transparency in the effectiveness of the equipment, enabling an understanding of the utilization and the reasons for non-productive time. This is essential in capital-intensive production sites, as it is expected that the slightest downtime will have a substantial effect on the financial performance, even though the PLS model is ambiguous in this respect. Since the non-productive time is decomposed by OEE*, it is possible to pinpoint the actual causes instead of merely accepting the presence of downtime, which creates the breeding ground for remedying the causes and thereby reducing the non-productive time. Furthermore, the focus on OEE* forces a reflection on losses, caused by reduced speed in production or non-conforming products, which is necessary in order to take action to improve the equipments ability to constantly produce at maximum capacity with minimized defective rate. In particular, it is important to obtain knowledge of the losses caused by reduced performance, since Arla Foods GIN prior to the OEE* initiative was, and to some extent still is, unaware of the maximum capacity of the equipment, which renders it impossible to determine, whether the equipment is producing at satisfactory speed. Even though the performance component of OEE* does not guarantee complete insight in the maximum capacity, the OEE* initiative indicates fluctuation in the performance rate and deviations from the defined potential maximum. 8.3.2 Best Practice
One of the additional benefits of the decomposition of OEE* is the ability to identify best practice within the respective types of effectiveness losses. Even though benchmarking must be interpreted with care, as there are many factors affecting OEE* that are not controllable by the respective production sites, comparing the spray towers
79
enables an identification of maximum utilization of the equipment, and serves as a learning platform for the underperforming production sites. In spite of the differences between the spray towers, there are many similar traits, which make it possible to determine and communicate best practices throughout the production sites, thereby achieving a higher overall utilization of the equipment in Arla Foods GIN. In order to determine the best practices, a comprehensive understanding of the production sites, and the constraints under which they operate, is essential. However, the analysis in section 8.2 reveals that a number of the spray towers distinguish themselves on a number of different parameters, and it is the causes of these characteristics, which serve as the basis for the best practice. Obviously, it is meaningless to attempt to identify best practices for the parameters uncontrollable by the production sites, however, it is still desired to understand the causes behind these factors and minimize the resulting losses. Furthermore, even though some factors do not yield significant differences between the spray towers, this does not entail that all sites produce in accordance with best practice, and overall optimization may therefore be justified. With the previously mentioned reservations regarding benchmarking in mind, it is seen that AKAFA 2 excels in several ways, cf. section 8.2, why it would be natural to begin with studying AKAFA 2, when attempting to identify the fundamental basis for future effectiveness. If the necessary measures are taken in regards to the limitations of internal benchmarking, comparisons between production sites may result in sound, internal competition. An OEE* implementation is therefore likely to stimulate an increased effort, derived from the aspiration to excel and the fear of underperforming relative to other production sites. 8.3.3 Concerns of OEE* Implementation
The main concern in relation to the OEE* initiative is the substantial amount of resources needed, if a successful implementation throughout Arla Foods should be possible, both in terms of manpower and technical measurement system in the implementation stage, as well as in the continuous attention necessary. Furthermore, the implementation is complicated by the fact that it is not possible for a single team to be responsible for the OEE* implementation, as comprehensive knowledge of all
80
production processes is necessary, and no individual staff member is likely to have the necessary thorough understanding of the various aspects across departments. Another important disadvantage regarding the OEE* initiative in Arla Foods GIN is the constraints associated with the critique of OEE discussed in section 3.3. Even though external influences are necessary to accept in OEE*, the measure is intended for individual stand-alone equipment, which indicates that all OEE* components must be measured only for the spray tower. This places high demands on the data collection, since it must be ensured that the data is measured in such a way that losses are attributed to the right equipment. Furthermore, in order to be able to monitor the development in OEE* over time, it is important that the components are always measured identically, and if benchmarking, in spite of all red flags, is attempted, it is essential that the data collection is homogeneous across sites. This harmonization, within each production site and across Arla Foods, demands time and resources. Even though it has been possible to pinpoint several benefits of as well as concerns with the OEE* initiative, the actual concrete value of implementing OEE* is difficult to determine due to its intangible nature. It is therefore attempted to quantify the value by establishing a clear relationship between OEE* and the financial performance of the organization in the PLS model. Even though such a relationship intuitively exists, this relationship is difficult to confirm in practice, as increased effectiveness of the equipment is likely to have a positive effect on revenue but ambiguous effects on costs. Considering the current state of OEE* and the inability to quantify the value of the OEE* implementation, it is uncertain whether the benefits outweigh the costs associated with the current implementation of OEE* Arla Foods GIN, and it is questionable whether today OEE* is a useful, value-creating KPI in the organization. However, as potential for improvement exists, recommendations for further OEE* roll-out in Arla Foods are provided in the following.
81
8.4.1
Harmonization
One of the key success factors for a successful OEE* project is harmonization of the production sites in Arla Foods, as the current conditions in the organization are characterised by the fact that Arla Foods is a merger of a number of smaller cooperative dairies, which all have their individual purpose, structure and routines. It is essential to ensure a common mindset and eliminate the isolated thinking, currently present at the production sites, so optimization of the overall production can be achieved as opposed to sub-optimizing the individual production sites. Such an alignment is necessary to exploit the advantages of any KPI implementation, and therefore also to achieve a successful OEE* execution. The production sites must be harmonized in terms of a common ERP system, which enables the production sites to log their operations identically, creating transparency throughout the organization. It is important that the processes are comparable, independently of the plant in question, and standard procedures for CIP, maintenance, repair etc. must be defined based on best practice, taking the constraints on the individual site into account. Through a harmonization of the production sites, better and more comparable measures of the OEE* components can be obtained. The current data basis does not allow for strict comparisons and benchmarking due to the differences in the registering and measuring of data. However, if it is defined how, when and where the variables are to be measured, the possibilities for meaningful benchmarking and comparison are improved. Eventually, it will be possible to easily monitor and compare the development in the effectiveness of the respective spray towers, if observations are logged directly in the same IT system, which must consider the individual constraints of each production site. Such an IT system is already in the process of being finalised at Arla Foods, but seems to stem from a top-down decision of the design of the OEE* initiative. It has become evident that this system does not take the various limitations of the plants into account11, and such a system is therefore difficult to implement in the current stage of the production sites. In order for such a system to be effective, it is necessary that there is buy-in from the users of the system in the process of designing it.
11
Discussed at the OEE workshop with PTD managers and OEE* Project Manager from Arla Foods.
82
8.4.2
Not only buy-in in the technical aspects of the OEE* initiative is important, but general creation of ownership throughout the organization is crucial, and both with regards to the frontline workers, the PTD managers at the respective production sites, the OEE* Project Manager, and corporate Arla Foods. One of the most important observations made, when collecting data at the respective production sites, is the fact that in the current stage of the OEE* project, the frontline workers feel little association to and responsibility for the OEE* initiative. Employee involvement is crucial, when attempting to implement new initiatives, which have important impact on the daily routines, and the project will not be successfully executed, if frontline workers are not included in the decision process (Liker 2004). Ownership at all levels can be achieved by incorporating top-down and bottom op approaches to ensure both strategic alignment and grassroots buy-in to the initiative (Eckerson 2009). Since it seems reasonable to assume that little employee involvement exists, due to the many different projects initiated simultaneously, causing confusion and inability to take ownership and meet all objectives at once, it is necessary to focus the effort, prioritize the resources and make a clear distinction between must win battles and merely potential optimization initiatives. 8.4.3 Bottlenecks and Interdependence of Production Sites
If successful roll-out of OEE* in all parts of the production of ingredients is achieved, it is possible to identify bottlenecks and enable a visualization of the critical part of the production chain. Currently, the spray towers are considered to be the constraining part of the production chain, however, it seems relevant to consider the possibility that other parts of the production chain also constitutes bottlenecks, since three out 10 spray towers have registered substantial amounts of standstill in the fourth quarter of 2009. In the short run, standstill can be difficult to eliminate, as in most cases it is due to lack of feed, which is uncontrollable by the production sites. However, in the longer run standstill can potentially be minimized by controlling the interdependence and considering the production sites as one collaborative unit, which should be optimized in its entirety, in order to ensure that production is levelled out across sites in critical periods, so no spray towers are unable to meet the required capacity while others are in standstill. It is important to acknowledge that due to the different nature of the raw material required and the product portfolio, the fluctuations are not expected to be
83
completely eliminated, but a smoother production flow could remedy some of the variation. Furthermore, it must be recognized that a spray tower in standstill several weeks in a row is an indication of low utilization, and it must therefore be recommended to analyse the actual supply and demand in order to discover whether all spray towers can be justified. 8.4.4 Data Basis
The OEE* measured in the fourth quarter 2009 is based on an insufficient data basis, and is therefore not detailed enough to truly picture all the nuances in the effectiveness losses. In order to gain the full benefit of an OEE* implementation, it is essential to incorporate all the variables defined in section 4.2, which are likely to affect the effectiveness of the equipment. At the current stage of the OEE* implementation, the various production sites are unable to register the necessary detail level, and it is therefore not possible to break down the OEE* components to a larger extent. There is therefore substantial uncertainty associated with the current OEE* values, as many unmeasured variables are expected to influence the effectiveness of the spray towers. Ideally, all the identified variables would be measured in the computation of OEE* at higher detail level, so it would be possible to pinpoint the exact reason for e.g. breakdown. Both intuitively and statistically it is essential to incorporate additional variables in OEE*, so a sufficient detail level can be obtained, and the causes behind the effectiveness losses can be determined and remedied. In addition to including all relevant variables in the OEE* measure, it would be valuable to incorporate additional background variables to enhance the understanding of OEE*, as the analysis in this thesis indicates that the number of CIP (NC) and product shifts (PS) have substantial explanatory power. It is therefore suggested to consider the possibility of defining separate OEE* targets according to the implication of the background variables. For instance, in the future it can be relevant to consider the OEE* level in relation to the maturity of the OEE* implementation, as it is well known that the maturity of TQM initiatives affect the level of performance (Hendricks and Singhal 2001). 8.4.5 Cost Perspective
The use of background variables in the PLS model indicates that OEE* cannot be evaluated in isolation. Another external aspect, which seems necessary to consider in
84
connection with OEE*, is the incorporation of a cost perspective, as mentioned in section 3.3 as one of the points of critique of OEE. This is particularly relevant in connection with the trade-offs between maintenance and breakdowns as well as between quality and production speed. It is important to make a conscious decision as to when to allow for repair of unforeseen breakdowns versus when to attempt preventing breakdowns by means of maintenance. Furthermore, there is a delicate balance between attempting to maximize the production speed without compromising the quality. It is therefore important that the production sites determine the optimal balance between these two components from a cost perspective, so it is possible to aim at the right production speed. Moreover, it would be valuable to be able to identify the severity of the quality loss, as the financial impact is assumed to differ, depending on whether the non-conforming products are scrapped or possible to sell at concession. If no cost perspective is incorporated, there is a risk that OEE* is wrongfully interpreted, and undesirable decisions based on inadequate information are made. 8.4.6 Must Win Battles
An additional important recommendation for Arla Foods in relation to the OEE* project is to consider and prioritize the interrelated initiatives, which are currently being implemented across the organization, by distinguishing between must win battles and secondary performance initiatives. At the moment, Arla Foods is in the process of developing an organization-wide performance management system including a dashboard of KPIs, just as a fundamental Lean mindset is introduced12, towards which OEE* to some degree has conflicting objectives. The introduction of Lean is intended to optimize the entire value chain, whereas the OEE* initiative focuses on optimization of the effectiveness of the individual equipment. From a Lean viewpoint, there is a risk that OEE* sub-optimizes parts of the production and thereby hinders an optimal Lean implementation. Whereas Lean covers many relevant and coherent aspects of the continuous improvement of production, OEE* attends to a more specific fragment, disregarding possibly important considerations. Since it is given that Arla Foods plans to roll-out OEE* throughout all production units, the purpose of this analysis is not to determine whether OEE* should be implemented, but rather to attempt to increase the possibility of a successful implementation. The
12
Knowledge gained from meetings with OEE Project Manager at Arla Foods GIN
85
above suggestions and recommendations are therefore intended to improve OEE* as a valuable effectiveness measure, which can fulfil the purpose intended by Arla Foods, when accepting the limitations in the possibilities for benchmarking. It is evident that OEE* is very valuable in terms of visualizing the current utilization of the equipment, as well as the current performance in terms of capacity and quality. However, OEE* only becomes truly value-creating, when Arla Foods determines how they intend to proceed with the knowledge gained from the OEE* results and take the corresponding appropriate action to increase the effectiveness. Arla Foods must therefore acknowledge the actions necessary to take, to reduce the effectiveness losses identified in the analysis, otherwise the resources spent on improving the OEE* measure and implementing the KPI may be in vain.
86
9 Further Studies
It has been the objective of the theoretical part of the present thesis to make an assessment of the equipment effectiveness, OEE. Even though a number of problems have constrained the ability to draw substantive conclusions on the findings, it is indicated that the current components and structure of OEE are questionable and interesting to investigate further. In future studies, it is therefore recommended to accommodate the issues that have not been possible to remedy in the present thesis, in order to achieve results, which can challenge the existing OEE theory and be generalized to a wider context. The assessment of OEE in this study is compromised by the fact that it is based on OEE* in Arla Foods GIN, which e.g. includes planned downtime, as opposed to OEE, and it is therefore interesting to apply the present structural equation model only on variables, originally contained in the theoretical OEE. If an adequate data basis was present, it would allow for the exclusion of planned downtime, as there would be a sufficient number of variables indicating the down time loss and thereby the availability component. This implies that it would in fact be the multiplication of availability, performance, and quality with equal weights that is scrutinized, enabling the findings to directly challenge and improve the existing OEE theory. Additionally, the twofold objective of the present thesis has entailed that the implications of the longitudinal data have been disregarded in the structural equation model, as the practical aspect dominates the theoretical problem. However, in future studies, the issues of correlated error terms as a consequence of the longitudinal data structure can be remedied by inclusion of dummy variables and time lags. The exact structural equation model for remedying the cross-sectional and longitudinal correlation is expected to be complex in structure as well as interpretation, due to the efforts required to suppress the correlation in the econometric analysis. It has therefore not been attempted to construct such a model in the present thesis, however, inspiration for future model designs can be found in the multiple-lag models and models with background variables suggested by Bagozzi and Yi (Bagozzi 1994). Multiple-lag models allow a variable to have effects for more than one subsequent period, and are expected to accommodate the autocorrelation identified in the econometric analysis. Incorporation of background variables, representing the spray and week effects in the
87
structural equation model, allows for a reduction in the omitted variable bias, which causes the error terms to correlate. It is expected that inclusion of several latent constructs is necessary to suppress the issues of correlated error terms completely, and it must therefore be assessed to which extend the correlation should be remedied, as the benefits of a statistically well specified model, must outweigh the disadvantages associated with the complex structure and interpretation. In future research, it must be kept in mind that there is a tradeoff between the theoretically ideal model, which is complex, but accommodates all issues, and a model, which serves the practical purpose of the OEE measure, but fails to satisfy all criteria.
88
10 Conclusion
The overall objective of the thesis is twofold: the theoretical part testing the components and structure of OEE, and the practical part analysing the implementation of OEE* in Arla Foods. The theoretical and empirical research has made it possible to identify the variables, which best depict OEE*. The availability component encompasses a number of detailed measures of reductions in the productive time, reflecting both the planned and the unplanned downtime. The loss in performance effectiveness has proven difficult to identify, as the reasons for speed loss are hard to quantify, and the performance loss is therefore measured as the rate of the actual performance to the maximum performance possible, though this optimum is often unknown. The quality rate ideally encompasses the different types of quality losses, which are characterised by different degrees of severity, and therefore should be handled individually. Common for the identified variables are high levels of detail and accuracy, as well as a comprehensive sample size, to ensure a valid and reliable study. The structural equation model reveals that the relations between the effectiveness losses and productivity are not statistically significant, but it is indicated that the performance and quality components affect OEE* positively, as expected. However, availability appears to relate negatively to OEE*, which contradicts the intuition that downtime decreases equipment productivity. This discrepancy may be explained by the idea that the equipment is more effective after downtime losses have occurred, meaning that the machinery operates more effectively after e.g. cleaning or repair of a breakdown. Furthermore, the analysis of the structural equation model has confirmed that a relationship between OEE and financial performance is difficult to establish. This is caused by the fact that financial performance is affected by numerous factors, for which data has not been possible to obtain, implying that the analysis has been based only on cost variables. However, the PLS model indicates, though not with statistical significance, that a negative relationship between OEE* and financial performance exists, which may be explained by the ambiguous effect of the cost variables. As already indicated by the empirical research, it is questionable whether the multiplication of availability, performance, and quality with equal weights is the best
89
expression of OEE. In spite of the statistically insignificant path coefficients in the structural equation model, it is indicated that availability, performance, and quality should not be multiplied with equal weights to best express OEE*, as the availability component appears to have a smaller effect than the other components and relate negatively on OEE*. The effectiveness of the spray towers in Arla Foods GIN fluctuates substantially in the fourth quarter of 2009. When evaluating the OEE* results according to the defined OEE* threshold of 56%, it is evident that approximately half of the observations, on the 10 spray towers during the 12 weeks, meet the target. It can therefore be argued that the effectiveness of the spray towers in Arla Foods GIN is currently not satisfactory, and that the production sites should therefore strive for improvements. The benchmarking of equipment effectiveness has proven difficult due to the nature of OEE* and the data basis of the present study, however, it is attempted to compare the effectiveness of the spray towers in Arla Foods GIN by means of a Repeated Measures ANOVA. It can be concluded that OEE* differs across spray towers as well as production sites, and the underlying reasons behind the different OEE* means are identified based on the tendencies in the measured variables and effectiveness components. Since the analysis has indicated that differences in the OEE* results exist, a FIMIX-PLS segmentation has been performed to discover if latent groups can be formed. It is revealed that segmentation into three groups is statistically feasible, however, since this grouping cannot be substantiated by an intuitive pattern based on week, production site, or OEE* level, it is not reasonable to segment the spray towers. The analysis of the applicability of and value of implementing OEE* in Arla Foods indicates that OEE* is a valuable key performance indicator, provided that the concerns of the OEE* initiative, in terms of the empirical critique points of OEE and the resources needed for the implementation, are taken into account. The implementation of OEE* increases the transparency of the utilization of the spray towers in Arla Foods GIN, and provides an understanding of the reasons behind the reduced effectiveness, which enables the improvement activities necessary to eliminate the effectiveness losses. Furthermore, the OEE* implementation sets the grounds for an identification and communication of the best practices in Arla Foods GIN, and drives the production sites to improve the aspects of the equipment utilization, which are currently below target.
90
In order to enhance the success of the OEE* implementation in Arla Foods, it is recommended that the company strives to harmonize the production sites, so a common mindset, as well as homogeneous structures and routines, can be ensured and any suboptimization eliminated. In addition, it is essential that buy-in and ownership of the OEE* initiative is created at all levels to secure the necessary commitment to the common objective throughout the organization. In order to optimize the OEE* implementation, it is recommended that Arla Foods identifies whether, in fact, the spray towers are the bottlenecks in the organization, as well as, assesses how the interdependence of the production sites should be controlled. In addition, one of the basic foundations for a valuable OEE* measure and successful implementation is a comprehensive and detailed data basis, which allows for a consistent monitoring and benchmarking of the equipment effectiveness. It is, though, important to acknowledge that the monitoring and evaluation of OEE* cannot be performed in isolation, and that a cost perspective in particular is relevant to assess in connection with OEE*. Finally, due to the many activities currently initiated in Arla Foods, it is crucial to identify the must win battles, so the improvement activities can be prioritized and the resources distributed accordingly. Overall, the thesis is unable to validate the current theoretical specification of the components and structure of OEE. Furthermore, it is concluded that OEE* is a valuable indicator of equipment effectiveness in Arla Foods, provided that the initiatives recommended in the thesis are considered in the current implementation of OEE* in Arla Foods GIN as well as in future roll-out in the entire Arla Foods.
91
11 Literature List
Amaratunga, D. & Baldry, D. 2002, "Moving from performance measurment to performance management", vol. 20, no. 5/6, pp. 217-223. Bagozzi, R.P. 1994, Advanced methods of marketing research, Repr. edn, Blackwell, Cambridge, Mass. Bourne, M., Mills, J., Wilcox, M., Neely, A. & Platts, K. 2000, "Designing, implementing and updating performance measurement systems", International Journal of Operations & Production Management, vol. 20, no. 7, pp. 754-754. Brinberg, D. & Hirschman, E.C. 1986, "Multiple Orientations for the Conduct of Marketing Research: An Analysis of the Academic/Practitioner Distinction", Journal of Marketing, vol. 50, no. 4, pp. 161-173. Brinberg, D. & McGrath, J.E. 1989, Validity and the research process, Sage Publications, Newbury Park, Calif. Cai, L. & Lee, T. 2009, "Covariance Structure Model Fit Testing Under Missing Data: An Application of the Supplemented EM Algorithm", Multivariate Behavioral Research, vol. 44, no. 2, pp. 281-304. Coolican, H. 1999, Research methods and statistics in psychology, 3. ed. edn, Hodder & Stoughton, London. Dal, B., Tugwell, P. & Greatbanks, R. 2000, "Overall equipment effectiveness as a measure of operational improvement", International Journal of Operations & Production Management, vol. 20, no. 12, pp. 1488. De Ron, A.J. & Rooda, J.E. 2006, "OEE and equipment effectiveness: an evaluation", International Journal of Production Research, vol. 44, no. 23, pp. 4987-5003. Durbin, J. & Watson, G.S. 1950, "Testing for Serial Correlation in Least Squares Regression I", Biometrika, vol. 37, no. 3-4. Eckerson, W.W. 2009, "Performance Management Strategies: How to Create an Deploy Effective Metrics", . Ferreira, A. & Otley, D. 2009, "The design and use of performance management systems: An extended framework for analysis", Management Accounting Research, vol. 20, no. 4, pp. 263-282. Fornell, C. & Bookstein, F.L. 1982, "Two Structural Equation Models: LISREL and PLS Applied to Consumer Exit-Voice Theory", Journal of Marketing Research (JMR), vol. 19, no. 4, pp. 440-452. Guba, E.G. 1990, The paradigm dialog, Sage Publications, Newbury Park, Calif.
92
Hair, J.F. 2005, Multivariate data analysis, 6. ed. edn, Pearson Prentice Hall, Upper Saddle River, N.J. Harris, D. 2009, "OEE: achieving real-time process visibility. (cover story)", Plant Engineering, vol. 63, no. 11, pp. A5-A6. Hendricks, K.B. & Singhal, V.R. 2001, "Firm characteristics, total quality management, and financial performance", Journal of Operations Management, vol. 19, no. 3, pp. 269-1285. Henseler, J., Ringle, C.M. & Sinkovics, R.R. 1994, "The use of partial least squares path modeling in international marketing", . Jeong, K.-. & Phillips, D.T. 2001, "Operational efficiency and effectiveness measurement", International Journal of Operations & Production Management, vol. 21, no. 11, pp. 1404-1416. Kaplan, R.S. & Norton, D.R. 2005, "The Balanced Scorecard: Measures That Drive Performance. (cover story)", Harvard business review, vol. 83, no. 7, pp. 172-180. Keegan, D.P., Eiler, R.G. & Jones, C.R. 1989, "Are your performance measures obsolete?", , no. June, pp. 45-50. Keselman, H.J. 1982, "Multiple Comparisons for Repeated Measures Means", Multivariate Behavioral Research, vol. 17, no. 1, pp. 87. Kzdi, G. 2003, "Robus Standard Error Estimation in Fixed-Effects Panel Models", . Lebas, M.J. 1995, "Performance measurement and performance management", International Journal of Production Economics, vol. 41, no. 1-3, pp. 23-35. Liang, J. & Pan, W.S.Y. 2007, "ANOVA for Business Data: What Can You Do If the Statistical Conditions Are Not Satisfied?", Journal of Global Business Issues, vol. 1, no. 2, pp. 205-215. Liker, J.K. 2004, The Toyota Way, McGraw-Hill. Ljungberg, O. 1998, "Measurement of overall equipment effectiveness as a basis for TPM activities", International Journal of Operations & Production Management, vol. 18, no. 5, pp. 495. Miller, D.M. 1984, "Profitability=productivity + price recovery", Harvard business review, vol. 62, no. 3, pp. 145-153. Muchiri, P. & Pintelon, L. 2008, "Performance measurement using overall equipment effectiveness (OEE): literature review and practical application discussion", International Journal of Production Research, vol. 46, no. 13, pp. 3517-3535. Nakajima, S. 1988, Introduction to TPM: Total productive maintenance, Productivity Press, Cambridge,Mass.
93
National Research Council 1979, Measurement and Interpretaion of Productivity, . Netemeyer, R.G., Bearden, W.O. & Sharma, S. 2003, Scaling Procedures: Issues and Applications, Sage Publications, United States. Parmenter, D. 2010, Key Performance Indicators (KPI) : Developing, Implementing, and Using Winning KPIs, 2nd edn, Wiley, Hoboken. Pophaley, M. 2010, "Revisiting OEE as an Assessment Methodology for TPM Activities: A Practical Analysis", IUP Journal of Operations Management, vol. 9, no. 1, pp. 35-42. Procurement Executives' Association 1999, , Guide to a Balanced Scorecard PErformance Management Methodology [Homepage of Procurement Executives' Association], [Online]. Raouf, A. 1994, "Improving Capital Productivity through Maintenance", International Journal of Operations & Production Management, vol. 14, no. 7, pp. 44-52. Rich, N. 1999, Total Productive Maintenance: The Lean Approach, 1st edn, Tudor Business Publishing Limited, Great Britain. Ringle, C.M. 2006, "Segmentation for Path Models and Unobserved Heterogeneity: The Finite Mixture Partial Least Squares Approach", . Ringle, C.M., Wende, S. & Will, A. "Customer Segmentation with FIMIX-PLS", . Shahin, A. & Mahbod, M.A. 2007, "Prioritization of key performance indicators: An integration of analytical hierarchy process and goal setting", International Journal of Productivity & Performance Management, vol. 56, no. 3, pp. 226-240. Simons, R. 2000, Performance Measurement & Control Systems for Implementing Strategy, Prentice Hall, Upper Saddle River, N.J. Smith, A.M. & Hinchcliffe, G.R. 2006, "Calculating availability to develop preventative maintenance plan", Plant Engineering, vol. 60, no. 3, pp. 41-42. Stehrer, R. & Woerz, J. 2009, "Industrial Diversity, Trade Patterns and Productivity Convergence", Review of Development Economics, vol. 13, no. 2, pp. 356-372. Taguchi, G. & Clausing, D. 1990, "Robust Quality", Harvard business review, vol. 68, no. 1, pp. 65-75. Tangen, S. 2005, "Demystifying productivity and performance", International Journal of Productivity & Performance Management, vol. 54, no. 1, pp. 34-46. Tenenhaus, M., Vinzi, V.E., Chatelin, Y.-. & Lauro, C. 2005, "PLS path modeling", Computational Statistics and Data Analysis, vol. 48, no. 1, pp. 159-205.
94
Tenenhaus, M., Vinzi, V.E., Chatelin, Y.-. & Lauro, C. 2004, "PLS path modeling", Computational Statistics & Data Analysis, vol. 48. Turkcan, A. 1999, "Machine Scheduling with Availability Constraints", . Wooldridge, J.M. 2006, Introductory Econometrics - A Modern Approach, Third edn, Thomson South-Western, United States.
Authors Anne Sophie Rahbek Hansen MSc. in Finance and IB Christina Bkkelund Breum MSc. in BPM
Structural Equation Modeling Approach to Overall Equipment Effectiveness in Arla Foods APPENDICES
Aarhus School of Business, Aarhus University Department of Marketing and Statistics August 2010
List of Appendices
Appendix A Optimal Measurement of OEE* in Arla Foods and Arla Foods GIN ....... 1 Appendix B Definitions of Spray Towers ..................................................................... 2 Appendix C Assumptions for OEE* Data ..................................................................... 3 Appendix D List of Variables ........................................................................................ 7 Appendix E Complete Data Set ..................................................................................... 8 Appendix F Descriptive Statistics of Series in Econometric Analysis ........................ 10 Appendix G Descriptive Statistics of Additional Series in SEM ................................ 18 Appendix H LS Estimation ......................................................................................... 22 Appendix I FD Estimation........................................................................................... 24 Appendix J Assumptions for FE Estimation ............................................................... 25 Appendix K RE Estimation ......................................................................................... 30 Appendix L LS Estimation with Factory Dummies .................................................... 32 Appendix M FE Estimation with Week Dummies ...................................................... 33 Appendix N Estimation with Lagged Endogenous Variables ..................................... 35 Appendix O FE Estimation with Interaction and Squared Variables .......................... 38 Appendix P Six Stage Approach to SEM .................................................................... 40 Appendix Q List of Symbols in Path Diagram ............................................................ 41 Appendix R List of Symbols in PLS Algorithm ......................................................... 42 Appendix S Overviews, Path Coefficients and Inner T-Statistics ............................... 43 Appendix T Model 1: PLS Algorithm and Bootstrapping .......................................... 46 Appendix U Availability, Performance and Quality Results ...................................... 47 Appendix V 24 hours Overview of Activities on Spray Towers ................................. 48 Appendix W Mean Differences Between Spray Towers............................................. 59 Appendix X Assumptions for Repeated Measures ANOVA ...................................... 61 Appendix Y SPSS Output of Repeated Measures ANOVA........................................ 62 Appendix Z FIMIX-PLS Segmentation .................................................................... 104
II
Appendix AA Segmentation Pattern based on FIMIX-PLS ...................................... 110 Appendix BB Assumptions for One-Way ANOVA.................................................. 111 Appendix CC One-Way ANOVA of OEE* based on FIMIX-PLS .......................... 113 Appendix DD CD with Dataset and Preliminary Calculations ................................. 114
Appendix A Optimal Measurement of OEE* in Arla Foods and Arla Foods GIN
AKAFA CIP CIP tower is assumed, if there are more than four hours from one batch ends to the next batch begins. Miscellaneous CIP is assumed, if there are less than four hours from one batch ends to the next begins. It is assumed that suction of the tower is part of CIP. Miscellaneous Production is only what is logged as production, not production start, water on feeding system and stop production. Setup It is assumed that setup time in connection with breakdowns is attributable to the breakdown and therefore considered part of breakdown time. Vimmerby Setup It is assumed that there are two hours of setup (one before, and one after the production) in connection with changeover of products. CIP CIP of the feeding system, i.e., CIP misc., is performed when there is more than 1.5 hours between two batches. Breakdown Breakdown is assumed when a comment is made in the columns Comments/Breakdowns in the spreadsheet forwarded by the PTD. Breakdowns with no accurate time period are placed in the middle of the production time. Idling Idling with no accurate time period attached is placed in the middle of the production time. If the spray tower stops in the middle of a batch, but no setup has been recorded after this batch, the standstill is assumed to be idling.
Quality It is assumed that products for rework belong to the batch to which it is registered. It is assumed that output sold for alternative use, is not part of the quality loss.
Denmark Protein (DP) Breakdown Breakdowns are assumed if there are more than 5 minutes between two logged processes. Setup It is assumed that sweeping is part of setup. Idling Idling is assumed when the spray is in standstill for less than 1.5 hours. Standstill Standstill is assumed when the spray stops for more than 1.5 hours. CIP It is assumed that when CIP is logged, it is CIP tower. Idling after CIP tower is considered part of CIP. Misc. CIP is performed 2-3 times a week and is assumed to be sweeping for 4-5 hours. Quality If the non-conforming output exceeds the total output produced, a positive waste is assumed, and the quality of this batch is assumed to be 100%. A positive waste can occur, if no CIP or sweeping is performed between the batches and powder is accumulated. ARINCO Breakdown Breakdown on spray 4 is only recorded, if a time period is indicated.
Breakdown on spray 5 is assumed to last two hours from the time planned in the spreadsheet Comments". Not all comments are acknowledged as breakdowns. When comparing to spray 4, it is decided to include the same number of breakdowns and these are selected based on the wording (problems, shift etc), so it is possible with great certainty to determine that the spray has been non-productive during this period. Repairs are not included as breakdown, as repair can be performed during CIP. Setup Setup time is assumed to be the first and last hour of CIP (if the CIP is long enough for this assumption to be feasible), since CIP, according to the PTD, contains setup. Idling If there is a gap between CIP and production start, idling is assumed (unless the gap is due to a breakdown).
10
CIP Misc.
11
CIP Tower
12
Setup
Standstill
13
Idling
14
Breakdown
Capacity
15
Non-conforming Products
16
Product Shift
Number of CIP
17
18
Productivity/raw material
19
Capital Cost
20
Material Cost
Salary Cost
21
22
Appendix H LS Estimation
Regression
Box Plot
23
Correlogram
24
Appendix I FD Estimation
Regression
Correlogram
25
where yit is the productivity of a given spray i in time t explained by the exogenous variables xit, j are the parameters to estimate and ai is the unobserved effect of the individual sprays. Assumption FE. 2: Random Sampling The initial population consists of 6 production sites with a total of 18 sprays and from this population a sample of 10 sprays is drawn. drawn. Even though one of the production sites is not represented in the sample and the sprays have been chosen by the PTDs on the respective plants, the sample can be considered random as there is no pattern or underlying reason for the selection of sprays. sprays Assumption FE. 3: No Perfect Collinearity The presence of multicollinearity is tested by means of a correlation matrix of the explanatory variables.
The matrix reveals that there is, as expected, a high correlation of 0.821 between CIP tower (CT) and number of CIP (NC). Assumption FE. 4: Strict Exogeneity The assumption of strict exogeneity
26
can be assessed by evaluating the correlations between the residuals and the covariates. The correlation matrix reveals correlations of approximately zero between bet the covariates and the residuals and strict exogeneity can therefore be assumed.
In order to assess whether the assumption of homoscedasticity is fulfilled, it is relevant to evaluate the scatter plots of the residuals.
It is evident that there is no pattern in the distribution of the residuals, residuals and that there is a more or less even spread of the residuals across the observations. However, it appears that there are at least four outlying observations, which will be evaluated in connection with assumption 7. Assumption FE. 6: Uncorrelated Idiosyncratic Errors
27
The correlation of the errors can be evaluated by means of the correlogram below. It is noticeable that significant correlation exists from lag 7 and onwards despite the correction for fixed effects, which is supposed to remedy some of the omitted variables bias. The remaining correlation must be accomodated in the attempts to improve the FE estimation.
Assumption FE. 7: Normality The assumption of normality in the error terms can be evaluated by means of a histogram showing the distribution of standardized residuals.
The histogram shows a bellshaped distribution with a mean value of approximately zero. It is evident that outliers occur in both tails, and it is therefore relevant to consider the Jarque-Bera statistic, which indicates whether the standardized residuals are actually normally distributed. The null hypothesis of normality is rejected as the Jarque-Bera
28
statistic yields a p-value of zero, which violates the assumption that the error terms follow the normal distribution. It is attempted to accommodate this violation by a log transformation, as this reduces the effect of outliers, of the variables for which it seems logical to perform such a transformation. However, since some of the covariates contain observations of zero and therefore are not suitable for log transformation, and the remaining transformations results in a poorer degree of explanation, log transformation is not applied for accommodating the problem of normality. The box plot of the residuals in model 0 indicats no outliers, but after performing FE estimation three far outliers appear, which must be removed in order to achieve normally distributed error terms.
After excluding the three far outlying observations from the model, the Jarque-Bera test is still significant, as seen above, indicating that some of the near outliers must also be removed to ensure normality. A fourth observation is removed, and normality is achieved, as seen below. It must be kept in mind that the removal of the four outliers (Vimmerby 1 and 2 in weeks 4 and 12) creates an unbalanced panel of 116 observations, and the unbalanced panel may constrain the options for improving the model specification.
29
30
Appendix K RE Estimation
Regression
31
Hausman Test
32
Correlogram
33
34
Correlogram
35
36
37
38
39
40
41
42
Inner Relations A vector of the endogenous latent variables A vector of exogenous latent variables A vector of residuals Path coefficient matrix Path coefficient matrix
43
Path Coefficients
Inner T-Statistics
44
Model 1 Overview
Path Coefficients
Inner T-Statistics
45
Model 2 Overview
Path Coefficients
Inner T-Statistics
46
Bootstrapping
47
Performance
Quality
48
The activities occuring in the following 24 hour overview is registered according to the characteristics of the activity and have been assigned the corresponding colour as seen above. The time recorded at the beginning of a new colour is the starting time for the specific activity. If a time is noted in the middle of a production period, this is an indication that a batch shift has occured, but the equipment produces the same product.
49
HOCO Spray 2
50
HOCO Spray 3
51
AFAKA Spray 2
52
Vimmerby Spray 1
53
Vimmerby Spray 2
54
DP Spray 1
55
DP Spray 2
56
DP Spray 3
57
ARINCO Spray 4
58
ARINCO Spray 5
59
Setup
Standstill
60
Availability
Performance
Since no data is recorded in the performance and quality measures for Vimmerby 2 in the first three weeks of the fourth quarter of 2009, the values for Vimmerby 2 are replaced by results based on a separate Repeated Measures ANOVA only including the last nine weeks of the period. It is worth mentioning that both Repeated Measures ANOVAs have been significant overall, and it is therefore chosen to incorporate them in a joint figure.
61
62
Pairwise Comparisons Measure:OEE (I) Spray (J) Spray Mean Difference (I-J)
dimension1
dimension2
,110
1,000
63
3 4 5 6 7 8 9 10 2 1 3 4 5
dimension2
-,129 ,155 ,267 ,000 ,229 ,046 -,068 ,025 -,110 -,239
*
,065 ,083 ,073 ,067 ,078 ,066 ,076 ,090 ,073 ,039 ,046 ,096 ,061 ,053 ,046 ,048 ,052 ,065 ,039 ,042 ,086 ,031 ,033 ,014 ,029 ,041 ,083 ,046 ,042 ,085 ,049 ,057 ,040 ,039 ,044 ,073 ,096 ,086
1,000 1,000 ,172 1,000 ,626 1,000 1,000 1,000 1,000 ,003 1,000 1,000 1,000 1,000 1,000 ,148 1,000 1,000 ,003 ,001 ,034 ,076 ,000 ,000 1,000 ,149 1,000 1,000 ,001 1,000 ,421 1,000 ,920 ,007 ,617 ,172 1,000 ,034
-,412 -,209 -,053 -,292 -,114 -,242 -,401 -,368 -,430 -,410 -,154 -,261 -,377 -,115 -,266 -,387 -,313 -,153 ,068 ,100 ,020 -,008 ,213 ,114 -,064 -,027 -,519 -,244 -,468 -,261 -,371 -,176 -,286 -,396 -,324 -,588 -,576 -,773
,153 ,519 ,588 ,292 ,572 ,334 ,264 ,418 ,210 -,068 ,244 ,576 ,157 ,353 ,138 ,030 ,143 ,412 ,410 ,468 ,773 ,266 ,504 ,236 ,186 ,335 ,209 ,154 -,100 ,486 ,061 ,324 ,067 -,051 ,064 ,053 ,261 -,020
,045 ,157 -,110 ,119 -,064 -,178 -,085 ,129 ,239 ,284 ,397
* * *
6 7 8 9 10
1 2 4 5
dimension2
6 7 8 9 10
1 2 3 5
dimension2
6 7 8 9 10
5
dimension2
1 2 3
64
4 6 7 8 9 10 6 1 2 3 4
dimension2
-,112 -,267 -,038 -,222 -,336 -,243 ,000 ,110 -,129 ,155 ,267 ,229
*
,085 ,079 ,104 ,083 ,083 ,094 ,067 ,061 ,031 ,049 ,079 ,047 ,026 ,038 ,051 ,078 ,053 ,033 ,057 ,104 ,047 ,033 ,042 ,035 ,066 ,046 ,014 ,040 ,083 ,026 ,033 ,021 ,038 ,076 ,048 ,029 ,039 ,083
1,000 ,267 1,000 ,984 ,088 1,000 1,000 1,000 ,076 ,421 ,267 ,023 1,000 1,000 1,000 ,626 1,000 ,000 1,000 1,000 ,023 ,007 ,001 ,005 1,000 1,000 ,000 ,920 ,984 1,000 ,007 ,010 1,000 1,000 ,148 1,000 ,007 ,088
-,486 -,612 -,492 -,585 -,700 -,656 -,292 -,157 -,266 -,061 -,077 ,022 -,067 -,235 -,200 -,572 -,353 -,504 -,324 -,415 -,436 -,326 -,482 -,358 -,334 -,138 -,236 -,067 -,142 -,159 ,040 -,207 -,187 -,264 -,030 -,186 ,051 -,028
,261 ,077 ,415 ,142 ,028 ,171 ,292 ,377 ,008 ,371 ,612 ,436 ,159 ,099 ,250 ,114 ,115 -,213 ,176 ,492 -,022 -,040 -,113 -,050 ,242 ,266 -,114 ,286 ,585 ,067 ,326 -,022 ,145 ,401 ,387 ,064 ,396 ,700
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9 10
1 2 3 4
dimension2
5 6 7 9 10
1 2
dimension2
3 4 5
,336
65
6 7 8 10 10 1 2 3 4
dimension2
,038 ,042 ,021 ,040 ,090 ,052 ,041 ,044 ,094 ,051 ,035 ,038 ,040
1,000 ,001 ,010 1,000 1,000 1,000 ,149 ,617 1,000 1,000 ,005 1,000 1,000
-,099 ,113 ,022 -,083 -,418 -,143 -,335 -,064 -,171 -,250 ,050 -,145 -,270
,235 ,482 ,207 ,270 ,368 ,313 ,027 ,324 ,656 ,200 ,358 ,187 ,083
5 6 7 8 9
,021 -,093
Based on estimated marginal means a. Adjustment for multiple comparisons: Bonferroni. *. The mean difference is significant at the ,05 level.
Availability
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 ,6467 ,7033 ,8183 ,4150 ,3475 ,6892 ,6050 ,7450 ,7858 ,7933 Std. Deviation ,25256 ,17489 ,03433 ,24262 ,29515 ,11057 ,14897 ,05584 ,07561 ,06035 N 12 12 12 12 12 12 12 12 12 12
66
Pairwise Comparisons Measure:Availability (I) Spray (J) Spray Mean Difference (I-J) 1 2 3 4 5
dimension2
95% Confidence Interval for Difference Std. Error ,051 ,079 ,074 ,055 ,062 ,098 ,074 ,066 ,071 ,051 ,055 ,061 ,066 ,027 ,072 ,052 ,052 ,046 ,079 ,055 ,070 ,089 ,034 ,041 Sig.
a a
Lower Bound -,281 -,517 -,094 ,057 -,314 -,387 -,424 -,429 -,456 -,167 -,355 ,021 ,067 -,104 -,218 -,267 -,310 -,293 -,174 -,125 ,096 ,080 -,020 ,035
Upper Bound ,167 ,174 ,557 ,541 ,229 ,470 ,227 ,151 ,163 ,281 ,125 ,556 ,645 ,132 ,414 ,184 ,145 ,113 ,517 ,355 ,710 ,861 ,278 ,392
1,000 1,000 ,443 ,010 1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,028 ,010 1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,006 ,012 ,136 ,013
6 7 8 9 10
1 3 4
dimension1
5
dimension2
6 7 8 9 10
1 2 4
dimension2
5 6 7
,129 ,213
*
67
8 9 10 4 1 2 3 5
dimension2
,023 ,026 ,021 ,074 ,061 ,070 ,099 ,063 ,075 ,071 ,073 ,065 ,055 ,066 ,089 ,099 ,070 ,118 ,090 ,084 ,080 ,062 ,027 ,034 ,063 ,070 ,059 ,038 ,038 ,032 ,098 ,072 ,041 ,075 ,118 ,059 ,042 ,050
,418 1,000 1,000 ,443 ,028 ,006 1,000 ,054 1,000 ,030 ,016 ,006 ,010 ,010 ,012 1,000 ,022 1,000 ,048 ,013 ,008 1,000 1,000 ,136 ,054 ,022 1,000 1,000 1,000 ,365 1,000 1,000 ,013 1,000 1,000 1,000 ,299 ,183
-,029 -,080 -,067 -,557 -,556 -,710 -,366 -,551 -,516 -,638 -,690 -,665 -,541 -,645 -,861 -,501 -,649 -,773 -,793 -,807 -,796 -,229 -,132 -,278 -,003 ,034 -,172 -,220 -,261 -,246 -,470 -,414 -,392 -,136 -,258 -,340 -,324 -,400
,175 ,145 ,117 ,094 -,021 -,096 ,501 ,003 ,136 -,022 -,052 -,092 -,057 -,067 -,080 ,366 -,034 ,258 -,002 -,070 -,095 ,314 ,104 ,020 ,551 ,649 ,340 ,109 ,067 ,037 ,387 ,218 -,035 ,516 ,773 ,172 ,044 ,038
6 7 8 9 10
1 2 3 4
dimension2
-,068 -,342
*
6 7 8 9 10
1 2 3 4
dimension2
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9
68
10 8 1 2 3 4
dimension2
,052 ,074 ,052 ,023 ,071 ,090 ,038 ,042 ,027 ,025 ,066 ,052 ,026 ,073 ,084 ,038 ,050 ,027 ,034 ,071 ,046 ,021 ,065 ,080 ,032 ,052 ,025 ,034
,184 1,000 1,000 ,418 ,030 ,048 1,000 ,299 1,000 1,000 1,000 1,000 1,000 ,016 ,013 1,000 ,183 1,000 1,000 1,000 1,000 1,000 ,006 ,008 ,365 ,184 1,000 1,000
-,416 -,227 -,184 -,175 ,022 ,002 -,109 -,044 -,159 -,158 -,151 -,145 -,145 ,052 ,070 -,067 -,038 -,078 -,157 -,163 -,113 -,117 ,092 ,095 -,037 -,040 -,061 -,142
,040 ,424 ,267 ,029 ,638 ,793 ,220 ,324 ,078 ,061 ,429 ,310 ,080 ,690 ,807 ,261 ,400 ,159 ,142 ,456 ,293 ,067 ,665 ,796 ,246 ,416 ,158 ,157
5 6 7 9 10
1 2 3 4
dimension2
5 6 7 8 10
10
1 2 3 4
dimension2
5 6 7 8 9
Based on estimated marginal means a. Adjustment for multiple comparisons: Bonferroni. *. The mean difference is significant at the ,05 level.
69
Performance
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 ,9358 ,7250 ,9308 ,8283 ,8792 ,8550 ,6542 ,8025 ,8817 ,7833 Std. Deviation ,10681 ,17016 ,02968 ,09379 ,06201 ,06922 ,13235 ,05429 ,08266 ,16870 N 12 12 12 12 12 12 12 12 12 12
Pairwise Comparisons Measure:Performance (I) Spray (J) Spray Mean Difference (I-J) 1
dimension1 dimension2
Sig.
2 3
,211
,042 ,033
,019 1,000
,005
70
4 5 6 7 8 9 10 2 1 3 4 5
dimension2
,044 ,036 ,042 ,048 ,033 ,041 ,060 ,042 ,050 ,070 ,058 ,061 ,049 ,052 ,052 ,054 ,033 ,050 ,029 ,021 ,023 ,040 ,017 ,027 ,052 ,044 ,070 ,029 ,033 ,026 ,045 ,028 ,040 ,061 ,036 ,058 ,021 ,033
1,000 1,000 1,000 ,005 ,086 1,000 1,000 ,019 ,081 1,000 ,969 1,000 1,000 1,000 ,524 1,000 1,000 ,081 ,200 1,000 ,337 ,001 ,001 1,000 ,759 1,000 1,000 ,200 1,000 1,000 ,122 1,000 1,000 1,000 1,000 ,969 1,000 1,000
-,084 -,100 -,102 ,071 -,011 -,126 -,110 -,396 -,426 -,410 -,406 -,396 -,143 -,303 -,383 -,292 -,151 -,015 -,023 -,040 -,026 ,102 ,053 -,069 -,082 -,299 -,204 -,228 -,196 -,141 -,024 -,096 -,227 -,221 -,213 -,098 -,143 -,095
,299 ,213 ,264 ,493 ,277 ,235 ,415 -,025 ,015 ,204 ,098 ,136 ,285 ,148 ,070 ,176 ,141 ,426 ,228 ,143 ,177 ,451 ,203 ,167 ,377 ,084 ,410 ,023 ,095 ,088 ,372 ,148 ,120 ,311 ,100 ,406 ,040 ,196
-,206 -,103 -,154 -,130 ,071 -,077 -,157 -,058 -,005 ,206 ,102 ,052 ,076 ,277 ,128
* *
6 7 8 9 10
1 2 4 5
dimension2
6 7 8 9 10
,049 ,147 -,107 ,103 -,102 -,051 -,027 ,174 ,026 -,053 ,045 -,057 ,154 -,052 ,051
1 2 3 5
dimension2
6 7 8 9 10
1 2
dimension2
3 4
71
6 7 8 9 10 6 1 2 3 4
dimension2
,024 ,225
*
,026 ,050 ,020 ,030 ,055 ,042 ,061 ,023 ,026 ,026 ,044 ,030 ,030 ,053 ,048 ,049 ,040 ,045 ,050 ,044 ,044 ,048 ,059 ,033 ,052 ,017 ,028 ,020 ,030 ,044 ,029 ,051 ,041 ,052 ,027 ,040 ,030 ,030
1,000 ,041 ,109 1,000 1,000 1,000 1,000 ,337 1,000 1,000 ,034 1,000 1,000 1,000 ,005 1,000 ,001 ,122 ,041 ,034 ,271 ,026 1,000 ,086 1,000 ,001 1,000 ,109 1,000 ,271 ,856 1,000 1,000 ,524 1,000 1,000 1,000 1,000
-,091 ,006 -,009 -,132 -,146 -,264 -,136 -,177 -,088 -,139 ,010 -,079 -,158 -,159 -,493 -,285 -,451 -,372 -,444 -,392 -,340 -,436 -,386 -,277 -,148 -,203 -,148 -,162 -,184 -,043 -,205 -,203 -,235 -,070 -,167 -,120 -,127 -,104
,139 ,444 ,162 ,127 ,338 ,102 ,396 ,026 ,141 ,091 ,392 ,184 ,104 ,302 -,071 ,143 -,102 ,024 -,006 -,010 ,043 -,019 ,127 ,011 ,303 -,053 ,096 ,009 ,079 ,340 ,047 ,241 ,126 ,383 ,069 ,227 ,132 ,158
5 7 8 9 10
1 2 3 4
dimension2
-,071 -,277
*
5 6 8 9 10
-,148 -,228
*
1 2 3 4
dimension2
-,026 -,077 -,053 ,148 -,079 ,019 -,054 ,157 -,049 ,053 ,002 ,027
5 6 7 9 10
1 2 3
dimension2
4 5 6
72
7 8 10 10 1 2 3 4
dimension2
,228
,048 ,029 ,037 ,060 ,054 ,052 ,061 ,055 ,053 ,059 ,051 ,037
,026 ,856 ,998 1,000 1,000 ,759 1,000 1,000 1,000 1,000 1,000 ,998
,019 -,047 -,063 -,415 -,176 -,377 -,311 -,338 -,302 -,127 -,241 -,260
,436 ,205 ,260 ,110 ,292 ,082 ,221 ,146 ,159 ,386 ,203 ,063
,079 ,098 -,152 ,058 -,147 -,045 -,096 -,072 ,129 -,019 -,098
5 6 7 8 9
Based on estimated marginal means *. The mean difference is significant at the a. Adjustment for multiple comparisons: Bonferroni.
Performance Vimmerby 2
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 ,9156 ,6889 ,9322 ,8156 ,8956 ,8511 ,6189 ,8011 ,8711 ,7522 Std. Deviation ,11759 ,18135 ,02774 ,10655 ,06386 ,07184 ,11921 ,05840 ,09239 ,18061 N 9 9 9 9 9 9 9 9 9 9
73
Pairwise Comparisons Measure:Performance_VI2 (I) Spray (J) Spray Mean Difference (I-J) 1 2 3 4 5
dimension2
95% Confidence Interval for Difference Std. Error ,055 ,042 ,059 ,041 ,054 ,061 ,042 ,055 ,080 ,055 ,063 ,093 ,068 ,076 ,062 ,065 ,066 ,070 ,042 ,063 ,037 ,025 ,025 Sig.
a a
Lower Bound -,047 -,224 -,192 -,183 -,205 -,006 -,091 -,229 -,231 -,501 -,554 -,588 -,542 -,540 -,238 -,432 -,508 -,412 -,191 -,067 -,066 -,089 -,042
Upper Bound ,501 ,191 ,392 ,223 ,334 ,599 ,320 ,317 ,558 ,047 ,067 ,334 ,129 ,215 ,378 ,208 ,144 ,285 ,224 ,554 ,300 ,163 ,205
,227 -,017 ,100 ,020 ,064 ,297 ,114 ,044 ,163 -,227 -,243 -,127 -,207 -,162 ,070 -,112 -,182 -,063 ,017 ,243 ,117 ,037 ,081
,155 1,000 1,000 1,000 1,000 ,056 1,000 1,000 1,000 ,155 ,210 1,000 ,710 1,000 1,000 1,000 1,000 1,000 1,000 ,210 ,604 1,000 ,522
6 7 8 9 10
1 3
dimension1
4 5
dimension2
6 7 8 9 10
1 2
dimension2
4 5 6
74
7 8 9 10 4 1 2 3 5
dimension2
,313 ,131
* *
,041 ,022 ,033 ,065 ,059 ,093 ,037 ,040 ,033 ,055 ,036 ,053 ,080 ,041 ,068 ,025 ,040 ,030 ,053 ,022 ,035 ,065 ,054 ,076 ,025 ,033 ,030 ,052 ,035 ,039 ,068 ,061 ,062 ,041 ,055 ,053 ,052 ,046
,003 ,016 1,000 1,000 1,000 1,000 ,604 1,000 1,000 ,339 1,000 1,000 1,000 1,000 ,710 1,000 1,000 1,000 ,035 ,113 1,000 1,000 1,000 1,000 ,522 1,000 1,000 ,091 1,000 1,000 1,000 ,056 1,000 ,003 ,339 ,035 ,091 ,177
,111 ,021 -,104 -,141 -,392 -,334 -,300 -,278 -,199 -,078 -,162 -,319 -,334 -,223 -,129 -,163 -,118 -,105 ,015 -,014 -,147 -,179 -,334 -,215 -,205 -,128 -,194 -,024 -,126 -,212 -,236 -,599 -,378 -,516 -,471 -,538 -,489 -,408
,516 ,241 ,226 ,501 ,192 ,588 ,066 ,118 ,128 ,471 ,191 ,208 ,461 ,183 ,542 ,089 ,278 ,194 ,538 ,203 ,196 ,466 ,205 ,540 ,042 ,199 ,105 ,489 ,226 ,172 ,434 ,006 ,238 -,111 ,078 -,015 ,024 ,043
,061 ,180 -,100 ,127 -,117 -,080 -,036 ,197 ,014 -,056 ,063 -,020 ,207 -,037 ,080 ,044 ,277
*
6 7 8 9 10
1 2 3 4
dimension2
6 7 8 9 10
,094 ,024 ,143 -,064 ,162 -,081 ,036 -,044 ,232 ,050 -,020 ,099 -,297 -,070 -,313
*
1 2 3 4
dimension2
5 7 8 9 10
1 2 3
dimension2
4 5 6 8
-,197 -,277
*
-,232 -,182
75
9 10 8 1 2 3 4
dimension2
,057 ,079 ,042 ,065 ,022 ,036 ,022 ,035 ,046 ,037 ,061 ,055 ,066 ,033 ,053 ,035 ,039 ,057 ,037 ,044 ,080 ,070 ,065 ,080 ,065 ,068 ,079 ,061 ,044
,095 1,000 1,000 1,000 ,016 1,000 ,113 1,000 ,177 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,095 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-,533 -,525 -,320 -,208 -,241 -,191 -,203 -,226 -,043 -,253 -,255 -,317 -,144 -,226 -,208 -,196 -,172 -,028 -,113 -,098 -,558 -,285 -,501 -,461 -,466 -,434 -,259 -,352 -,336
,028 ,259 ,091 ,432 -,021 ,162 ,014 ,126 ,408 ,113 ,352 ,229 ,508 ,104 ,319 ,147 ,212 ,533 ,253 ,336 ,231 ,412 ,141 ,334 ,179 ,236 ,525 ,255 ,098
-,014 -,094 -,050 ,182 -,070 ,049 -,044 ,182 -,061 ,056 -,024 ,020 ,252 ,070 ,119 -,163 ,063 -,180 -,063 -,143 -,099 ,133 -,049 -,119
5 6 7 9 10
1 2 3 4
dimension2
5 6 7 8 10
10
1 2 3 4
dimension2
5 6 7 8 9
Based on estimated marginal means a. Adjustment for multiple comparisons: Bonferroni. *. The mean difference is significant at the
76
Quality
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 ,8800 ,9867 ,9867 ,9858 ,8583 ,9833 ,9417 ,9675 ,9875 ,9800 Std. Deviation ,29180 ,02570 ,00492 ,02275 ,27149 ,04097 ,11543 ,05754 ,00622 ,00953 N 12 12 12 12 12 12 12 12 12 12
Pairwise Comparisons Measure:Quality (I) Spray (J) Spray Mean Difference (I-J) 1
dimension1 dimension2
95% Confidence Interval for Difference Std. Error ,086 ,084 Sig.
a a
2 3
-,107 -,107
1,000 1,000
77
4 5 6 7 8 9 10 2 1 3 4 5
dimension2
-,106 ,022 -,103 -,062 -,088 -,108 -,100 ,107 ,000 ,001 ,128 ,003 ,045 ,019 -,001 ,007 ,107 ,000 ,001 ,128 ,003 ,045 ,019 -,001 ,007 ,106 -,001 -,001 ,127 ,003 ,044 ,018 -,002 ,006 -,022 -,128 -,128 -,127
,085 ,125 ,087 ,096 ,082 ,084 ,084 ,086 ,008 ,011 ,080 ,014 ,036 ,020 ,008 ,008 ,084 ,008 ,006 ,079 ,012 ,033 ,017 ,002 ,003 ,085 ,011 ,006 ,080 ,011 ,033 ,020 ,007 ,006 ,125 ,080 ,079 ,080
1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-,478 -,524 -,485 -,481 -,445 -,475 -,467 -,268 -,034 -,047 -,221 -,058 -,112 -,067 -,034 -,027 -,262 -,034 -,027 -,216 -,051 -,101 -,054 -,009 -,008 -,266 -,049 -,029 -,224 -,048 -,101 -,067 -,033 -,022 -,567 -,478 -,473 -,479
,266 ,567 ,278 ,357 ,270 ,260 ,267 ,482 ,034 ,049 ,478 ,065 ,202 ,106 ,033 ,040 ,476 ,034 ,029 ,473 ,057 ,191 ,092 ,008 ,021 ,478 ,047 ,027 ,479 ,053 ,189 ,104 ,029 ,034 ,524 ,221 ,216 ,224
6 7 8 9 10
1 2 4 5
dimension2
6 7 8 9 10
1 2 3 5
dimension2
6 7 8 9 10
1 2
dimension2
3 4
78
6 7 8 9 10 6 1 2 3 4
dimension2
-,125 -,083 -,109 -,129 -,122 ,103 -,003 -,003 -,003 ,125 ,042 ,016 -,004 ,003 ,062 -,045 -,045 -,044 ,083 -,042 -,026 -,046 -,038 ,088 -,019 -,019 -,018 ,109 -,016 ,026 -,020 -,013 ,108 ,001 ,001 ,002 ,129 ,004
,082 ,060 ,077 ,079 ,080 ,087 ,014 ,012 ,011 ,082 ,037 ,023 ,012 ,012 ,096 ,036 ,033 ,033 ,060 ,037 ,040 ,034 ,034 ,082 ,020 ,017 ,020 ,077 ,023 ,040 ,017 ,018 ,084 ,008 ,002 ,007 ,079 ,012
1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-,482 -,346 -,447 -,474 -,470 -,278 -,065 -,057 -,053 -,232 -,119 -,083 -,058 -,048 -,357 -,202 -,191 -,189 -,180 -,203 -,202 -,194 -,188 -,270 -,106 -,092 -,104 -,229 -,115 -,150 -,095 -,092 -,260 -,033 -,008 -,029 -,215 -,050
,232 ,180 ,229 ,215 ,227 ,485 ,058 ,051 ,048 ,482 ,203 ,115 ,050 ,054 ,481 ,112 ,101 ,101 ,346 ,119 ,150 ,102 ,112 ,445 ,067 ,054 ,067 ,447 ,083 ,202 ,055 ,067 ,475 ,034 ,009 ,033 ,474 ,058
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9 10
1 2 3 4
dimension2
5 6 7 9 10
1 2 3
dimension2
4 5 6
79
7 8 10 10 1 2 3 4
dimension2
,046 ,020 ,007 ,100 -,007 -,007 -,006 ,122 -,003 ,038 ,013 -,007
,034 ,017 ,002 ,084 ,008 ,003 ,006 ,080 ,012 ,034 ,018 ,002
1,000 1,000 ,246 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,246
-,102 -,055 -,002 -,267 -,040 -,021 -,034 -,227 -,054 -,112 -,067 -,017
,194 ,095 ,017 ,467 ,027 ,008 ,022 ,470 ,048 ,188 ,092 ,002
5 6 7 8 9
Quality Vimmerby 2
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 ,9511 ,9822 ,9856 ,9900 ,8244 ,9856 ,9256 ,9667 ,9867 ,9811 Std. Deviation ,10481 ,02863 ,00527 ,01658 ,31013 ,04333 ,13087 ,06225 ,00707 ,01054 N 9 9 9 9 9 9 9 9 9 9
80
Pairwise Comparisons Measure:Quality_VI2 (I) Spray (J) Spray Mean Difference (I-J) 1 2 3 4 5
dimension2
95% Confidence Interval for Difference Std. Error ,037 ,034 ,036 ,117 ,040 ,063 ,045 ,033 ,033 ,037 ,010 ,012 ,106 ,017 ,048 ,025 ,010 ,009 ,034 ,010 ,005 ,104 ,015 Sig.
a a
Lower Bound -,214 -,205 -,216 -,455 -,233 -,289 -,238 -,197 -,194 -,152 -,054 -,068 -,369 -,088 -,180 -,109 -,054 -,046 -,136 -,048 -,028 -,355 -,075
Upper Bound ,152 ,136 ,138 ,708 ,164 ,340 ,207 ,126 ,134 ,214 ,048 ,052 ,685 ,081 ,294 ,140 ,045 ,048 ,205 ,054 ,019 ,677 ,075
-,031 -,034 -,039 ,127 -,034 ,026 -,016 -,036 -,030 ,031 -,003 -,008 ,158 -,003 ,057 ,016 -,004 ,001 ,034 ,003 -,004 ,161 ,000
1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
6 7 8 9 10
1 3
dimension1
4 5
dimension2
6 7 8 9 10
1 2
dimension2
4 5 6
81
7 8 9 10 4 1 2 3 5
dimension2
,060 ,019 -,001 ,004 ,039 ,008 ,004 ,166 ,004 ,064 ,023 ,003 ,009 -,127 -,158 -,161 -,166 -,161 -,101 -,142 -,162 -,157 ,034 ,003 ,000 -,004 ,161 ,060 ,019 -,001 ,004 -,026 -,057 -,060 -,064 ,101 -,060 -,041
,044 ,021 ,003 ,004 ,036 ,012 ,005 ,105 ,015 ,042 ,023 ,006 ,007 ,117 ,106 ,104 ,105 ,107 ,080 ,102 ,104 ,105 ,040 ,017 ,015 ,015 ,107 ,048 ,028 ,015 ,015 ,063 ,048 ,044 ,042 ,080 ,048 ,053
1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-,157 -,085 -,014 -,016 -,138 -,052 -,019 -,356 -,072 -,144 -,091 -,028 -,024 -,708 -,685 -,677 -,687 -,693 -,500 -,645 -,678 -,678 -,164 -,081 -,075 -,081 -,371 -,176 -,118 -,076 -,069 -,340 -,294 -,277 -,273 -,297 -,296 -,303
,277 ,123 ,012 ,025 ,216 ,068 ,028 ,687 ,081 ,273 ,138 ,034 ,041 ,455 ,369 ,355 ,356 ,371 ,297 ,361 ,354 ,364 ,233 ,088 ,075 ,072 ,693 ,296 ,156 ,073 ,077 ,289 ,180 ,157 ,144 ,500 ,176 ,221
6 7 8 9 10
1 2 3 4
dimension2
6 7 8 9 10
1 2 3 4
dimension2
5 7 8 9 10
1 2 3
dimension2
4 5 6 8
82
9 10 8 1 2 3 4
dimension2
-,061 -,056 ,016 -,016 -,019 -,023 ,142 -,019 ,041 -,020 -,014 ,036 ,004 ,001 -,003 ,162 ,001 ,061 ,020 ,006 ,030 -,001 -,004 -,009 ,157 -,004 ,056 ,014 -,006
,044 ,045 ,045 ,025 ,021 ,023 ,102 ,028 ,053 ,022 ,023 ,033 ,010 ,003 ,006 ,104 ,015 ,044 ,022 ,002 ,033 ,009 ,004 ,007 ,105 ,015 ,045 ,023 ,002
1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-,281 -,277 -,207 -,140 -,123 -,138 -,361 -,156 -,221 -,127 -,127 -,126 -,045 -,012 -,034 -,354 -,073 -,159 -,087 -,006 -,134 -,048 -,025 -,041 -,364 -,077 -,166 -,098 -,018
,159 ,166 ,238 ,109 ,085 ,091 ,645 ,118 ,303 ,087 ,098 ,197 ,054 ,014 ,028 ,678 ,076 ,281 ,127 ,018 ,194 ,046 ,016 ,024 ,678 ,069 ,277 ,127 ,006
5 6 7 9 10
1 2 3 4
dimension2
5 6 7 8 10
10
1 2 3 4
dimension2
5 6 7 8 9
83
CIP Tower
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 73,1667 14,7500 29,1667 42,9167 53,5833 62,6667 80,0833 62,8333 102,2500 54,6667 Std. Deviation 76,25297 51,09550 43,36071 102,60911 97,00371 61,66085 81,58705 86,92613 121,07483 69,72196 N 12 12 12 12 12 12 12 12 12 12
Pairwise Comparisons Measure:CIP_tower (I) Spray (J) Spray Mean Difference (I-J) 1
dimension1 dimension2
95% Confidence Interval for Difference Std. Error 13,401 29,938 Sig.
a a
2 3
58,417 44,000
,051 1,000
84
4 5 6 7 8 9 10 2 1 3 4 5
dimension2
30,250 19,583 10,500 -6,917 10,333 -29,083 18,500 -58,417 -14,417 -28,167 -38,833 -47,917 -65,333 -48,083 -87,500 -39,917 -44,000 14,417 -13,750 -24,417 -33,500 -50,917 -33,667 -73,083 -25,500 -30,250 28,167 13,750 -10,667 -19,750 -37,167 -19,917 -59,333 -11,750 -19,583 38,833 24,417 10,667
40,006 40,134 29,649 35,304 31,226 42,337 27,192 13,401 21,271 34,786 33,844 26,504 31,417 25,984 41,393 27,736 29,938 21,271 31,267 35,000 19,065 29,425 33,460 42,008 27,170 40,006 34,786 31,267 33,394 30,230 45,354 44,692 35,581 41,341 40,134 33,844 35,000 33,394
1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,051 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-144,713 -155,942 -119,170 -161,317 -126,234 -214,241 -100,426 -117,027 -107,445 -180,301 -186,848 -163,831 -202,735 -161,723 -268,531 -161,218 -174,931 -78,612 -150,495 -177,487 -116,882 -179,606 -180,003 -256,806 -144,328 -205,213 -123,967 -122,995 -156,716 -151,959 -235,521 -215,375 -214,946 -192,552 -195,109 -109,182 -128,653 -135,383
205,213 195,109 140,170 147,484 146,901 156,074 137,426 ,194 78,612 123,967 109,182 67,998 72,068 65,556 93,531 81,384 86,931 107,445 122,995 128,653 49,882 77,772 112,669 110,639 93,328 144,713 180,301 150,495 135,383 112,459 161,187 175,541 96,279 169,052 155,942 186,848 177,487 156,716
6 7 8 9 10
1 2 4 5
dimension2
6 7 8 9 10
1 2 3 5
dimension2
6 7 8 9 10
1 2
dimension2
3 4
85
6 7 8 9 10 6 1 2 3 4
dimension2
-9,083 -26,500 -9,250 -48,667 -1,083 -10,500 47,917 33,500 19,750 9,083 -17,417 -,167 -39,583 8,000 6,917 65,333 50,917 37,167 26,500 17,417 17,250 -22,167 25,417 -10,333 48,083 33,667 19,917 9,250 ,167 -17,250 -39,417 8,167 29,083 87,500 73,083 59,333 48,667 39,583
37,699 28,520 32,128 28,685 34,778 29,649 26,504 19,065 30,230 37,699 35,493 38,332 33,088 23,646 35,304 31,417 29,425 45,354 28,520 35,493 23,557 39,110 27,806 31,226 25,984 33,460 44,692 32,128 38,332 23,557 45,592 35,442 42,337 41,393 42,008 35,581 28,685 33,088
1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-173,960 -151,231 -149,759 -174,120 -153,185 -140,170 -67,998 -49,882 -112,459 -155,793 -172,643 -167,811 -184,292 -95,414 -147,484 -72,068 -77,772 -161,187 -98,231 -137,810 -85,775 -193,214 -96,193 -146,901 -65,556 -112,669 -175,541 -131,259 -167,477 -120,275 -238,811 -146,840 -156,074 -93,531 -110,639 -96,279 -76,787 -105,126
155,793 98,231 131,259 76,787 151,018 119,170 163,831 116,882 151,959 173,960 137,810 167,477 105,126 111,414 161,317 202,735 179,606 235,521 151,231 172,643 120,275 148,880 147,027 126,234 161,723 180,003 215,375 149,759 167,811 85,775 159,978 163,173 214,241 268,531 256,806 214,946 174,120 184,292
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9 10
1 2 3 4
dimension2
5 6 7 9 10
1 2 3
dimension2
4 5 6
86
7 8 10 10 1 2 3 4
dimension2
22,167 39,417 47,583 -18,500 39,917 25,500 11,750 1,083 -8,000 -25,417 -8,167 -47,583
39,110 45,592 34,705 27,192 27,736 27,170 41,341 34,778 23,646 27,806 35,442 34,705
1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-148,880 -159,978 -104,197 -137,426 -81,384 -93,328 -169,052 -151,018 -111,414 -147,027 -163,173 -199,363
193,214 238,811 199,363 100,426 161,218 144,328 192,552 153,185 95,414 96,193 146,840 104,197
5 6 7 8 9
CIP Misc.
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 26,8333 222,8333 121,0000 22,4167 12,9167 46,7500 46,8333 48,8333 112,2500 80,6667 Std. Deviation 11,92273 149,35853 23,95830 15,35317 11,39743 30,99303 41,88910 45,50291 22,75612 52,57606 N 12 12 12 12 12 12 12 12 12 12
87
Pairwise Comparisons Measure:CIP_misc (I) Spray (J) Spray Mean Difference (I-J) 1 2 3 4 5
dimension2
Sig.
Lower Bound -385,141 -131,405 -18,508 -10,329 -53,992 -81,295 -79,597 -121,475 -114,843 6,859 -79,026 14,810 12,870 -15,052 -15,144 -9,048 -83,801 -53,172 56,929 -282,693 68,521 72,600 20,097
Upper Bound -6,859 -56,929 27,342 38,163 14,158 41,295 35,597 -49,358 7,177 385,141 282,693 386,024 406,963 367,218 367,144 357,048 304,967 337,506 131,405 79,026 128,646 143,566 128,403
-196,000 -94,167
43,247 8,515 5,242 5,544 7,791 14,015 13,170 8,245 13,950 43,247 41,354 42,439 45,055 43,703 43,705 41,854 44,446 44,665 8,515 41,354 6,874 8,113 12,382
,038 ,000 1,000 1,000 1,000 1,000 1,000 ,000 ,120 ,038 1,000 ,028 ,031 ,089 ,090 ,072 1,000 ,392 ,000 1,000 ,000 ,000 ,004
6 7 8 9 10
-53,833 196,000
*
1 3
dimension1
4 5
dimension2
6 7 8 9 10
1 2
dimension2
4 5 6
88
7 8 9 10 4 1 2 3 5
dimension2
74,167
11,012 17,302 10,235 15,650 5,242 42,439 6,874 5,507 11,309 14,395 15,888 8,377 14,114 5,544 45,055 8,113 5,507 10,100 12,241 15,296 7,615 15,897 7,791 43,703 12,382 11,309 10,100 16,804 13,621 10,442 15,633 14,015 43,705 11,012 14,395 12,241 16,804 19,228
,001 ,070 1,000 1,000 1,000 ,028 ,000 1,000 1,000 1,000 1,000 ,000 ,076 1,000 ,031 ,000 1,000 ,292 ,819 1,000 ,000 ,060 1,000 ,089 ,004 1,000 ,292 1,000 1,000 ,003 1,000 1,000 ,090 ,001 1,000 ,819 1,000 1,000
26,008 -3,505 -36,013 -28,110 -27,342 -386,024 -128,646 -14,584 -73,791 -87,373 -95,903 -126,471 -119,976 -38,163 -406,963 -143,566 -33,584 -78,006 -87,452 -102,815 -132,638 -137,277 -14,158 -367,218 -128,403 -25,124 -10,339 -73,577 -61,653 -111,168 -102,289 -41,295 -367,144 -122,326 -38,540 -19,619 -73,410 -86,094
122,326 147,839 53,513 108,777 18,508 -14,810 -68,521 33,584 25,124 38,540 43,069 -53,195 3,476 10,329 -12,870 -72,600 14,584 10,339 19,619 30,981 -66,029 1,777 53,992 15,052 -20,097 73,791 78,006 73,410 57,486 -19,832 34,455 81,295 15,144 -26,008 87,373 87,452 73,577 82,094
6 7 8 9 10
1 2 3 4
dimension2
6 7 8 9 10
1 2 3 4
dimension2
5 7 8 9 10
1 2 3
dimension2
4 5 6 8
89
9 10 8 1 2 3 4
dimension2
-65,417 -33,833 22,000 -174,000 -72,167 26,417 35,917 2,083 2,000 -63,417
*
15,064 19,778 13,170 41,854 17,302 15,888 15,296 13,621 19,228 13,358 23,553 8,245 44,446 10,235 8,377 7,615 10,442 15,064 13,358 16,067 13,950 44,665 15,650 14,114 15,897 15,633 19,778 23,553 16,067
,053 1,000 1,000 ,072 ,070 1,000 1,000 1,000 1,000 ,027 1,000 ,000 1,000 1,000 ,000 ,000 ,003 ,053 ,027 1,000 ,120 ,392 1,000 ,076 ,060 1,000 1,000 1,000 1,000
-131,297 -120,331 -35,597 -357,048 -147,839 -43,069 -30,981 -57,486 -82,094 -121,839 -134,840 49,358 -304,967 -53,513 53,195 66,029 19,832 -,464 4,994 -38,684 -7,177 -337,506 -108,777 -3,476 -1,777 -34,455 -52,665 -71,173 -101,851
,464 52,665 79,597 9,048 3,505 95,903 102,815 61,653 86,094 -4,994 71,173 121,475 83,801 36,013 126,471 132,638 111,168 131,297 121,839 101,851 114,843 53,172 28,110 119,976 137,277 102,289 120,331 134,840 38,684
5 6 7 9 10
-31,833 85,417
*
1 2 3 4
dimension2
5 6 7 8 10
65,417 63,417
*
31,583 53,833 -142,167 -40,333 58,250 67,750 33,917 33,833 31,833 -31,583
10
1 2 3 4
dimension2
5 6 7 8 9
Based on estimated marginal means *. The mean difference is significant at the a. Adjustment for multiple comparisons: Bonferroni.
90
Setup
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 82,7500 78,2500 49,4167 56,5833 51,2500 68,4167 76,6667 120,0833 78,4167 121,6667 Std. Deviation 35,27586 31,51947 5,71216 30,62815 43,86576 25,51812 33,07384 34,08668 24,52998 30,45812 N 12 12 12 12 12 12 12 12 12 12
Pairwise Comparisons Measure:Setup (I) Spray (J) Spray Mean Difference (I-J) 1
dimension1 dimension2
95% Confidence Interval for Difference Std. Error 12,557 10,302 12,785 Sig.
a a
2 3 4
91
5 6 7 8 9 10 2 1 3 4 5
dimension2
31,500 14,333 6,083 -37,333 4,333 -38,917 -4,500 28,833 21,667 27,000 9,833 1,583 -41,833 -,167 -43,417 -33,333 -28,833 -7,167 -1,833 -19,000 -27,250 -70,667 -29,000 -72,250
* * *
15,885 12,258 17,645 17,378 10,347 14,836 12,557 9,547 12,135 12,876 11,741 16,066 11,754 10,805 14,153 10,302 9,547 8,133 12,964 6,831 9,448 10,442 6,313 8,837 12,785 12,135 8,133 13,011 10,912 12,395 15,378 9,526 11,891 15,885 12,876 12,964 13,011 15,951
1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,524 1,000 1,000 1,000 1,000 ,202 1,000 ,482 ,357 ,524 1,000 1,000 ,804 ,669 ,001 ,035 ,000 1,000 1,000 1,000 1,000 1,000 1,000 ,075 1,000 ,009 1,000 1,000 1,000 1,000 1,000
-37,972 -39,275 -71,086 -113,335 -40,920 -103,802 -59,417 -12,921 -31,403 -29,315 -41,517 -68,680 -93,238 -47,420 -105,312 -78,388 -70,588 -42,737 -58,530 -48,876 -68,569 -116,334 -56,608 -110,898 -82,083 -74,737 -28,404 -51,571 -59,557 -74,291 -130,754 -63,494 -117,086 -100,972 -83,315 -54,863 -62,238 -86,926
100,972 67,941 83,252 38,668 49,587 25,969 50,417 70,588 74,737 83,315 61,184 71,847 9,571 47,087 18,479 11,721 12,921 28,404 54,863 10,876 14,069 -24,999 -1,392 -33,602 29,750 31,403 42,737 62,238 35,890 34,124 3,754 19,827 -13,080 37,972 29,315 58,530 51,571 52,593
6 7 8 9 10
1 2 4 5
dimension2
6 7 8 9 10
1 2 3 5
dimension2
6 7 8 9 10
1 2
dimension2
3 4 6
92
7 8 9 10 6 1 2 3 4
dimension2
-25,417 -68,833
*
16,810 15,559 15,511 11,572 12,258 11,741 6,831 10,912 15,951 12,338 13,191 9,282 12,410 17,645 16,066 9,448 12,395 16,810 12,338 14,661 12,190 9,180 17,378 11,754 10,442 15,378 15,559 13,191 14,661 13,846 14,783 10,347 10,805 6,313 9,526 15,511 9,282 12,190
1,000 ,046 1,000 ,004 1,000 1,000 ,804 1,000 1,000 1,000 ,108 1,000 ,057 1,000 1,000 ,669 1,000 1,000 1,000 ,582 1,000 ,021 1,000 ,202 ,001 ,075 ,046 ,108 ,582 ,535 1,000 1,000 1,000 ,035 1,000 1,000 1,000 1,000
-98,934 -136,881 -95,003 -121,027 -67,941 -61,184 -10,876 -35,890 -52,593 -62,208 -109,358 -50,594 -107,526 -83,252 -71,847 -14,069 -34,124 -48,100 -45,708 -107,536 -55,062 -85,149 -38,668 -9,571 24,999 -3,754 ,786 -6,024 -20,703 -18,889 -66,236 -49,587 -47,087 1,392 -19,827 -40,670 -30,594 -51,562
48,100 -,786 40,670 -19,806 39,275 41,517 48,876 59,557 86,926 45,708 6,024 30,594 1,026 71,086 68,680 68,569 74,291 98,934 62,208 20,703 51,562 -4,851 113,335 93,238 116,334 130,754 136,881 109,358 107,536 102,223 63,070 40,920 47,420 56,608 63,494 95,003 50,594 55,062
-27,167 -70,417
*
-14,333 -9,833 19,000 11,833 17,167 -8,250 -51,667 -10,000 -53,250 -6,083 -1,583 27,250 20,083 25,417 8,250 -43,417 -1,750 -45,000
*
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9 10
1 2 3 4
dimension2
63,500 68,833
*
5 6 7 9 10
1 2 3
dimension2
4 5 6 7
93
8 10 10 1 2 3 4
dimension2
13,846 10,453 14,836 14,153 8,837 11,891 11,572 12,410 9,180 14,783 10,453
,535 ,074 1,000 ,482 ,000 ,009 ,004 ,057 ,021 1,000 ,074
-102,223 -88,965 -25,969 -18,479 33,602 13,080 19,806 -1,026 4,851 -63,070 -2,465
18,889 2,465 103,802 105,312 110,898 117,086 121,027 107,526 85,149 66,236 88,965
5 6 7 8 9
53,250 45,000
*
1,583 43,250
Based on estimated marginal means a. Adjustment for multiple comparisons: Bonferroni. *. The mean difference is significant at the
Standstill
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 324,0000 92,9167 45,0833 617,5833 747,0833 245,9167 327,3333 121,2500 ,0000 16,6667 Std. Deviation 345,67220 167,88711 13,42629 382,20329 510,38194 188,13895 204,48620 49,07346 ,00000 42,17999 N 12 12 12 12 12 12 12 12 12 12
94
Pairwise Comparisons Measure:Standstill (I) Spray (J) Spray Mean Difference (I-J) 1 2 3 4 5
dimension2
95% Confidence Interval for Difference Std. Error 96,530 98,772 112,798 110,102 93,455 129,802 100,411 99,787 101,576 96,530 50,323 93,632 124,322 31,980 86,815 54,927 48,465 41,856 98,772 50,323 110,922 148,032 55,844 58,365 Sig.
a a
Lower Bound -191,086 -153,058 -786,901 -904,609 -330,640 -571,020 -236,396 -112,415 -136,907 -653,253 -172,252 -934,164 -1197,886 -292,862 -614,099 -268,556 -119,043 -106,807 -710,891 -267,918 -1057,613 -1349,411 -445,064 -537,508
Upper Bound 653,253 710,891 199,734 58,442 486,806 564,354 641,896 760,415 751,574 191,086 267,918 -115,169 -110,448 -13,138 145,266 211,889 304,876 259,307 153,058 172,252 -87,387 -54,589 43,398 -26,992
231,083 278,917 -293,583 -423,083 78,083 -3,333 202,750 324,000 307,333 -231,083 47,833 -524,667 -654,167 -153,000
* * *
1,000 ,745 1,000 ,123 1,000 1,000 1,000 ,350 ,519 1,000 1,000 ,007 ,012 ,026 ,929 1,000 1,000 1,000 ,745 1,000 ,014 ,027 ,189 ,024
6 7 8 9 10
1 3 4
dimension1
5
dimension2
6 7 8 9 10
1 2 4
dimension2
5 6 7
-200,833 -282,250
*
95
8 9 10 4 1 2 3 5
dimension2
-76,167 45,083
* *
14,218 3,876 14,366 112,798 93,632 110,922 151,104 107,624 126,567 109,214 110,333 106,238 110,102 124,322 148,032 151,104 127,948 180,335 152,569 147,335 142,834 93,455 31,980 55,844 107,624 127,948 83,374 56,468 54,311 49,458 129,802 86,815 58,365 126,567 180,335 83,374 53,463 59,030
,010 ,000 1,000 1,000 ,007 ,014 1,000 ,243 1,000 ,038 ,007 ,007 ,123 ,012 ,027 1,000 ,108 1,000 ,079 ,016 ,015 1,000 ,026 ,189 ,243 ,108 1,000 1,000 ,039 ,032 1,000 ,929 ,024 1,000 1,000 1,000 ,121 ,008
-138,348 28,132 -34,413 -199,734 115,169 87,387 -790,347 -99,025 -263,285 18,690 135,047 136,289 -58,442 110,448 54,589 -531,347 -58,409 -368,941 -41,423 102,720 105,737 -486,806 13,138 -43,398 -842,358 -1060,743 -446,051 -122,293 8,389 12,947 -564,354 -145,266 26,992 -843,785 -1208,441 -283,218 -27,737 69,167
-13,985 62,034 91,246 786,901 934,164 1057,613 531,347 842,358 843,785 973,977 1100,119 1065,544 904,609 1197,886 1349,411 790,347 1060,743 1208,441 1293,090 1391,447 1355,096 330,640 292,862 445,064 99,025 58,409 283,218 371,626 483,444 445,553 571,020 614,099 537,508 263,285 368,941 446,051 439,903 585,500
6 7 8 9 10
1 2 3 4
dimension2
6 7 8 9 10
1 2 3 4
dimension2
-78,083 153,000
*
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9
96
10 8 1 2 3 4
dimension2
310,667
61,772 100,411 54,927 14,218 109,214 152,569 56,468 53,463 14,166 19,353 99,787 48,465 3,876 110,333 147,335 54,311 59,030 14,166 12,176 101,576 41,856 14,366 106,238 142,834 49,458 61,772 19,353 12,176
,017 1,000 1,000 ,010 ,038 ,079 1,000 ,121 ,000 ,010 ,350 1,000 ,000 ,007 ,016 ,039 ,008 ,000 1,000 ,519 1,000 1,000 ,007 ,015 ,032 ,017 ,010 1,000
40,510 -641,896 -211,889 13,985 -973,977 -1293,090 -371,626 -439,903 59,294 19,942 -760,415 -304,876 -62,034 -1100,119 -1391,447 -483,444 -585,500 -183,206 -69,919 -751,574 -259,307 -91,246 -1065,544 -1355,096 -445,553 -580,823 -189,225 -36,586
580,823 236,396 268,556 138,348 -18,690 41,423 122,293 27,737 183,206 189,225 112,415 119,043 -28,132 -135,047 -102,720 -8,389 -69,167 -59,294 36,586 136,907 106,807 34,413 -136,289 -105,737 -12,947 -40,510 -19,942 69,919
5 6 7 9 10
1 2 3 4
dimension2
5 6 7 8 10
10
1 2 3 4
dimension2
5 6 7 8 9
16,667
Based on estimated marginal means a. Adjustment for multiple comparisons: Bonferroni. *. The mean difference is significant at the
97
Idling
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 8,6667 ,0000 ,0000 43,6667 75,9167 7,4167 6,1667 1,1667 ,0000 16,6667 Std. Deviation 12,90760 ,00000 ,00000 56,37752 172,29280 7,35414 5,60573 1,26730 ,00000 42,17999 N 12 12 12 12 12 12 12 12 12 12
Pairwise Comparisons Measure:Idling (I) Spray (J) Spray Mean Difference (I-J) 1
dimension1 dimension2
95% Confidence Interval for Difference Std. Error 3,726 3,726 15,703 Sig.
a a
2 3 4
98
5 6 7 8 9 10 2 1 3 4 5
dimension2
-67,250 1,250 2,500 7,500 8,667 -8,000 -8,667 ,000 -43,667 -75,917 -7,417 -6,167 -1,167 ,000 -16,667 -8,667 ,000 -43,667 -75,917 -7,417 -6,167 -1,167 ,000 -16,667 35,000 43,667 43,667 -32,250 36,250 37,500 42,500 43,667 27,000 67,250 75,917 75,917 32,250 68,500
50,835 3,504 3,963 3,825 3,726 11,575 3,726 ,000 . 16,275 49,737 2,123 1,618 ,366 ,000 . 12,176 3,726 ,000 . 16,275 49,737 2,123 1,618 ,366 ,000 . 12,176 15,703 16,275 16,275 57,804 16,777 16,184 16,209 16,275 20,965 50,835 49,737 49,737 57,804 49,285
155,077 16,573 19,832 24,227 24,963 42,624 7,629 ,000 27,511 141,605 1,868 ,911 ,433 ,000 36,586 7,629 ,000 27,511 141,605 1,868 ,911 ,433 ,000 36,586 103,675 114,844 114,844 220,555 109,624 108,282 113,390 114,844 118,691 289,577 293,438 293,438 285,055 284,045
6 7 8 9 10
1,000 1,000
1 2 4 5
dimension2
6 7 8 9 10
1,000 1,000 ,958 ,958 1,000 1,000 1,000 1,000 ,958 1,000 1,000 1,000 1,000 1,000 1,000
-69,919 -33,675 -27,511 -27,511 -285,055 -37,124 -33,282 -28,390 -27,511 -64,691 -155,077 -141,605 -141,605 -220,555 -147,045
1 2 3 5
dimension2
6 7 8 9 10
1 2
dimension2
3 4 6
99
7 8 9 10 6 1 2 3 4
dimension2
69,750 74,750 75,917 59,250 -1,250 7,417 7,417 -36,250 -68,500 1,250 6,250 7,417 -9,250 -2,500 6,167 6,167 -37,500 -69,750 -1,250 5,000 6,167 -10,500 -7,500 1,167 1,167 -42,500 -74,750 -6,250 -5,000 1,167 -15,500 -8,667 ,000 ,000 -43,667 -75,917 -7,417 -6,167
49,288 49,879 49,737 51,996 3,504 2,123 2,123 16,777 49,285 2,118 2,333 2,123 12,998 3,963 1,618 1,618 16,184 49,288 2,118 1,723 1,618 12,497 3,825 ,366 ,366 16,209 49,879 2,333 1,723 ,366 12,049 3,726 ,000 . ,000 . 16,275 49,737 2,123 1,618
1,000 1,000 1,000 1,000 1,000 ,226 ,226 1,000 1,000 1,000 ,964 ,226 1,000 1,000 ,130 ,130 1,000 1,000 1,000 ,648 ,130 1,000 1,000 ,388 ,388 1,000 1,000 ,964 ,648 ,388 1,000 1,000
-145,810 -143,393 -141,605 -168,153 -16,573 -1,868 -1,868 -109,624 -284,045 -8,014 -3,952 -1,868 -66,095 -19,832 -,911 -,911 -108,282 -285,310 -10,514 -2,537 -,911 -65,155 -24,227 -,433 -,433 -113,390 -292,893 -16,452 -12,537 -,433 -68,195 -24,963 ,000 ,000
285,310 292,893 293,438 286,653 14,073 16,701 16,701 37,124 147,045 10,514 16,452 16,701 47,595 14,832 13,244 13,244 33,282 145,810 8,014 12,537 13,244 44,155 9,227 2,767 2,767 28,390 143,393 3,952 2,537 2,767 37,195 7,629 ,000 ,000 27,511 141,605 1,868 ,911
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9 10
1 2 3 4
dimension2
5 6 7 9 10
1 2 3
dimension2
4 5 6 7
100
8 10 10 1 2 3 4
dimension2
-1,167 -16,667 8,000 16,667 16,667 -27,000 -59,250 9,250 10,500 15,500 16,667
,366 12,176 11,575 12,176 12,176 20,965 51,996 12,998 12,497 12,049 12,176
,388 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-2,767 -69,919 -42,624 -36,586 -36,586 -118,691 -286,653 -47,595 -44,155 -37,195 -36,586
,433 36,586 58,624 69,919 69,919 64,691 168,153 66,095 65,155 68,195 69,919
5 6 7 8 9
Breakdown
Descriptive Statistics Mean HO_2 HO_3 AK_2 VI_1 VI_2 DP_1 DP_2 DP_3 AR_4 AR_5 ,5833 21,8333 16,8333 48,8333 ,0000 1,3333 23,0833 1,1667 13,2500 13,3333 Std. Deviation 1,44338 62,17692 15,60206 84,75079 ,00000 3,25669 60,71312 1,26730 25,70506 11,00688 N 12 12 12 12 12 12 12 12 12 12
101
Pairwise Comparisons Measure:Breakdown (I) Spray (J) Spray Mean Difference (I-J) 1 2 3 4 5
dimension2
95% Confidence Interval for Difference Std. Error 17,855 4,591 24,509 ,417 1,067 17,570 ,379 7,445 3,380 17,855 20,016 20,272 17,949 18,097 22,357 17,933 19,349 19,534 4,591 20,016 23,684 4,504 4,886 16,222 Sig.
a a
Lower Bound -99,339 -36,329 -155,437 -1,239 -5,416 -99,344 -2,239 -45,228 -27,533 -56,839 -82,537 -115,661 -56,666 -58,646 -99,028 -57,763 -76,041 -76,932 -3,829 -92,537 -135,582 -2,864 -5,868 -77,198
Upper Bound 56,839 3,829 58,937 2,406 3,916 54,344 1,072 19,894 2,033 99,339 92,537 61,661 100,332 99,646 96,528 99,097 93,208 93,932 36,329 82,537 71,582 36,531 36,868 64,698
-21,250 -16,250 -48,250 ,583 -,750 -22,500 -,583 -12,667 -12,750 21,250 5,000 -27,000 21,833 20,500 -1,250 20,667 8,583 8,500 16,250 -5,000 -32,000 16,833 15,500 -6,250
1,000 ,209 1,000 1,000 1,000 1,000 1,000 1,000 ,139 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,209 1,000 1,000 ,148 ,400 1,000
6 7 8 9 10
1 3 4
dimension1
5
dimension2
6 7 8 9 10
1 2 4
dimension2
5 6 7
102
8 9 10 4 1 2 3 5
dimension2
15,667 3,583 3,500 48,250 27,000 32,000 48,833 47,500 25,750 47,667 35,583 35,500 -,583 -21,833 -16,833 -48,833 -1,333 -23,083 -1,167 -13,250 -13,333 ,750 -20,500 -15,500 -47,500 1,333 -21,750 ,167 -11,917 -12,000 22,500 1,250 6,250 -25,750 23,083 21,750 21,917 9,833
4,637 8,877 5,308 24,509 20,272 23,684 24,465 24,714 28,115 24,548 24,788 25,124 ,417 17,949 4,504 24,465 ,940 17,526 ,366 7,420 3,177 1,067 18,097 4,886 24,714 ,940 17,615 1,058 7,691 3,177 17,570 22,357 16,222 28,115 17,526 17,615 17,635 20,065
,277 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,148 1,000 1,000 1,000 ,388 1,000 ,067 1,000 1,000 ,400 1,000 1,000 1,000 1,000 1,000 ,138 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000
-4,615 -35,242 -19,714 -58,937 -61,661 -71,582 -58,166 -60,586 -97,209 -59,695 -72,826 -74,379 -2,406 -100,332 -36,531 -155,832 -5,445 -99,734 -2,767 -45,703 -27,230 -3,916 -99,646 -36,868 -155,586 -2,778 -98,789 -4,459 -45,555 -25,893 -54,344 -96,528 -64,698 -148,709 -53,568 -55,289 -55,211 -77,922
35,948 42,409 26,714 155,437 115,661 135,582 155,832 155,586 148,709 155,029 143,993 145,379 1,239 56,666 2,864 58,166 2,778 53,568 ,433 19,203 ,563 5,416 58,646 5,868 60,586 5,445 55,289 4,792 21,722 1,893 99,344 99,028 77,198 97,209 99,734 98,789 99,044 97,589
6 7 8 9 10
1 2 3 4
dimension2
6 7 8 9 10
1 2 3 4
dimension2
5 7 8 9 10
1 2 3 4
dimension2
5 6 8 9
103
10 8 1 2 3 4
dimension2
9,750 ,583 -20,667 -15,667 -47,667 1,167 -,167 -21,917 -12,083 -12,167 12,667 -8,583 -3,583 -35,583 13,250 11,917 -9,833 12,083 -,083 12,750 -8,500 -3,500 -35,500 13,333 12,000 -9,750 12,167 ,083
17,864 ,379 17,933 4,637 24,548 ,366 1,058 17,635 7,412 3,123 7,445 19,349 8,877 24,788 7,420 7,691 20,065 7,412 9,159 3,380 19,534 5,308 25,124 3,177 3,177 17,864 3,123 9,159
1,000 1,000 1,000 ,277 1,000 ,388 1,000 1,000 1,000 ,112 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 1,000 ,139 1,000 1,000 1,000 ,067 ,138 1,000 ,112 1,000
-68,380 -1,072 -99,097 -35,948 -155,029 -,433 -4,792 -99,044 -44,498 -25,826 -19,894 -93,208 -42,409 -143,993 -19,203 -21,722 -97,589 -20,331 -40,140 -2,033 -93,932 -26,714 -145,379 -,563 -1,893 -87,880 -1,493 -39,973
87,880 2,239 57,763 4,615 59,695 2,767 4,459 55,211 20,331 1,493 45,228 76,041 35,242 72,826 45,703 45,555 77,922 44,498 39,973 27,533 76,932 19,714 74,379 27,230 25,893 68,380 25,826 40,140
5 6 7 9 10
1 2 3 4
dimension2
5 6 7 8 10
10
1 2 3 4
dimension2
5 6 7 8 9
104
105
106
Three Segments
107
108
Two Segments
109
110
111
Variance Homogeneity The variances of the respective samples must be homogeneous (Liang and Pan 2007), 2007) and this condition can be tested by means of Levenes Test in SPSS, as seen below.
Since the Levene statistic is insignificant at = 0.05, the hypothesis of variance homogeneity cannot be rejected, and the assumption is fulfilled. fulfi Independent Samples The samples must be independently collected from the population (Liang and Pan 2007), , meaning that no correlation of the error terms or correlation of the error terms and the covariates is tolerated. Since the OEE values are observed on the same spray towers in the same weeks, it can be questioned whether the samples are independent,
112
but since the OEE pattern retrieved from the FIMIX-PLS segmentation does not appear to depend on spray tower or week, the assumption is assumed to be fulfilled.
113
Multiple Comparisons OEE Bonferroni (I) Group (J) Group Mean Difference (I-J) 1,00
dimension3
95% Confidence Interval Std. Error ,04331 ,04557 ,04331 ,05481 ,04557 ,05481 Sig. ,874 ,174 ,874 1,000 ,174 1,000 Lower Bound -,0593 -,0234 -,1512 -,0918 -,1980 -,1746 Upper Bound ,1512 ,1980 ,0593 ,1746 ,0234 ,0918
2,00 3,00
2,00
dimension2 dimension3
1,00 3,00
3,00
dimension3
1,00 2,00
114