[Err @BUSINESS Operations Management Roy D. Shapiro, Series Editor (ESS + neractive itiustrations Managing Quality with Process Control ROY D. SHAPIRO. HARVARD BUSINESS SCHOOL 8020 | Published: September 12, 2013 Copyright © 2014 Harvard Business School Publishing Corporation. All rights reserved1 INTRODUCTION he importance of quality—whether in the production of goods or the delivery of services—should be self-evident. Meeting or exceeding customers’ expectations is a prerequisite for successfully participating in any competitive market. Quality is an intuitive concept. However, because it may be defined in part by customers’ subjective assessments and because many factors contribute to quality, measuring it can be complicated. Determining the appropriate level of quality, delivering it consistently, and identifying ways to improve it require managers to exercise judgment and use the right analytical tools. This is the essence of quality control, the term we use to describe all the activities undertaken to control materials, processes, and products in order to ensure quality. Quality control has broad application. A producer of jet engines must determine and meet the specifications of every component and understand the criteria for rejecting a part. A beer manufacturer needs to ensure that every bottle has the same taste and freshness and is filled to the same level. A fast-food restaurant should provide a consistent experience for its ‘customers by processing orders quickly and accurately, preparing food correctly, and delivering it at the right temperature. Note that in these examples we are concerned with conformance quailty. Indeed, that is the primary goal of quality control, Performance quailty is typically within the purview of product design. Quality control requires management to answer the following questions: ‘* How should we measure quality? ‘* How and when should we test for quality? Are our processes under our control, and do they deliver consistent results? ‘© Do our processes deliver the quality customers expect? ‘* How do we improve our quality? ‘As we discuss in the reading on quality management, a central principle of Total Quality Management is that instead of inspecting a product only at the end of a process (an activity we can think of as “product control"), we must monitor individual process steps to ensure that the process is running the way we want it to. This is the essence of process control, a crucial aspect of quality control and the focus of our discussion in this reading. No business intentionally creates a low-quality product or service (one that does not conform to specifications). When poor quality results, the culprit is often unexpected variability. Variability in the inputs (such as raw materials), process parameters (machine settings, forinstance), or execution will cause variability—and hence, quality problems—in the output. If one shipment of hops is not as fresh as others, the beer manufacturer might produce a subpar batch. If workers on one shift are not as highly skilled as workers on another, the jet engine manufacturer might produce some flawed products. Process control has two objectives. First, it seeks to detect variability measuring and controlling it, establishing systems to flag situations where it exceeds defined bounds, and thus discovering problems before they reach the customer. By carefully monitoring process steps, we can detect process problems before they result in product defects, Second, process control seeks to reduce variability by improving product and/or process design, inputs, or process execution, thereby decreasing the probability that unexpected variance would lead to quality problems. This reading will introduce key concepts and considerations in applying process control and will provide an overview of the most commonly used statistical and analytical tools, which we refer to as Statistical Process Control (SPC). SPC is applicable to all processes, whether in manufacturing or services, whenever there is a measurable quantity that is important in ensuring a high-quality product or customer experience. The thickness of a precision metal part, the temperature of a meal, the length of a customer service call, the number of potholes in a city street—all can be measured and controlled. 2 ESSENTIAL READING Process control starts with the measurement of an important variable. This can be a product attribute such as the diameter of a metal component, the number of customers who say they are “very satisfied’ on a customer satisfaction survey, the weight of a bag of potato chips, and 0 on. Alternatively, we might choose to measure a process attribute, such as the temperature ina restaurant's oven, the pressure applied in a glass molding operation, the wait time to speak with a customer service rep, or the length of the line at a ticket booth, Statistical Process Control is mostly concerned with the chosen variable’s mean (the process mean) and variability, The usual measure for variability of a probability distribution is its standard deviation (the process standard deviation), which measures the dispersion—or “spread’—of data around the mean.1 In practice, it's impossible to measure the process mean and standard deviation because that would require measuring all possible process outcomes, Instead, we take a number of different samples over time, measure the sample means and measures of sample variation, and estimate the process mean and standard deviation from these, Imagine a manufacturing process that requires a component's length to be 100 millimeters + 1 millimeter (in other words, between 99 and 101 millimeters). This specification ensures that the component wil fit well with other parts of the product. Interactive illustration 1 illustrates theimpact of changes in the mean and standard deviation of a quality metric (the component's length) on the quality of production, as measured by the defect rate. Q INTERACTIVE ILLUSTRATION 1 Defect Rate ‘eases soecteg aioe roe len mt + coor tan 925 mm ane (cites ae ated yee Fee poaucton ot 1.000,000 units: Detect Rate ne tat ar te ong “otal defecive units ‘PRODUCTION PROCESS MEAN ‘PRODUCTION PROCESS STANIARD DEVIATION The probability distribution curve in Interactive Illustration 1 characterizes the production of this ‘component and provides a useful demonstration of variability. You can change the mean and the standard deviation and observe how this changes defects and defect rates. Greater variability (that is, a larger standard deviation) leads to a flatter curve with longer other words, production is less consistent overall and there is a higher probability that the product falls outside the specifications (the areas in red). Reduced variability leads to a narrower distribution curve with shorter tails—produetion is clustered more closely around the ‘mean and fewer units fall outside specifications. in Interactive Illustration 1 demonstrates that ‘© The defect rate increases as variability (standard deviation) increases; more variability means that a higher number of units fall above or below the specification limits. * The defect rate increases as the mean of the measurements moves away from the mean of the specification range, toward either the upper or lower specification limit; this increases the probability that the measures of the quality metric will fall outside the limits.Itis helpful to think of process control as measuring, managing, and reducing the talls—or the variabilty—of the distribution curve and ensuring that the mean does not shift. The key concepts in process control pertain to the sources of variabilily and its implications for business processes and customers. Before we look at the tools of Statistical Process Control, we must say a few words about product inspection, which in most processes is done in tandem with process control The most common form of product inspection is acceptance sampling, the inspection of goods before, during, or after production, such as the testing of a circuit board to check ‘connections before it is incorporated into a computer or the examination of manufactured garments for stitching quality. This is an important aspect of quality control but is very different from process control, which is the ongoing monitoring of a process to ensure that its performance is stable and meets certain criteria, PW Noted) o) ela Br yl] Acceptance sampling requires managers to make two key decisions: when to conduct inspections and how rigorously to test. The answers depend entirely on the individual business, which must weigh the costs of inspection against the risks (and costs) of defective product. Because acceptance sampling involves inspection, problems are spotted after they have occurred. Thus, quality problems that are missed are passed on either to a subsequentprocess, which can be highly disruptive to operations, or to customers directly, which may be very costly. These dual costs—of defects and of inspections—are the key considerations for determining how a business conducts acceptance sampling: ‘* When in the process to inspect depends on the point at which an inspection has the greatest benefit at an acceptable cost. Typical inspection points include upon receipt of raw material inputs, at the end of a process, before a particularly costly step, or before a step that will make later inspection difficult (for example, a step that covers or seals a set of components). '* How rigorously to inspect can be divided into two related matters: the frequency with which to test and how many products to test each time. The rigor depends on the degree of confidence needed concerning output quality. Less than 100% acceptance sampling is, employed when testing is destructive to the product, the cost of 100% inspection is too high, or 100% inspection takes too long ‘A number of statistical methods can be used in acceptance sampling, such as simple random sampling, systematic sampling, and cluster sampling. The details of these are beyond the scope of this reading, but all methods share the same overarching objective: minimizing the ‘sum of inspection and defect costs. The traditional view of quality is that acceptance sampling is optimized when the total cost of inspections and of passing on defectives to customers is minimized, as illustrated in Figure 1. FIGURE 1 Optimal Amount of Inspection Cost optimal Total Cost v. Cost of Inspection Cost of Passing Defectives Amount of Inspection‘The shapes of these curves differ significantly by situation. The costs of a defective paper lip, a dry ballpoint pen, or an egg cracked in the carton are relatively low. Because conducting ‘exhaustive inspections would probably not be worthwhile in these cases, a final inspection of a ‘small sample of products might suffice. On the other hand, the costs of passing on a defective airbag deployment system in an automobile or a faulty implantable cardiac defibrillator would be very high indeed, so every product would be tested. 2.2 Statistical Process Control (SPC) SPC encompasses a set of statistical tools that allows us to monitor a process, see whether it is stable, and distinguish pure “randomness" from variation that we can track down and eliminate. SPC rests on four key concepts: Common variation (also known as natural variation oF noise variation): This is the “random” variation that is inherent in a system or process. Causes may be difficult to distinguish or wholly unidentifiable, and the resulting degree of variation may be minor. An example of common variation would be a bow! of soup that is hotter or colder than normal owing to inherent (and uncontrollable) fluctuations in the ambient temperature, or the variable length of a component made on an old machine that is not able to maintain the desired tolerances. Reducing common variation usually requires management intervention and may require the investment of capital (to buy a new machine, for instance). Assignable variation (also known as special variation or signal variation): This is variability that is not inherent in a process or system. It arises when something has changed in the process itself, and the change, typically after careful study, can be assigned to a specific cause. Reducing this type of variation usually involves identifying the root cause and making process changes either to eliminate it or, ifit cannot be eliminated, to allow the process to ‘accommodate it. A bow! of soup that is too hot owing to a new oven setting or a component that is too long because of a worker's inattention are both examples of changes to the process that lead to assignable variation. Whereas common variation is pervasive and affects all production unless the variability is addressed, assignable variation ocours when something in the process has changed. Control (or in control or stable): A process that exhibits only common variation is said to be in control, or stable. A process is deemed out of control when it exhibits assignable variation. When a process that is in control goes out of control, it means that something has happened (such as a new, less experienced worker has replaced a more experienced worker, there was a machine malfunction, a machine setting was changed, and so on). Process control allows us to detect this change and eliminate it (or incorporate itif it is good).Capabilily: A process is capable if it consistently meets design and/or customer specifications (often called folerances). In such a process, variability falls within an acceptable range, usually determined by customers’ needs, Note that a process that is in control (exhibits only common. variation) may still not be capable if its common variation is too great. In formal statistical terms, a process in control exhibits a mean and variation that are consistent with a stable probability distribution for the variable that we are measuring, When a process goes out of control, the implication is that an assignable cause has changed the distribution. Figure 2 depicts distributions of a process variable, as sampled over time, and illustrates several possible situations. FIGURE 2 Process Variable Distributions over Time £21 Spec A 2A shows a process that is in controthat is, its probability distribution doesn't change over time (there is no assignable cause that changes it). This process is also capable: product specification limits fall within the upper and lower tails of the distribution (almost always). 28 shows a process that is also in contro! but is incapable—some measurements fallin the tails of the distribution, outside the specification range.2C. is also in control and incapable; measurements on the high side will be out of spec. 2D is out of control. An assignable cause is changing the distribution over time. It appears to be capable, but beware: Until we find the reason for the shifting distributions, the process is unstablec and could become incapable at any time. 2E is both out of control and incapable. 2.2.1 Control Charts for Variables The fundamental tool in SPC is the control chart, which is used to assess and distinguish sources of variabilty.« This is sometimes described as “the voice of the process.” (Product specifications, often derived from customer specifications, in contrast, are sometimes described as “the voice of the customer.”) Conceptually, the control chart is the visual depiction of quality measurements from samples taken from a process, with multiple samples taken over time. For ‘example, imagine a bottling process in a beer plant. A variable of interest would be the fill evel —the amount of beer in each bottle—for which a narrow set of tolerances would be set. Samples would be taken at regular intervals, say at the beginning of each hour, and recorded In some cases, the distribution of sample measurements might follow a normal distribution (a bell-shaped curve), but this isn’t necessarily the case. For example, the bottling equipment might be more likely to overfill than to underfill the bottle. In this case, the distribution of sample measurements would have a tendency to “lean” to the right, resulting in a skewed rather than symmetric distribution Regardless of the shape of the distribution of sample measurements (the "sampling distribution’), the central limit theorem ensures that the distribution of sample means will be a normal distribution. In other words, for the purposes of SPC, we take a series of samples and calculate the mean of each set of samples. Regardless of how the underlying population is. distributed, the distribution of the sample means will be a normal distribution (see Interactive Mustration 2), This normal distribution, generated by the means of the individual samples, provides the theoretical foundation for SPC and the control chart.@Q irteractive ILLUSTRATION 2 Central Limit Theorem ‘Simulation speed: 2 uniform dition Inimeunm antes, eb) avai beeen sar ae hiv atave tes ig messes t= | In Interactive Illustration 2, first choose a distribution. This will be the distribution of individual quality measuments, as shown on the left. Note that this distribution is not normal, and may be quite asymmetrical. The mean of each set of samples will be calculated and placed in the appropriate “bin’ in the histogram on the right. As more and more sample means are added, the distribution of sample means begins to emerge. Note that, whatever the distribution of an individual sample measurement, the distribution of sample means will assume the familiar bell shape of a normal distribution. ‘An SPC analysis typically relies on two control charts for variables: mean charts show process averages and range charts show process dispersion. Together, they indicate whether a process is in control. They should always be used together, because they provide different information about process stability. The central limit theorem tells us that the mean of the sample means (the mean of the normal distribution above) will be the same as the process mean. The standard deviation of the sample means is directly related to the process standard deviation. The Mean Chart Also known as an X (x-bar) chart, the mean chart captures averages from a series of samples. Itis a sequential (time-ordered) plot of the sample means. For example, for a beer-bottle-filingline, if we measure the average volume of liquid in each bottle in a series of samples, we can chart the sample means, ‘Suppose that at the beginning of every hour, five samples are drawn from this bottle-filing line and the exact volume of liquid in each bottle is measured. Consider 25 of these samples (Table 1). TABLE 1 Sample Measurements from Beer-Filling Line Sample # 10 a) 12 3 14 15 16 7 12.03 12.09 187 12.04 1.99 n.92 12.12 12.10 187 20 12.01 11.98 12.04 11.98 12.10 12.08 11.92 12.05 12.1 12.02 11.93 12.02 1.87 12.09 12.1 11.95 12.04 1.96 11.96 11.93 11.96 12.07 11.93 12.11 11.92 11.94 12.06 11.96 12.00 12.02 1.98 12.03 12.04 12.01 11.87 12.01 11.99 11.90 11.96 12.06 11.96 Sample Measurements (ounces) n91 N99 12.08 1.90 197 12.03 11.96 n99 12.00 11.95 1.97 12.05 12.01 12.05 12.02 1.98 11.97 12.01 12.05 12.00 1.98 12.05 1.98 12.02 198 1.97 12.02 12.03 12.00 12.05 11.99 12.03 12.01 11.99 Mean 11.98 12.04 12.01 11.96 12.01 11.96 12.03 12.04 11.97 12.03 1.97 12.00 12.00 11.98 12.04 12.01 11.99 Range 0.14 0.17 0.21 0.14 0.08 0.16 0.16 0.13 0.17 0.16 0.16 0.09 0.12 0.15 0.14 0.15 0.198 19 20 21 22 23 24 25 12.10 12.08 12.06 12.04 12.02 12.00 11.98 11.96 N94 nz 11.90 Mean 12.07 1209 99 1210 1200 12.05 on 12.05 12.01 11.94 1.98 11.92 11.98 0.13 1.99 12.07 12.05 12,09 12.01 12.04 0.10 12.0 12.04 1.99 1.97 12.04 12.03 0.14 N93 11.99 12.01.95 1.98.97 (0.08 1.97 12.05 12.08 1.99 12.00 12.02 on 12.06 12.04 1205 11.99 11.97 12.02 0.09 1.92 1.99 12.01 11.96 1.93 11.96 0.09 averages 12.00 0.13 The mean of each sample is calculated. Figure 3 plots the sample means. FIGURE 3 Sample Means from Beer-Filling Line 3 5 7 9 N HB BW 7 9 A 2 2s Because the sample means are distributed according to a normal distribution, we can use Sample Number the standard deviation to define upper and lower bounds, which are called control limits. Center lin« Upper control limit (UCL) = ¥4 7. Lower control limit (LCL) SsWhere: (<-double bar) = mean of the sample means standard deviation of the sample means (calculated from historical data or estimated from the sample ranges, as we describe below) lumber of standard deviations The key managerial decision is setting z, the acceptable number of standard deviations. Because the central limit theorem guarantees that the sample means fit a normal distribution, the z-value and percentage between the control limits are known exactly, as shown in Table 2. TABLE 2 Common z Values Target Control Limit (%) z-Value 90.00 1.65 95.00 1.96 99.00 2.58 99°73) 3.00 Itis common practice to use three standard deviations, which means that 99.73% of the sample means wil fall between the UCL and LCL if the process is stable (operating as it normally does). The standard deviation of sample means can be reasonably estimated from historical data if the sample sizes are large enough. When this data is not available, or sample sizes are small, another approach is to use a well-established relationship between the range of a sample (R ‘equals the largest value minus the smallest value in the sample), The random variable W = Rio is called the relative range, where cis the process standard deviation. Its distribution depends on the sample size n, and its mean is denoted d2. An estimator for the standard deviation of the sample means, s, is (aah)R. called R-bar, is the average of the sample ranges. Table 3 gives the values of d2 for sample sizes up to 10. TABLE 3 a2 Values Sample size n dz 1.128 1.693 2.059 2.326 2.534 2.704 2.847 CON AH FWD 2.970 10 3.078 Plugging the estimator for s into the control limit equations above gives us \ 3 avn UCL x( LCL ( ae avn) The averages and ranges (x and R) of each of the samples are calculated, and then these are again averaged. As Table 1 shows, X = 12.00 ounces, and R = 0.13 ounces. For a sample size of 5, d2 = 2.326. If we want 99.73% confidence that process vari caloulate the upper and lower control limits as, ion is common, we Center line = 12.00 ounces (3-0.13) (2.326-¥5) (3-0.13) (2.326- V5) UCL =12 ounces + 2.075 ounces LCL =12 ounces — =11.925 ounces?Adding these control limits to the plot of sample means results in the mean chart shown in Figure 4. FIGURE 4 Mean Chart for Samples from Beer-Filling Line 12.10 12.08 UCL= 12.075 ounces 12.06 12.04 12.02 12.00 11.98 11.96 N94 Mean m.92 LCL= 11.925 ounces 11.90 1 3 5 7 9 NH B Bb 7 9 2 2 2 Sample Number From the construction of the chart and the central limit theorem, we can say that there is a 99.73% probability that, if the process is stable, all the points will fall between the control limits, as they do in the example above. With respect fo sample means, this process is in control. However, if one sample mean (a point on the chart) were to fall outside the control limits, then something would be wrong, This should happen less than once in every three hundred such samples, if the process is continuing to operate as it normally does. However, it may happen more frequently if there are assignable causes. In other words, we can be nearly sure that any sample mean that falls outside the control limits is the result not of common variation but of assignable variation; something in the process has changed, and we should investigate further to determine the cause. Often, we will shut down the process temporarily in order to prevent further defects. In practice, control limits are calculated when a process is stable, and then samples are regularly taken and sample means are plotted on the control chart. When a sample mean falls outside of those previously calculated control limits, we know something in the process has changed and some action needs to be taken to bring the process back into control. The purpose of the mean chart is to evaluate the stability of the sample means. If the sample means are consistent, that suggests the process is in control (Figure 5A). However, if the means are erratic, itis an indicator of assignable variability (Figure 58). FIGURE 5 Mean Charts for Two Processeswv & Variable Consider the mean charts for two different beer-fling lines, line A and line B (Figure 6 FIGURE 6 Mean Charts for Two Beer-Filling Lines Line A Uct= 12.075 ounces Variation due to natural Mean TeL= 11928 ounces 38 7 9 fh &% ® 7 1 mM 2 25 Sample Number Line BVariation due to assignable UcL= 12.078 ounces ‘causes Variation due to natural 1 3 8 7 9 NW &B 8 wv wm a 2 Sample Number assignable How do we interpret these mean charts? Line A is in control, with all the sample means falling between the control limits. However, Line B appears to be out of control, with sample means falling below the lower control limit. This suggests assignable variation—apparently the process has become unstable between the time that sample 23 was taken and the time that ‘sample 24 was taken, This might signal an equipment malfunction or operator fatigue. ‘The mean chart captures the averages of every set of samples. However, itis an incomplete reflection of the process, We are also interested in the Kind of instability in which the mean of the distribution has remained stable but the variability has changed. The range chart provides that information by showing the stability of sample ranges and thereby indicating whether variability within sample measurements is stable. The Range Chart Like standard deviation, the range of a sample is a measure of its dispersion around the mean —that is, a measure of its variability, Traditionally, the range was used rather than the standard deviation because it could easily be calculated by workers, with pencil and paper, on the floor of the factory. For the range chart, we plot the sample ranges and add a center line and upper and lower control limits, calculated as Center line=R UCL(R)=R +3-0, =R+3-d, (R/d,) LCL(R)=R R-3-d,-(R/d,) =R-3-6, where ds is another parameter of the relative range distribution that we use to estimate the standard deviation of the ranges, oR. Like d2, d3 is a function of n. IFLCL(R), as calculated above, is below zero, then LCL(R) is set at zero. This will happen whenever the sample size is six or less. For sample sizes of seven or more, the LCL(R) will begreater than zero, It may seem odd for a range chart to have any LCL other than zero—after, all its sensible to assume that no dispersion within a sample is a good thing. However, itis important to remember that the purpose of control limits is to identify assignable variation. In most situations, some dispersion is expected in each sample. If some samples show too little dispersion, this is certainly an area for further investigation: Was something done differently that reduced the process sample range? Is the testing equipment faulty? Is someone getting pressure to mask the true data? If sample ranges fall unexpectedly, this may be a good thing, but it is incumbent upon quality control managers and employees to determine why. To simplify the equations above, let D, =1+3-(d,/d,) D, =1-3-(d,/d,) Then UCL(R)=R-D, LCL(R)=R-D, Table 4 gives values of D3 and Dé for sample sizes up to ten. D3 has been set to zero whenever the calculated LCL would be negative. TABLE 4 D3 and D4 FactorsR Chart Factors Sample Size Lower Limit Upper Limit n D; Da 2 0.00 3.27 3 0.00 2.57 4 0.00 2.28 5 0.00 21 6 0.00 2.00 7 0.08 1.92 8 0.14 1.86 9 0.18 1.82 10 0.22 178 For a sample size of five, corresponding to the samples drawn on our bottle-filing line (Table 4). Ds = 0, and D4 = 2.11 So, for a range chart, the Center line = 0.13 LeL=0 UCL =0.275 The range chart is shown in Figure 7. FIGURE 7 Range Chart for Samples from Beer-Filling Line0.30 UCL= 0.275 0.25 0.20 Ounces ° 1 3 5 7 9 u 13 15 7 19 2 230 «25 Sample Number Our beer filing line is in control. Both the mean and the standard deviation are stable. The mean chart and range chart must be viewed together, The former reflects the “central tendency’ and shows any changes in the averages across the samples. The latter measures dispersion, or “spread,” within each sample. Neither chart on its own is sufficient to determine if a process is in control. Consider two different processes. Figure 8 demonstrates how the mean chart and the range chart are complementary. FIGURE 8 Stylized Control Charts for Two Different Processes ~ bbbp s pbhP X-Chart X-Chart Does not Detects detect the shift in increase in —-@——-® ~~ central -- © -- dispersion tendency e LcL cL RChart yo, © Chart yey Detects Does not increase in detect dispersion change in mean Let LetBoth processes are out of control, but in different ways. In Figure 8A, the range chart shows increasing inconsistency within each sample of five bottles, even though the sample means are stable. This could be because filing equipment has become more and more inconsistent, overfiling some bottles and underfiling others. In Figure 8B, in contrast, the ranges within the different samples are consistent, but the increasing averages on the mean chart indicate that bottles are more and more overfilled. A useful analogy to the mean and range charts is illustrated by target shooting, Imagine two shooters, each of whom shoots five bullets at a target. Figure 9 shows the results of one such competition. FIGURE 9 Target Shooting Results Shooter A Shooter B &¢ Shooter A shoots (consistently) around the bull’s-eye, but with very litle accuracy. Shooter B is very accurate (al five bullets go to approximately the same spot) but is consistently off-target. Shooter A might not be holding the rifle steady. Shooter B might be very steady, but perhaps the rifle sight is out of alignment, so all the shots are off the bulls-eye. Both shooters are out of control For a process to be in control, both the mean chart and the range chart must be in control. Suspicious PatternsSo far, we have labeled a process “out of control” when sample data points fall outside the ‘control limits. Even if all sample points are within the control limits, certain patterns indicate that the process is out of control. Here are two important patterns: 1. Ten or more consecutive points are either all above or all below the chart’s center line. Given that we know that the sample points follow a normal distribution around the mean, the probability that the next sample point is above or below the mean line is ¥. Thatis, if the process is stable, then the probability that two consecutive sample points are on the same side of the mean line is 4, The probability that three consecutive samples are on the same side of the line is (14)2 = % (¥4 that the second point is on the same side of the line as the first point and % that the third point is also on that side). The probability that ten consecutive points, each drawn independently, are on the same side of the line is (74)9 = 0.002. In words, there is less than a 14% chance of that occurring. Earlier, we labeled a process “out of controf” when a sample mean fell outside the control limits, because the probability of that occurring for a process that has only common (random) variation is very small (0.0027) and, therefore, it's most likely that we're observing assignable variation. Using the same logic, we label a process that has ten or more consecutive points on the same side of the mean line “out of controf” because it is extremely unlikely that such a situation would be due simply to random variation 2. A series of points show an upward or a downward trend. The logic here is the same as above. A series of consecutive points that trend in one direction is very unlikely if the process is subject only to random fluctuations. It's far more likely that something has changed and that the process has gone out of control In general, what we expect from a process in control is randomness in its control charts. If the pattern of plotted points looks systematic in any way, there is reason for suspicion 2.2.2 Interpreting Control Charts SPC charts are drawn using historical data and then monitored over time to ensure that the process stays in control. As the “voice of the process,” control charts tell us if a process is stable. However, when a process goes out of control, they may not tell us why. SPC charts are simply an alarm bell that signals, "Something has happened, and the process is now running differently than it was before.” We need to dig further to understand the root cause of the instability This requires careful observation to detect what has changed or what's different from one time period to the next. Consider the following example. In a factory that runs 24 hours per day (in three 8-hour shifts), a sample is taken in the middle of each shift at a process step in which an operator is running a complex piece ofmachinery, resulting in the X-bar and R charts shown in Figure 10, The sample points outside the control limits indicate that the process is out of control. FIGURE 10 Control Charts for Factory Ae ABCA A A A A AA A AA’ A AAA Range (mm) ABCA A A A A A A A A A A A A A Shift Indeed, the sample means are all over the place. The sample ranges, in contrast, seem in control, showing only common variation. Why is this? The question a manager would ask is, “What changes occur over time as the process is running?" One possible explanation is that the three operators, on the three different shifts, are each running the machinery in a different way. With that hypothesis in mind, we can redraw the control charts, separating the different shifts, and see the pattern shown in Figure 11. FIGURE TI Factory Control Charts, One for Each ShiftX Charts Range Charts ‘Sampio Means ov. ‘Sariple Means 25. uct=2.210 uct=0.62 %=2.060 LeL=1.91 07. 0.0 The operators on all three shifts show remarkable precision (indicated by the fact that variation in all three R charts is very low), but like the target shooter, the three operators seem to be “aiming” differently. The operator on shift A tends to overshoot (with sample means above the center line); the operator on shift C is consistently undershooting; and the operator on shift B seems to do both at random—but, whether undershooting or overshooting, that operator's samples show very ttle variance. While an out-of-control chart doesn't tell us why a process has changed, certain patterns point to potential hypotheses. Indeed, for a process that is in control, the control chart can be used as a leading indicator, flagging the likelihood of future problems before the process is fully out of control (see Figure 12). FIGURE 12 Examples of Control Chart Patterns c E412A This abrupt upswing in X-bar suggests a changed machine setting or a new employee unfamiliar with the specifications. Or perhaps a new material was introduced. 42B This abrupt change in R suggests an untrained employee, a machine needing maintenance, or a new material supplier. 12C This gradual upward drift in X-bar indicates a tool being gradually worn down or a gradual slippage in machine settings 12D This gradual rise in R means greater randomness. It might be the deterioration of a machine, but it can also indicate a decline in operator discipline 12E One might suspect that we have three different operators who set up and operate at different levels. Also, this might indicate a “sticky” setting in the process. 12F This looks like an example of a raw worker who eventually learned the job and approaches normal results. But it could also be the same worker who is distracted and is now ‘coming back into control 2.2.3 Process Capability Control charts provide an indication of process stability. However, as we discussed earlier, even a stable process may not be capable—that is, it may not be able to meet design or customer specifications. It is not terribly useful for a process to operate within its internal controf limits but not meet customers’ specification limits. Obviously, a sustainable business requires processes that are both in control and capable. Two measures are most commonly used to determine if a process is capable: the process capability ratio (Cp) and the process capability inclex (Cpk). The process capability ratio (Cp) is calculated using specification limits and the process standard deviation Upper specification limit (USL) - lower specification limit (LSL) Leo Cc, a where, specification limits (USL and LSL) = limits as defined by design requirements or customer needs; z= number of standard deviations required. Typically, again, z is set to 3; © = standard deviation of the process. This ratio compares the full allowable range (between the upper and lower specification limits) to process variability. Because process results (such as the volume of liquid in a bottle) can fall above or below the mean, the number of standard deviations (z) is multiplied by two. If the process capability ratio is above 1, then the process is capable—in other words, with z = 3,there is at least a 99.73% probability that the process will produce a product that is within ‘customer specification limits. To illustrate, let us return to the bottle-fillng example in which the mean beer volume is 12 ‘ounces. Suppose the firm has set a target that the volume of liquid should never fall below 11.9 ‘ounces or exceed 12.1 ounces. From the sample data we collected (in Table 1), we can calculate that the process has a standard deviation of 0.058. If we want to deliver against these specification limits with 99.73% confidence, we calculate c, ~)2Lounces=11.9 ounces __ 0.2 ounces a 2 The process capability ratio is less than 1, showing that itis not capable of meeting this product specification, Indeed, from the sample data collected, we can see that 11 of the 125, samples collected (8.8%) are defective relative to this specification == 0575 (0.058 ounces 0.348 ounces Managers have two choices when a process is shown to be incapable: Either they (or their customers) must relax their expectations and broaden the specification range, or the common variability in the bottle-filing process must be decreased. For a process in control, decreasing ‘common variation may require management intervention and significant capital investment. For the bottler, it might require more operator training or a new, more precise filing technology. If management is willing to reexamine and relax its specifications, we can calculate that spec range cc, =SPecrangs 1 when spec range > 0.348 ounces. * 0.348 Thus, if management is willing to accept beer bottles that contain between 11.83 (12% - 0.348) ounces and 12.17 (12 + 4 - 0.348) ounces of beer, then the process will be deemed capable. In many circumstances, however, management is not free to relax the specifications, such as in cases where two parts of a product need to fit tightly together. The process capability index (Cpt) is similar to the process capability ratio, but it accounts for the possibility that the process mean is not centered between the specification limits. It measures the distance between the process mean and the nearest specification limit. Cpk is calculated as Upper specification limit-¥ _¥ - Lower specification limit | | ‘4 = Minimum of 30 30 where x = process mean 0 = standard deviation of the process population A process mean may not be the same as the target mean because of the way the process is performing, Alternatively, a process mean may not be centered between the specification limits when the cost of nonconformance is not symmetric, For example, if the cost incurred when the quality metric is higher than the mean is greater than the cost when it is equally lower than themean, then the specification mean will be closer to the lower specification limit. In other words, the process mean could be targeted to be closer to the specification limit with a lower cost of nonconformance, For example, a beer bottling company that sells millions of bottles of beer each year may view slightly underfiling the bottles as more desirable than overfiling them by the same amount, and so it might set specifications in the range 11.85 ounces to 12.1 ounces but seek to ‘maintain a mean of 12 ounces. Suppose that the standard deviation for a new, more precise filing line is 0.03 ounces. The process capability index would be ot (12.Lounces-12 ounces) (12ounces- 11.85 ounces) C, = Minima of HE T—rroororr'es', 3-0.03 ounces 3-0.03 ounces = Minimum of (0.1/0.09, 0,15/0.09) = 1.11, so the process would be capable. Interactive Illustration 3 depicts a probability distribution for a precision part and illustrates the circumstances in which a process is considered capable for a component whose length specifications are 100 mm + 0.5 mm. Q INTERACTIVE ILLUSTRATION 3 Process Capability Index ity ‘oecimn USL szrangy The cto = come = 1 come Come = SUBIR tsa capn G5 Pp ae = come PRODUCTION PROGESS MEAN ‘PRODUCTION PROGESS STANDARD DEVIATION By changing the process mean and standard deviation, you can see how the capability indices change. This illustrates the following:A higher standard deviation leads to a higher defect rate, as reflected by a lower capability ratio and index (Cp and Cok), This is because higher variability creates a wider distribution curve, with more probability in the distribution’s tails; in other words, the probability is higher that output falls outside the specification limits ‘* The capability ratio and index (Cp and Cok) are exactly the same when the distance between the specification limits is equal to six standard deviations, assuming the distribution is symmetric, ‘* Both capability ratio and index (Cp and Cok) are the same when the mean is centered to the process; with a noncentered mean, Cpk must be used to test capability. ‘* With a process mean shifted away from the center, a process may be capable to one specification limit, even if itis not capable overall 2.3 Diagnosing and Resolving Quality Issues As we've said, the tools of statistical process control we've discussed warn us when processes become unstable and cannot meet design or customer specifications, And they can alert us to potential problems—before the process produces defects—if we are alert to certain patterns, To discover why a process has gone out of control, and then to correct the problem, we need to rely on a wide array of problem-solving tools. This section provides a brief overview of some of the most commonly used tools.2 2.3.1 The Five Whys ‘One useful investigative tool is the Five Whys, which was developed by Sakichi Toyoda, the founder of Toyota, and subsequently became an important part of the Toyota Production System. The basic premise is this: Asking why a certain effect occurs will reveal an apparent ‘cause; asking why that cause exists will reveal yet another cause, an underlying cause. Repeating this questioning will ultimately uncover the root cause, helping to ensure that the problem is truly fixed. For example, the problem: My car will not start. 1 Why? The battery is dead.2 Why? The alternator is not functioning 3 Why? The alternator belt has broken. 4 Why? The alternator belt was well beyond its useful service life and had not been replaced. 5 Why? The vehicle was not maintained according to the recommended service schedule (the root cause). A solution to the root cause is replacing the alternator belt and starting to maintain the Vehicle according to the recommended service schedule, Note that just replacing the alternator belt may not be adequate, Without regular maintenance, the belt may break again. Replacing the alternator belt is reacting to the problem; regular maintenance addresses the root cause. ‘The use of the number five is meant to be evocative, not precise, For some problems, the root cause may be revealed with three whys; for others, it could take six. For example, the problem: The stone on the Jefferson Memorial is deteriorating. 1 Why? Itis being cleaned with abrasive cleaners. 2 Why? Abrasive cleaners are required to remove the bird droppings that have accumulated. 3 Why? Bird droppings have accumulated because birds tend to cluster around the Jefferson Memorial, 4 Why? Birds like to eat the midges that go into a mating frenzy at dusk and are attracted by the Jefferson Memorial lights (the root cause), A solution to the root cause is to turn on the lights 30 minutes after dusk, after the midges’ mating frenzy has ended. In many cases, quality problems arise from multiple root causes, so it is useful to identify as, many of these problems as possible and then work to remedy the ones that have the greatest negative impact. Two tools are useful in accomplishing this. 2.3.2 The Cause-and-Effect Diagram The cause-and-effect diagram or Ishikawa diagram (commonly known as a fishbone diagram because of its shape) is a method for brainstorming potential root causes and developing initial hypotheses. As shown in Figure 13, it starts by naming the undesirable effect (the “problem” to be solved), which becomes the “head” of the diagram. Major categories of potential causes become the “spine.”FIGURE 13 Cause-and-Effect Diagram Person Machines Root Cause Root Cause Problem Root Cause Root Cause Root Cause Materials Analogous to The Five Whys, the items on the spine are the answers to the first “why.” These are often grouped into four categories: People, Machines, Materials, and Methods. Possible answers to the next set of ‘whys" are the “bones” protruding from the spines. Managerial judgment is applied to develop hypotheses about the most likely causes, which merit further investigation. The benefits of the cause-and-effect diagram are that it can lead to ‘a comprehensive review of a full range of possible causes and it provides a framework to prioritize the most likely culprits. Most major quality problems have multiple root causes, and organizations do not have ‘enough resources to address all of them. Therefore, it is important to work on those that have the most negative impact. A useful tool for doing this is Pareto analysis. 2.3.3 Pareto Analysis FIGURE 14 Pareto ChartFrequency Causes Pareto analysis ranks problems, issues, or causes from most to least important. The operating concept here is the Pareto rule, typically referred to as the “80-20 rule,” which posits that approximately 80% of a result can be attributed to 20% of the possible causes. If we can determine the 20% of the causes that constitute 80% of the damage, we can better focus our resources on implementing fixes that will have the most impact. There are also a number of useful statistical tools that help us understand relationships ‘among factors or causes. For example, correlation analysis can be used to examine the relationship between two variables. Data points are shown on a scatter plot, and simple regression analysis assesses the strength and nature of the correlation. Note that correlation alone does not prove that two factors are causally related. They may both have a more fundamental root cause. Interactive Illustration 4 shows what data might look like for different values of a correlation coefficient.© interactive ILLUSTRATION 4 Correlation Coefficient “CORRELATION COEFFICIENT For more complex sets of factors, multiple regression can be used to understand which factors (causes) are most strongly correlated with the quality problem in question. Each tool that we've discussed here serves a different function, but they do not need to be used in a predefined sequence or as a group. The objective is to develop a strong hypothesis, validate the hypothesis, and effectively and efficiently push toward a sustainable improvement. This could involve one or all of these tools. 2.3.4 Six Sigma A popular problem-solving approach that uses these and other tools is the Six Sigma system. It was initially developed by Motorola in the 1980s to focus on eliminating defects in the ‘manufacturing of complex systems with many interrelated components. Six Sigma became well known after CEO Jack Welch made it a central focus of his business strategy at General Electric in 1995, and today it is used in a wide variety of industries. In our prior examples, we have assessed control and capability on the basis of three standard deviations on either side of the mean. If quality characteristics follow a normal distribution, and the process mean plus or minus three standard deviations falls within the specification limits, then 99.73% of the output will be good and 0.27% will be defective. That is,of one million parts, 2,700 will be defective. In a complex system in which out-of-tolerance parts interact with one another, this may mean a much larger number of system defects, For example, suppose that if one component in a system with 100 components is defective, then the system will be defective. if each component has a 0.27% defect rate, nearly one- quarter of the 100-component systems will be defective (0.973100 = 0.76; 1 — 0.997310 = 0,24), As its name suggests, Six Sigma sets a much more rigorous target of six standard deviations, which would imply only 3.4 defects per million, when the process mean plus or lus six standard deviations falls within the specification limits.) Achieving this while holding the specification limits constant requires a significant reduction of variability. Interactive Ilustration 5 shows the deciine in defects and improvement in capability as a process moves toward Six Sigma. INTERACTIVE ILLUSTRATION 5 Six Sigma Compliance and Process Capability egamate speci aeardng rato eect) PRODUCTION PROCESS MEAN [ Pocessstamsannevanon | While this was the concept behind Six Sigma, this extremely ambitious quality target is not what's most important about the approach. What really matters is how an organization achieves a high level of quality. As the concept has evolved, so has a data-driven problem- solving methodology, involving many of the tools discussed above, that seeks to improve the quality of process outputs by identifying and removing the causes of defects and by minimizing variability. Quality improvement projects are initiated and led by people in the organization who have been rigorously trained in statistical methods and problem-solving tools and have achieved various levels of competence ("Champions,” “Black Belts,” “Green Belts,” “YellowBelts,” etc.). Each Six Sigma project relies on line workers and may follow a defined sequence of steps or have quantified financial targets. It need not have Six-Sigma reliability as its target. 3 SUPPLEMENTAL READING 3.1.1 Control Charts for Attributes In some applications or for some customers, the measure of quality is binary—either the ‘component or product being sampled is defective or it's not. For example, parachutes used by the U.S. military during World War Il either opened or they did not. Control charts in such situations are generally called “control charts for attributes.” There are two types of control charts for attributes: the p-chart measures the percentage of defective units and is more frequently used; the o-chart measures the count of defective unit. ‘The procedure for the p-chart is similar to that of the mean chart for variables. Each sample is inspected for defective units, and the percentage of defects is calculated Number of defective units Total units inspected where p is calculated for multiple samples, and p (p-bar) is the average of the different sample p’s. As with the variable control charts, the central limit theorem ensures a normal distribution, and we can calculate a mean defect percentage and a standard deviation and set control limits using UCL(p)=5 + Be, } LCL(p)=p-(e, ) where gp = standard deviation of the sampling distribution. op can be estimated as where n, again, is the number of observations in each sample. Because we're plotting percentages, as with range charts, LCL(p) can never be less than zero. The second type of attribute control chart is the c-chart, which counts the number of defective units. Its often useful in situations where there are many opportunities for defects but we can't know about all of them. In such cases, we can't calculate the percentage of opportunities for defects that ended up as actual defects; we can only count the number of defects. For example, a police precinct might develop o-charts for the number of crimes ‘committed at night. They can easily count the number of crimes, but not the number thatweren't committed. Or we might count the number of scratches on a set of microscope slides; we can't know about all the scratches that weren't made. To create a c-chart, the mean number of defects, c-bar, is calculated as Number of defective units Output unit of measure where the output unit of measure could be batches, production runs, days, or weeks, for ‘example. The control limits are then simply, UCL(C)=7+3Ve LCL(C)=t-3VeE 4 KEY TERMS acceptance sampling Inspection of products before, during, or after production. assignable variation (also called special variation or signal variation) Variation in process that is attributable to a change (or instability) in the process, cause-and-effect diagram (also known as an Ishikawa diagram or a fishbone diagram) A diagram used to facilitate root cause analysis, which drills down to assign causes to observable problems, and causes for those causes, ete, Its purpose is to identify the root causes of an identified problem, central limit theorem A foundational theorem in probability and statistics that states that the distribution of the averages (or sums) of sets of samples drawn from any probability distribution converges to the normal distribution. common variation (also called natural variation or noise variation) The random variation that is inherent in any system or process when it is stable. conformance quality The degree to which a product or service meets the specifications of its design and achieves specified performance. control (or in control or stable) A process that exhibits only common variation. If the process exhibits assignable variation, it is out of control. Five Whys An investigation method that assists in uncovering the root cause of a problem. Much like the cause-and-effect diagram, it drills down from an observed problem to its causes and the causes for those causes, etc.mean chart (also known as an ¥ chart) A time-ordered plot of the averages of a series of samples, Pareto analysis A method for ranking the impacts of causes for a problem in a process. process capability Ability of a process to meet design or customer specifications, measured using the process capability ratio or process capability index. process mean Actual average value of a quality variable in a process; it can be estimated from sample means. process standard deviation Actual standard deviation of a quality variable in a process; it can be estimated from sample standard deviations or sample ranges. range charts A time-ordered plot of the difference between the highest and lowest value in a series of samples. sample mean The average value from a sample, defined as Tn for a sample of size n, where the x's are the individual sample measurements. ‘Six Sigma A managerial approach, which uses a variety of statistical and problem-solving tools to solve process problems and reduce defects. tolerances (also called specifications) The acceptable range for a product or process variable, as defined by designer or customer specification 5 ENDNOTES 11 Standard deviation is the square root of the variance in a set of data and is represented by the symbol o (the Greek letter “sigma’). It can be calculated in Excel using the ‘STDEV function, based on the underlying formula, where nis the sample size: 2 For more detail, see Nancy R. Tague, The Quality Toolbox (Milwaukee, Wisconsin: ASQ Quality Press, 1995), p. 15.3 Kaoru Ishikawa, Guide to Quality Control, trans. Gemba no QC Shuho, 2nd rev. ed. (Tokyo: Asian Productivity Organization, 1986). FOOTNOTES a See Core Reading: Managing Quality, HBP No. 8025 (Boston: Harvard Business School Publishing, 2013), b This example assumes that the component's length varies according to a normal (or bell-shaped) distribution. However, our observations about the impact of changes in variability and the shifting of the mean defect rates hold more generally. © We say that a distribution that doesn't change over time is stable, The ones in Figures 2D and 2E are unstable. ¢ In situations where inputs are not expected to be homogeneous, the process is not expected to be repeatable, or where outputs are expected to be unique, the application of SPC may be inappropriate. © Note that this use of the central limit theorem applies for a variable control chart, where the exact ounces or milliliters are captured, as well as for an attribute control chart (see Supplemental Reading: Control Charts for Attributes), where the percentage of bottles meeting specification is measured. { Many textbooks and other references give these formulas as % +4 7.5 +A2R, where 3 ee os 9 Our estimate of s forthe aistibuton of sample means = 73-355 7g) = 0.025 h Today, with laptops and tablets ubiquitous on many factory floors, S charts (which are derived directly from sample standard deviations) are used rather than R charts. S charts require larger samples, and for small samples, the range method yields a good estimate of the sample variances. For the sake of simplicity, we discuss only the R chart in this reading | Demonstrating this mathematically is outside the scope of this reading. | This is the standard deviation of the distribution of the sampled bottles. Note that this is different from the standard deviation of the distribution of sample means (which we calculated to be 0.025). k http:/Awww.itsmsolutions.com/newsletters/DITYvolSiss39.htm| For the 100-component system, 3.4 defects per million components translates to 0.03% system defects