The Future of Statistics in Quality Engineering and Management
 By Professor Tony Bendell

1. Does statistics have a future in supporting quality improvement?
Although statisticians themselves are clear on the contributions that their discipline has made historically in various aspects of engineering and industry, others are less sure and statisticians themselves continue to be concerned that engineers do not take statistical literacy sufficiently seriously.

Is this fair?
The birth of the Royal Statistical Society owed much to the need to understand, estimate and control variation in industry. The application of experimental design methodology and analysis of variance, as a single example, has been crucial to understanding, optimising and controlling complex industrial processes. Statistical quality control and process control through the contributions of Shewhart, Deming and others have provided the basis for stable controlled industrial processes, and the development of statistical forecasting methods has provided an ability to forecast future sales, profits and difficulties. And there is so much more.

But somehow it has not all worked.
The current state of industrial application of statistics does not live up to its glorious creative past. We have statistical methodology but not the clarity of purpose or the market image that facilitates its use. When statistically based Quality and Reliability methods are used they are not always used correctly.

Taking stock of the current and future state of the application of statistics in quality improvement in industry and commerce, the first question must be do the old needs for statistics still exist? Have the questions changed? Technological change and a change of mind set have indeed clearly made their mark. Just as for most purposes we no longer need tables of logarithms, computing power and on-line instrumentation have converted some of the analysis, forecasting and control issues of the past into history. Of course, change has also created new problems, as old methodologies prove inadequate and new needs are revealed. But the problems are more complex, more complicated and more academic – and therein lies the real problem.

From being a well-founded routine analysis process, employing relatively large numbers of statistical assistants to carry out the laborious calculations of numbers according to an assured well-defined unchanging enumerative framework, statistics has evolved into an elitist, remote, obtuse – and for many in industry and commerce, unnecessary – set of approaches and people. There are two dimensions of this change. Firstly, the work of statistics has changed: computational power together with the switch in statistical emphasis from enumerative to analytic studies, which draw inferences about future but currently badly defined processes, mean that statistics has become less automatic and hence less accessible. Secondly, statisticians have not adequately tackled the consequent public relations challenge and, by some of their more “academically interesting” work, have added to the “bad press”. For example, O’Connor (1991), in arguing that statistical methods for quality and reliability prediction and measurement are counter-productive and should be discarded in favour of a return to traditional engineering and quality values, stated

“They lead to over-emphasis on expensive, bureaucratic and esoteric approaches to quality and reliability. Many successful equipment designers and manufacturers generate highly complex yet reliable products without recourse to these methods.”

Similar points were made by some respondents in the recent study by the Engineering Quality Forum on the Quality Education of Engineers (Cullen et al, 1997). The real question therefore is that, with all this bad press, does statistics and statisticians, have a future in industry and business at all?

The gap between the potential of statistics to help quality improvement and what is actually achieved is not a specifically British or specifically Western phenomenon. It was present also in the early days of the introduction of statistically based quality improvement methods into Japan and for much the same reasons (Ishikawa, 1985). There is evidence also of a similar situation in the USA and Germany (McMunigal et al, 1990; Bendell, 1994). The subsequent success of statistically based quality improvement methods in helping to transform the Japanese economy is evidence not just that this problem is solvable but that in bringing the messages and methods of statistics to the people in business and industry they do need to be clarified, simplified, communicated and “packaged”. The emphasis in Ishikawa’s work of the simplification, “packaging”, mass education, team basis and consequent mass use of statistical tools by all or many employees (eg the “seven tools of quality control”) reappears also in the work of other Japanese quality gurus such as Taguchi and Shingo (Bendell, 1991). The emphasis is on making statistical techniques understandable and usable by the customer and not on blaming the customer for not understand or not using them.

It can be argued that the need is not just to re-educate others: the engineers, managers and other professionals who just do not appreciate the importance of variation and the part that statistics has to play. Even more importantly, the need is to re-educate the statistical community. For statistics to have a future, now is the time for statisticians to come out of their closets, to cross the boundaries into the real work problems, to avoid unnecessary complexity, to start their role earlier in the project and to end it later, to become fully integrated, to lose their “statistician” stigma, to become the facilitators of large-scale simplistic routine application of statistical methods by all workers (like Ishikawa’s seven tools of quality control). There are few enough statisticians left in industry (eg Greenfield, 1996) and the need and the opportunities are strong; maybe never stronger. All around we see examples of the lack of use, misuse and abuse of statistics.

The responsibility is not only with those in industry. The role of academic statisticians, and the academic tradition of statistics, is also in much need of attention. It is here that the greatest elitism and barriers are created and carried on through education to future generations. And it is here that the greatest opportunity exists for a new ethos of statistical service; of removing the jargon, complexity and elitist barriers and of creating clarity, simplicity and focus.

To illustrate many of the points in this section, we shall now discuss three key areas of statistical application in quality.

2. Quality and performance

2.1 The impact of Taguchi
The principal role of statistics and statistical methods in industry and business is to improve performance and to provide support for quality improvement, including increasing productivity and product quality in manufacturing and improving service quality. Throughout the 20th century developments in the application of appropriate statistical methods have been made by Shewhart (1931), Deming (1986), Ishikawa (1976), Fisher (1925, 1935) and Box et al (1978) among others.

The quality revolution of the 1980s and 1990s in the West provided an opportunity for statistically based quality improvement techniques to gain a larger foothold and Taguchi methods appeared in Europe and the USA at this time. Consequently, owing to their timely arrival, they achieved more prominence in the industrial and business community than they may have otherwise achieved; one outcome of this has been that they have subsequently attracted more critique from the statistical community than the important, statistically well founded but less generally publicised, work of Fisher (1935), Box et al (1978) and Wheeler (1987).

Taguchi’s major achievement, however, has been in making experimental design and statistical techniques accessible to engineers without the need to understand the detail of statistical theory (Taguchi, 1987). His techniques and their application (Bendell et al, 1989) have been critically reviewed (eg Box et al, 1988, Logothetis and Wynn, 1989, and Nelder and Lee, 1991, among others) and the Quality Improvement Committee of the Royal Statistical Society has initiated debate on their viability. However, the fact remains that Taguchi has introduced many engineers to experimental design for the first time and made them aware of the importance of systematically designing quality into a product or process rather than inspecting it out. Before Taguchi’s impact, many engineers and managers engaged in fire-fighting exercises to provide a short-term solution when the need arose, or conducted experiments involving one factor at a time (as some still do!), which ignored all interactions or sometimes focused their attention on just a small number of factors which could be handled within a full factorial design.

Nevertheless, Taguchi experimentation has not always succeeded in the West. Many of the organisations that have implemented Taguchi methodology or other experimental design techniques have been disappointed with the results and have consequently blamed the techniques when in many cases it is the management of the implementation and the quality culture of the organisation which are at fault. Bendell et al (1990) cited an automotive component manufacturer and a chemical processor who both experienced failure with Taguchi methods due to an over-optimistic expectation of their ability to handle experimentation alongside other demands on manufacturing plant and time, whereas Disney (1996) described a paint manufacturer whose experimentation was ruined by an inability to measure the performance indicator accurately.

Taguchi experiments are more manageable and predictable and require fewer resources than full factorial designs, so it is much easier to plan their implementation, in that results are available at the most opportune time.

It would, however, be misleading to suggest that all Taguchi’s proposed techniques are the most efficient or powerful tools or that they are over-appropriate in many situations. Furthermore many of the companies which have implemented Taguchi methods have only embraced the basics. Even highly successful initial experiments are not always followed up, as illustrated by the food packaging case-study discussed by Disney (1996). Many experiments that are conducted use only the most popular orthogonal arrays such as L4, L8, L9, and L16.

2.2 Simple statistical tools
There appears to be a tendency for statisticians to use overcomplicated tools and techniques. This may result from their awareness of the sensitivity of the underlying assumptions or it may be seen as an attempt to justify their role in an organisation! However, rather than convincing the organisation of the merit of having a statistician, this often has the opposite effect if no-one else can understand the statistician’s output.

The seven tools of quality (Ishikawa, 1976) and basic exploratory data analysis techniques (Tukey, 1977) remain the most powerful and effective tools, together with graphical data presentation methods, yet they are often neglected as trivial. The seven new management tools of quality control (Mizuno, 1988) have a technical organisational basis rather than a statistical foundation, yet they link in very well with many basic statistical techniques and should not be ignored by statisticians. These new tools are as follows: the relations diagram method; the K3 method (affinity diagram); the systematic diagram method; the matrix diagram method; the matrix data analysis method; the process decision programme chart method; the arrow diagram method.

Many statisticians find themselves called on to “troubleshoot” or “fire-fight” and their first task is to identify the major problems. The Pareto chart, applied to data obtained through basic data collection methods, is a simple and effective tool and has the advantage of being visible and easily explained. The next task is to identify the cause of these problems and brainstorming will lead to a cause-and-effect diagram; drawing a flow chart of the process will often lead to the identification of bottle-necks that may in themselves be causes of trouble. The identification of common and special causes of variation in the process can also lead to significant improvements, especially if the effects of common causes can be reduced (Deming, 1982). Many errors can be prevented by using the poka yoke approach to error prevention developed by Shingo (1986), who after 20 years of statistical quality control declared in 1977 that he was “finally released from the spell of statistical quality control methods” (Shingo, 1985).

SPC techniques can be applied to most processes in a manufacturing setting and many service sector activities. Although statisticians may argue over the precise meaning of charts (Alwan and Roberts (1995) and contributors to the discussion), the fact remains that SPC is a powerful tool for identifying stable and unstable processes. In many organisations, however, it is used as a “historic tool” to show where problems occurred, yet if it is applied on the shop-floor it is a simple method for signalling the need for corrective action which can be taken by the operator, possibly preventing the production of defective items in a manufacturing line or the delivery of a substandard service.

To properly promote the use of these tools, statisticians should communicate their benefits with clear examples of quality and productivity improvements coupled with bottom line savings. With the current range of software that is available, there has never been a better time to introduce simple statistical methods into manufacturing and service organisations, yet there is still little take-up of these techniques in most companies.

Just as operatives are now expected to be multiskilled, maybe working in a cellular manufacturing system rather than on a conventional assembly line, so Quality statisticians should be multitalented. They must be able to assist in the design, planning, production, quality control, distribution, sales and customer service functions, being equally adept at designing experiments and analysing results, forecasting sales figures and interpreting market research. They must be willing to use the seven new management tools, Kanban (Wild, 1995), the five Ss (seiri, seiton, seiso, seiketso and shitsuke, translated as CAN-DO meaning cleanliness, arrangement, neatness, discipline and order), and Kaizen (Japan Human Relations Association, 1992) in addition to more overtly statistical tools in order not only to sustain but also to prosper.

2.3 Process analysis and control
Statisticians need to be careful not to become more involved with the detail rather than the message behind the details. With this line of reasoning what has SPC to do with the normal distribution? As Henry Neave has frequently and eloquently pointed out, Shewhart never intended that one should be the foundation of the other. Statistics and statistical education have become obsessed with the minutiae, missing the message. Although theoretically based, W Edwards Deming’s and Walter Shewhart’s contribution in spreading control charts was about controlling real processes not about statistical algebra (Palm et al, 1997). Strangely, in the way that SPC is so frequently taught at UK universities and abroad, it is the algebra that is retained. We then seem surprised that SPC is not applied or is applied clearly incorrectly. Indeed, the current extent of bad application is extensive (Shaw et al, 1997).

It is now widely recognised that there is a very real need for process analysis management and control, not just in the context of manufacturing processes, but of all business processes. This process emphasis is the centre of the modern concept of quality management and is reflected in the formulation of the EFQM Excellence Model, business process re-engineering methodology and benchmarking activity (as described, for example, in Bendell et al, (1993)). But the idea is not new – it was inherent in the early approaches of Deming and the other “missionaries of quality” to Japan in the early 1950s and later in the early concepts of total quality management by the US Department of Defence. Two quotes from W Edwards Deming illustrate the simplicity, clarity and importance of this message (Deming, 1993):

“Draw a flow chart for whatever you are doing. Until you do, you do not fully understand what you are doing. You just have a job.”

“The first step in any organisation is to draw a flow diagram to show how each component depends on others. Then everyone may understand what his job is. If people do not see the process they cannot improve it.”

Clearly, with the current emphasis on improving business processes, the implication is that there is a very real potential for a widely developing use of process analysis, process effectiveness, measurement and control. Equally clearly, some statisticians would see this as “not proper statistics”. That is unfortunate: it is consistent with Shewhart’s and Deming’s original purpose and takes us beyond the level of statistical algebra.

3. Standards and awards
At the beginning of this paper, we asked the question whether statistics has a future in supporting quality improvement and argued that statisticians’ attitudes were potentially the biggest danger to the existence of that future. The debate about BS EN ISO 9000 is a classic illustration of that problem.

In January 1994 Adrian Stickley and Alan Winterbottom read a paper to a meeting jointly organised by the Business and Industrial Section of the Royal Statistical Society and City University on “The nature of quality assurance and statistical methods in BS 5750” (Stickley and Winterbottom, 1994). Rather than concentrating on the opportunities that the international quality systems standard, BS EN ISO 9000 (formerly known in the UK as BS 5750 and referred to hereafter as ISO 9000) offers for statistics and statisticians to be employed in industry and commerce, the paper attacks the standard! Actually, attacking ISO 9000 is very easy, and indeed very commonplace in the quality improvement literature, but does not contribute to gaining entry points for statistical application. The standard is here, and here to stay, so why debate whether it should be? Would it not be more productive to examine what it contains of a statistical nature and to use that as an entry point for statistical application, while still arguing for further developments?

In fact, as pointed out in the discussion of Stickley and Winterbottom (1994) by us and others, ISO 9000 does not potentially contain substantive statistical requirements! Clause 4.20 now reads as follows (British Standards Institution, 1994):

4.20 Statistical Techniques

4.20.1 Identification of need
The supplier shall identify the need for statistical techniques required for establishing, controlling and verifying process capability and product characteristics.

4.20.2 Procedures
The supplier shall establish and maintain documented procedures to implement and control the application of the statistical techniques identified in 4.20.1.

This requirement for the establishment and maintenance of process capability provided a wonderful opportunity for statisticians to extend the application of statistics in industry, making use of the enormous growth of the number of companies and other organisations which are certified to ISO 9000. It must be said that following the 1994 revision to ISO 9000 the accredited certification bodies have been slow to realise the full implications of this clause and companies can and do evade the statistical requirements.

The new version if ISO 9000 – ISO 9000:2000 – places even more emphasis on measurement based improvement but probably sensibly does not mention statistics.

As well as ISO 9000 itself, there are an increasing number of sector-specific standards which are based on ISO 9000 but go further in terms of statistical requirements. Two examples are QS 9000 systems for automotive component manufacturers, promoted by Chrysler Corporation, Ford Motor Company and General Motors Corporation (1995), and the technology and process approval procedures for the Cenelec Electronic Component Committee 90,000 system for electronic component manufacturing.

More interestingly perhaps are the implications of the EFQM Excellence Model for the use of statistics. This is a nine-criteria, 32 sub-criteria, basis for organisational self-assessment that is increasingly being used in business, industry and the public sector. The Model requires process analysis and measurement, trend analysis, benchmarking and customer and employee perception surveys. In the USA the Malcolm Baldrige National Quality Award Model has a longer history and has already had a major effect on the introduction of good statistical practice.

4. Reliability
At present, although not readily acceptable to all engineers, much of the application of statistics in engineering reliability is through the use of exploratory methods to show the data structures which appear in test and field failure data. The work of Walls and Bendell (1995) in exploratory data analysis, Bayesian methods (Bunday, 1991) and counting processes (Fleming and Harrington, 1991; Thompson, 1988; Crowder et al, 1991; Ansell and Phillips, 1989), which includes proportional intensities (Lawless, 1987), additive hazards (Pijnenberg, 1991), proportional hazards (Cox, 1972) and generalised linear model (McCullagh and Nelder, 1989), provides approaches which have aided the engineer or manager in understanding the possible causes of failure in systems. Examples of these analyses have been carried out by Drury et al (1987), Kumar and Klefsjo, Lawless (1987) and Wightman and Bendell (1985, 1995).

However, as systems become more reliable because of improving technology, there will be an increasing paucity in data to be analysed; hence the approach to statistics must change to more diagnostic analysis within the systems development process, Feed-back of statistical results back into systems design is like shutting the door after the horse has bolted because the implication is that, if a statistical analysis on failure data has been carried out, the systems were unreliable anyway!

There must be a movement away from statistical reliability requirements within system specifications that are difficult for engineers and managers to interpret and cost. For instance, instead of asking for a mean time between failures figure of 1,000 hours, a statement should be made about the failure-free operation period (Knowles, 1996). Thus instead of laboriously calculating some failure rates from a standard such as MIL-HDBK-217, scientific method should be used to determine and reduce the causes of in-service failure in the design stage. This will supplement the total quality management philosophy of failure prevention that currently utilises techniques such as Pugh design selection criteria, quality function deployment (QFD), failure modes, effects and criticality analysis (FMECA), design review and designed experimentation (Clausing, 1994; Rommel et al, 1996).

The cost implications of reliability (or lack of it) are not fully understood. The identification and removal of failures at the design stage requires a good customer-supplier interface, reliable internal communications and a quick turnaround of information. There is a need to develop a statistically based management approach to reduce variability (Baer and Dey, 1989) and an initiative to train staff to think of variability reduction not only to business and industry but also to the statistical community at large involved in teaching undergraduate engineers and business students. This approach requires a detailed knowledge of total quality management and elements of statistics such as causes of bias, effects of strong interaction, stratification, correlation and the scientific method to highlight the difference between confounding, common response and causality.

Some problems from not finding failures sufficiently early arise for the following reasons: the product configuration has been decided without the input from the reliability engineer or statistician, a lack of understanding of the use of redundant versus highly reliable systems; hazards have not been identified sufficiently early for preventive action (there is a lack of emphasis in using reliability techniques such as FMECA rather than after-the-fact testing such as Weibull analysis); reliability has been insufficiently costed into the contract (how much does it cost to run a sequential probability ratio test (SPRT) given that the time to completion is not fixed?); reliability demonstrations have been run before all design problems have been removed. The papers by Bain and Engelhardt (1982) for reliability growth SPRT and Harter and Moore (1976) and Vujanovic (1994) for Weibull distribution SPRT have not been used in industry to alleviate these demonstration problems in the author’s experience. Add to this list that the manager has not allocated sufficient funds to reliability activities, the reliability activities have started too late, repeated activities and important information has not been disseminated, and you can see why the reliability engineering community is not as well respected as other engineering disciplines. The main problem is that statistics on these problems are not available and so possible solutions are not addressed.

Training in risk management techniques is greatly important for a prospective manager as an understanding of the failure concerns of an experienced reliability engineer or statistician must be taken into consideration (see the commentary on the Challenger disaster in Feynman (1989)). This training required in statistics and risk inevitably must come about from the cost and legal implications of product recall, reduction in product development time, liability claims and Government and company directives such as the management of health and safety at work (MHSW) regulations (Health and Safety commission, 1992), QS 9000 and BS EN ISO 9001 which companies do have to comply with.

The standard practice of making equipment suppliers know that reliability is important to a customer is to include some statement of reliability in the specification or contract. This statement can be wide ranging and may include various tasks to be carried out by the supplier such as reliability prediction, FMECA, fault tree analysis and/or reliability demonstration. Many small suppliers do not have the knowledge or expertise to carry out these tasks, let alone to cost them into a contract. In the main, lip-service is still paid to reliability as can be seen in many company brochures or marketing pamphlets with no effort to quantify statements, such as “We offer a highly reliable service”.

Large companies are specifying reliability statements into contracts now because of bad experience of unreliable subsystems in the past. However, unreliable subsystems usually come to the notice of a customer when the system is well into use. The problem is that, unless there is a commitment on the behalf of the supplier’s management to use the reliability tasks to their advantage, they will only do the bare minimum to satisfy the customer requirements. In the same way that the Deming and Shewhart philosophies have been watered down by companies to reduce costs in the short term, so also have reliability activities.

The approach of automobile manufacturers and other large manufacturing companies to make all suppliers satisfy QS 9000 or BS EN ISO 9001 is the first step to improving supplier management commitment and hence reliability because all the reliability techniques such as failure reporting and corrective action systems and reliability development testing (military standard MIL-HDBK-338; US Army Communications Research and Development Command (1984)) require a well-documented quality system as a prerequisite.

QS 9000 is a good start for suppliers to learn and use techniques such as QFD and FMECA. However, suppliers having the knowledge of a technique and their proper use of that technique are very different. The use of FMECA, for example, in the automotive sector does not suggest that things will improve in the future. The main problems lie in the fact that it is applied too late to have much effect, the risk priority numbers are manipulated to reduce preventive action costs not to implement preventive actions at the design stage, the method is not managed well, the technique is too time consuming and it requires a detailed knowledge of not only how a product works but how it fails as well, something that only experienced engineers will know about the prospective product.

One solution that various Japanese automobile manufacturers use is to have a resident engineer on the supplier’s premises to deal with problems as they arise. Two other solutions are to tie in the FMECA with the requirements of the MHSW regulations 1992 since every company now must legally carry out a risk assessment (Health and Safety Commission, 1992) and use the technique as a prerequisite for design experimentation to reduce variability and incorporate life-cycle cost considerations into the FMECA format. The approach of making the risk priority numbers (or developing a risk approach) more robust against misuse has not yet been considered in the research literature. Larger companies should incorporate FMECA usage within their quality, environmental and safety and health strategy as hazard analysis critical control point, product, process and design FMECA, and the risk identification procedures specified in Croner Publications (1998) to meet the MHSW regulations are all similar in format.

Some of the areas where statistics and statisticians can provide improvements in the reliability field are given below.

All statistical reliability requirements which are put forward by major companies should be addressed by a common standardised approach with the cost implications. This will reduce the use in specifications of misleading terms such as failure rate, availability and even the definition of failure.

Reliability prediction using standard databases such as MIL-HDBK-217 should be discouraged as the analysis should only be used as a basis for comparing supplier specifications. However, the use of component reliability models and derating criteria for components listed in MIL-HDBK-217 should be encouraged as they are extremely useful in determining and reducing the causes of failure. Databases of in-service failure data are useful for determining the actual life characteristic of components and more work is required in this area (Wightman and Bendell, 1994; Landers and Kolarik, 1987).

Failure reporting and corrective action systems are usually incomplete and do not usually run well. This is a possible growth area for teaching statisticians as well-run, well-documented failure reporting and corrective systems with the prevention action approach of Shewhart will save companies the cost of failure in service use and will put the statistics into context with engineering graduates.

All testing is expensive and in many cases irrelevant (Nelson, 1995) as it does not reflect in-service use. Bayesian techniques incorporating test results from different tests, less emphasis on meeting requirements and more on system improvement (exploratory statistical tools) and new modelling approaches to the use of environmental stress screening and burn-in are all possible research areas for system improvement.

The approach of demonstrating the reliability of a product is carried out too late for some types of failure that occur. For instance, if the failure is due to some design problem on a power supply it may require a complete redesign to eliminate the cause of failure. The management of reliability at every stage of development should be audited to determine where and why trade-offs were made, eg performance, cost, which tools were used and how effective they were at improving reliability. Reliability must be designed in by considering it at the requirements stage (Akao, 1990) by the use of QFD and then making the product more robust to its intended environment (Clausing, 1994).

Tools such as FMECA, failure modes and effects analysis and fault tree analysis will come to the fore before equipments are even built and more statistical research input is required into the relationship between the risk and the cost of hazardous events within these methods. The use of the functional analysis system technique and reverse fault trees (Clausing, 1994; Fox, 1993) are techniques that lead to designed experimentation for a possible optimum design solution. Not enough work has been done on integrating these tools with designed experiments.

So far as designed experiments are concerned, engineers still use ad hoc methods to determine optimum solutions in finite element analysis, electronic circuit simulation, automatic dynamic analysis of mechanical systems and parametric feature-based solid modelling (Bigelow, 1995). The application of designed experiments with simulation in computer-aided design or computer-aided engineering needs to be addressed with the aid of statistical computer software to aide the designers to find the most robust product solution (Grove, 1997). The work of Quinlan (1987) and Logothetis and Wynn (1989), pages 334-345, provides approaches for optimising a design using designed experiments applied to finite element analysis and electronic circuit simulation respectively. At a recent validation of the Masters’ course in engineering at The Nottingham Trent University, a module on designed experiments and simulation was specifically required by a major automotive supplier. There needs to be more work in this area.

The future of reliability engineering lies in the hands of those companies who can integrate their design activities with their failure prevention activities and statisticians need to become involved in these activities.

To quote O’Connor (1991):

“Recognise that high quality and reliability are achieved by good management, in the widest sense. Good management of engineering includes paying attention to excellence in design and production, adopting a totally integrated product directed team approach, and a commitment to training at a level that is far in excess of that currently practised by western companies. This training must include a thorough grounding in appropriate industrial statistical methods and applications.”

He adds:

“eliminate all methods that distract from the pursuit of excellence. Statisticians working in the Q&R field should review their work against this criterion.”

5. The future
Much of this paper has been about the mismatch between the wants and needs of industry and the statistician’s view. The former want simplicity and parsimony whereas the latter, at least sometimes, wants academic respectability, interest and hence complexity of these needs.

There has been much recent debate in the profession about exactly which quality concepts and tools an engineer needs. A recent investigation by the Engineering Quality Forum on behalf of the Engineering Institutions (Cullen et al, 1997) reported that an undergraduate degree course should provide all engineers with a basic grounding in a broad range of quality issues, but this teaching should be integrated with the engineering disciplines rather than be taught as specialist stand-alone modules.

The survey also found that currently practising engineers would benefit from a broad quality education primarily focused on addressing business and organisational issues and directed at improving the management and implementation capability. Overall the findings of this survey strongly suggest that existing education does not deliver such quality tools and techniques in a manner which facilitates their application.

In the business community the need for classical statistics has been replaced by a need for the simple application of tools and problem solving techniques. This is in line with Ishikawa’s ranking of statistical methods (Ishikawa, 1985):

a) Elementary – every worker knows how to use the seven tools of quality control (Ishikawa, 1976).
b) Intermediate – managers additionally have knowledge of the seven new tools (Mizuno, 1988), theory of sampling surveys, statistical hypothesis testing, methods of sensory testing and basic design of experiments.
c) Advanced – the elite few have knowledge of advanced experimental design techniques, multivariate analysis and operational research methods.

The need for computational skills has been made redundant by advances in computing; indeed the worker at any level no longer needs to know how to perform a hypothesis test but how to interpret the outcome of it.

The result of this is that, although business still has a need for statistics, it no longer in general has a need for statisticians. In the USA, General Electric, Motorola and Milliken have created the role of the “quantitative engineer” and have been very successful in obtaining highly competent facilitators with statistical skills at the engineer’s level. Such facilitators have implemented variation reduction programmes as proposed by Deming and Ishikawa. Indeed the “six sigma” programme originated by Motorola has often been undertaken without statisticians being involved in the training, adding weight to the argument that, if you want a successful quality programme, keep the statisticians away!

The role for specialist statisticians is at the “top of the pyramid” for higher education, research and training. They have proved that they have no part to play in mass education or use of statistics. In discussing the US situation Hahn and Hoerl (1998) state that the “moves towards pro-activity in statistical quality methods have left the statisticians who have been singularly non-proactive behind”.

The same arguments raised in the USA apply to Europe and the UK in particular. One of the positive developments in the UK has been the involvement of the Royal Statistical Society with other bodies investigating statistical methods in quality management and engineering, eg the Engineering Quality Forum. However, those activities have been insignificant when compared with the momentum over the past decade of the entire quality movement. There is a need to research the developing themes in quality to offer statistical support in the client’s own language and without complexity.

How can we shape the future dimensions and role of quality in engineering and management? Who should the players be? The older generation of statisticians (and it is apparently an ageing profession) will not find this erosion of academic and methodological rigour acceptable; we need to attract a new generation of young statistical activists, but where will we find them or how shall we create them?

The Royal Statistical Society has a dearth of such competent practitioners. We need to challenge a few practices about the roles of both statistics and statisticians. There is a need to create a new kind of education programme within our universities leading to the emergence of the competent qualified quantitative facilitator who truly understands the nature of variation. Perhaps now is finally the time to drop statistics (or at least the statistician) from our quality vocabulary.

Juran (1994) has forecasted that the 21st century will be referred to by future historians as the “Century of Quality”. Quality improvement will be dependent on the application of statistical techniques irrespective of who facilitates this implementation.

This paper is based on the paper by Bendell et al in the Statistician (1999) 48, p299-326

top of page