STATISTICAL TOOLS FOR PROCESS IMPROVEMENT

Applying DOE to Microwave Popcorn

Design of experiments identifies which factors matter and which ones don’t, as well as helping find optimal settings. BY MARK J. ANDERSON and HANK P. ANDERSON

ook it hot enough, not too long, and experience, it provides a good example range-finding operation, cooking the pop- a little bit off the floor of the oven. for understanding how to apply DOE in corn pouch for 2-5 min. on a high setting CAnd preheating the oven by heating adjusting process industry recipes. until the rate of popping subsides to an a glass of water for 1 min. has no effect. In particular, cooking microwave pop- interval of about one pop every three Don’t even bother. corn demonstrated how DOE helps apply seconds. Those were the conclusions we made the Pareto principle. In other words, it Two unusual instructions caught our from applying the design of experiments helps to identify what Juran calls the vital attention. One involved having the mi- (DOE) technique to the problem of few factors from among the trivial many. crowave bag resting on a microwave-safe preparing microwave popcorn. The study rack at about the center of the chamber, as was conducted at home, using a common Independent variables opposed to resting on the floor of the microwave oven designed for the cons- To begin, a brainstorming session was oven. The second involved the pre-heating umer market. Since the study examined held to identify all possible factors that step. One of these unusual factors did something with which everyone has some could be studied as independent variables. influence the results, the other didn’t. For this study, five major factors were Our study was designed around a two- Table 1: Factors and Levels selected from a broader range of ideas. level factorial model. Some of the factors Factor Low High The five factors were brand, cooking time, were by their nature discrete and binary; (-) (+) microwave oven temperature, preheat others were continuous. All variables had Price Generic Brand time, and tray elevation. only two values. To limit the continuous Time 4 min. 6 min. The idea for the study grew out of the variables, range-finding trials were Power Medium High last two factors. A quick study of micro- conducted to set low and high levels for Preheat No Yes wave popcorn instructions at the local su- each of the experiments. Elevate No Yes permarket showed that all packages pret- During some of the range-finding runs, ty much say the same thing. The instruc- the popcorn was seriously over- tions advise that the consumer perform a

Normal plot Normal plot

DESIGN-EASE Plot DESIGN-EASE Plot Unpopped 99 Taste 99

A: Brand 95 A: Brand 95 B: Time 90 B: Time 90 C: Setting 80 C: Setting 80 D: Preheat 70 D: Preheat 70 E: Elevation 50 E: Elevation 50

30 30 20 C 20 B C 10 10 Normal % probability A Normal % probability BC 5 5 E B 1 1

-0.95 -0.66 -0.37 -0.08 0.21 -2.82 -1.95 -1.09 -0.22 0.64

Effect Effect Figure 1: Analysis of unpopped kernels Figure 2: Analysis of taste

1 PI QUALITY July / August 1993 (revised 2/98 by MJA) z:originals/flyers/popcorn STATISTICAL TOOLS FOR PROCESS IMPROVEMENT cooked. A kitchen filled with smoke, we responses. To estimate pure Table 3: Array of factors and responses found, was a small price to pay for the error, two repeat runs were Std Run A B C D E Bullets Taste education gained. planned. These extra exper- The brand factor was selected based iments were meant to be run at 112-1-1-1-11 1.57.5 on the central intent of the study, to mid-level (coded as zero) of the 2 9 1-1-1-1-1 1.4 8.0 determine if there is a strong correlation time factor, with the other 3 6-11-1-1-1 1.9 9.0 between the quality of the finished prod- factors fixed at low (-) or high 41811-1-11 0.66.5 uct and the price of the package on the (+). However, the runs were not 5 1 -1 -1 1 -1 -1 1.8 7.0 grocery store shelf. The brands tested executed as planned. Also, one 6141-11-11 0.37.5 were selected to contrast a nationally run in the standard array was 7 7 -1 1 1 -1 1 0.2 2.5 distributed big-name brand against a lo- botched and another one was 8 5 1 1 1-1-1 0.9 1.0 cal grocery store (generic) brand of missed. The software accom- microwave popcorn. The national brand modated these accidental vari- 917-1-1-11-11.77.0 was purchased at $1.79 per package, the ations, and they had no impact 10 15 1 -1 -1 1 1 0.8 6.0 generic brand for $1.25 per package. (A on the results. 11 3 -1 1 -1 1 1 0.6 4.5 complete listing of the two-level factors 12 16 1 1 -1 1 -1 0.9 4.0 can be found in Table 1.) Response analysis 13 4 -1 -1 1 1 1 0.6 9.0 Most of these factors should be famil- To measure the effects of the 14 13 1 -1 1 1 -1 1.3 7.5 iar to the reader. The preheating variable 15 NA -1 1 1 1 -1 Missing ------may be unusual to some, so let us explain variable factors in each run, 16 NA 1 1 1 1 1 Missing ------it in more detail. It was in fact a part of three response factors were what initially raised our curiosity. considered. First the unpopped x 2 -1 -1 -1 1 -1 3.2 8.5 The instructions on one package of kernels (bullets) were weighed x 810111 0.14.0 popcorn that we had tried suggested that and the weight recorded. Like- x 11 1 0 1 -1 -1 0.8 5.0 using a preheating step could increase the wise, burnt popcorn was col- x10-1011-11.65.5 yield of the cooking process. If the occur- lected from each sample run and rence of corn that remains unpopped (we weighed. However, this response turned out to be unreliable. outcome because it means we don’t have call these bullets) is high, the instructions to wait an extra minute for the popcorn. suggested, the yield can be increased by The third response - taste -was subjective, but finding people willing to The four remaining factors (brand, operating the oven with a glass of water time, temperature, and elevation) inside for a period of one minute. serve on a judging panel was not difficult in this case. Taste evaluations were significantly affected the bullets (see Our question was, does this preheat- Figure 1). Residual analysis by Design- ing step—which also would raise the recorded using a scale from 1-10, with 10 being high or good. Observed values Ease revealed the possibility that run 2 humidity inside the oven—really help? was an outlier for bullets. This experiment We shall see. ranged from 1.0 to 9.0. Observations from the 18 runs were produced an unusually low amount of A statistically desirable array of com- popcorn, but since no special cause could binations of the low and high levels was then entered in the Design-Ease package. The software calculated the effect each be attributed to this, and it did not greatly built, for a total of 16 runs, half the total affect the findings, it’s included in the number (32) of combinations possible. independent variable and combination of variables had on the responses. results. Such a fractional factorial design is suffi- Figure 2 shows the normal plot of cient to learn all we needed to know about effects for the taste response. It reveals a popping popcorn. In fact, making more What yield told us… The software automatically produced a highly significant interaction between time runs would not add to our know-ledge. It (B) and power (C). The biggest effect is not necessary to run all 32 combinations graph, called the normal plot of effects, that helped isolate the factors that were comes from the time alone, but its impact to study the interactions between factors. depends on the level of power. As the The runs were randomized to protect the key to determining the yield - the percentage of unpopped bullets. Figures 1 interaction plot in Figure 3 shows, when study against lurking variables—such as the time was limited to its low (-) level of changes in the environment—that could and 2 show the main effects and two- factor interactions for the two measurable 4 minutes, the predicted taste responses otherwise confound the study. To simplify were roughly equal, around 7.5. (The the administration of such a study, we responses. The trivial many factors, which had no influence, fall on a straight line points fall within the 95 % confidence used a Design-Ease® software for design “least significant difference” bars of experiments. It handled randomizing near the zero effect level. One of these factors was the pre- displayed by the software.) With time set the samples and the statistical analysis. at its high (+) level of 6 minutes, however, Table 2 shows the standard (“Std”) heating step (D). Preheating thus had no impact the responses. This is an important the taste response varies significantly array for five factors and 16 experiments. depending on other factors It also shows the run order and observed

2 PI QUALITY July / August 1993 (revised 2/98 by MJA) z:originals/flyers/popcorn STATISTICAL TOOLS FOR PROCESS IMPROVEMENT

- in this case, the temperature or setting of The results also suggest that a name although our tests covered only the two the microwave oven. When set on high, brand performs better than a generic one, brands. Clearly, more investigation is enough of the popcorn called for before changing burned to pull the taste one’s brand preference. response value down to As a result of this study, under 2. Set on medium- Interaction Graph however, we were able to high, taste response 9.0 reduce the presence of bullets dropped some-what less or unpopped kernels by 80 to around 6. percent, a significant gain in 7.7 C+ With this infor- C- yield. mation, we feel - that 6.3 In addition to possibly preheating the micro- C- making home movie or sports wave oven is a waste of 5.0 viewing more enjoyable, this time. On the other hand, study was intended as a elevating the pouch in Actual Taste 3.7 learning opportunity for all the oven is a good idea. participants. When talking No matter how powerful 2.3 about process improvement, your home oven is, C+ this is the kind of analysis that cooking microwave 1.0 has to be done to make popcorn at a high setting breakthrough changes. Look- and for a shorter rather B- B+ ing at one factor at a time in a than a longer time traditional approach will not probably produces a Interaction of B:Time and C:Setting work. DOE provides the tools tastier result. to uncover special causes. Figure 3: Interaction of time vs heat on taste ------Follow-up Study Reveals More Secrets for Making Popcorn

Our DOE on microwave popcorn be, we did a follow-up factorial at plus caused a significant increase in bullets, unintentionally turned out to be a des- and minus times around the factory setting perhaps because it absorbs energy. tructive test. The heat and smoke gen- (factor C), and two additional factors: Increasing the time caused a small but erated at the upper limits of time and prechilling the bag (A), and putting it on a significant reduction in bullets. However, power degraded the chamber to a point wind-up carousel (B). We did all the this was counteracted by a reduction in where we decided it might be best to get a combinations plus 4 centerpoints on time taste (see Figure 5). Prechilling did not new machine. We purchased a more for a total of 8 runs. significantly impact either the bullets or powerful and sophisticated microwave We thought the carousel (factor B) the taste, so it’s unnecessary. As a result that included a pre-programmed setting would distribute the microwaves more of this study, we decided to use the fac- for popcorn. Not content to leave things evenly, but as can be seen in Figure 4, it tory setting for popcorn and no carousel.

Normal plot Normal plot

DESIGN-EASE Plot DESIGN-EASE Plot Unpopped 99 Taste 99

A: Prechilled 95 B A: Prechilled 95 B: Carousel 90 B: Carousel 90 C: Time 80 C: Time 80 70 70

50 50

30 30 20 20 10

Normal % probability 10 Normal % probability 5 C 5 B

1 1

-1.70 -0.73 0.24 1.21 2.19 -1.84 -1.04 -0.24 0.57 1.37

Effect Effect Figure 4: Follow-up DOE – Effects on Bullets Figure 5: Follow-up DOE – Effects on Taste

3 PI QUALITY July / August 1993 (revised 2/98 by MJA) z:originals/flyers/popcorn