A number of methodological issues in QFD that I have described have been discussed in the literature on QFD before. Arrow’s theorem, for example, has been the subject of discussion since Hazelrigg argued that this theorem “proves that currently popular approaches to design optimisation such as Total Quality Management (TQM) and QFD, are logically inconsistent and can lead to highly erroneous results” (Hazelrigg
1996, p. 161). Various authors have suggested methods for dealing with this fundamental methodological problem (Scott and Antonsson
1999; Lowe and Ridgway
2000; Dym et al.
2002), which have again been criticized (Franssen
2005). Also a number of other methodological problems that I have described have drawn attention in the literature on QFD (Ramaswamy and Ulrich
1992; Wasserman
1993; Matzler and Hinterhuber
1998; Park and Kim
1998; Vairaktarakis
1999; Cook and Wu
2001). In several cases, this has led to proposals for improved or more sophisticated QFD methods.
5.1 Sophisticated QFD approaches
In the literature on QFD, a whole range of more sophisticated approaches has been proposed. Some of these are intended to deal with the methodological problems I have sketched; others mainly aim at a more precise and mathematically sophisticated formulation of QFD. My aim is not to give a complete overview, but only to sketch some current developments and to indicate whether these are promising for eventually overcoming the earlier sketched methodological problems or not.
A first development is the integration of Kano’s model for customer satisfaction into QFD (Matzler and Hinterhuber
1998). Kano’s model makes a distinction between three types of user demands:
-
Must be requirements. If these are not me, customers will be extremely dissatisfied, but these requirements do not positively contribute to perceived customer satisfaction.
-
One-dimensional requirements. Customer satisfaction is supposed to be proportional to the degree to which these requirements are fulfilled.
-
Attractive requirements. These are extra product features. Customers are not dissatisfied if these requirements are not met, but if these requirements are met, the rate of customer satisfaction is disproportional.
This distinction can be seen as an attempt to address the methodological problem that tradeoffs between customer demands are usually not constant (Sect.
3.2). Distinguishing between these three types of requirements helps to avoid this oversimplified assumption. Matzler and Hinterhuber (
1998) propose different indexes for customer satisfaction and customer dissatisfaction for customer demands. They do, however, not offer a method for translating these into priorities among the engineering characteristics or into target values. Still, although Kano’s model does not address the more fundamental methodological problems in QFD, it goes some way in addressing the issues described in Sect.
3.2.
A second development is the use of more sophisticated rating scales for the relation between customer demands and engineering characteristics. Park and Kim (
1998), for example, criticize the choice of rating scales like 1–3–9. They propose to use a cardinal scale instead of an ordinal scale for these ratings, and so try to address the methodological issue discussed in Sect.
4.1. They fail to argue, however, how it would be possible to measure the correspondence between customer demands and engineering characteristics on a cardinal scale and on one that is uniform for all the different correspondences. Another proposal to deal with this methodological problem is sensitivity analysis (e.g. Shen et al.
1999). Again, this does not solve the fundamental problem. Sensitivity analysis applies different rating scales and tests whether this results in different outcomes. In doing so, it presupposes that the (relative) weight of the engineering characteristics can be expressed as a
linear
additive function of the (relative) importance of the customer demands. However, the distorting effect of this assumption might well be larger than the mere choice of the rating scale. For such reasons, the added value of sensitivity analysis is limited.
A third development is more sophisticated methods for target setting (e.g. Fung et al.
1998,
2003, Kim et al.
2000; Tang et al.
2002). Such more sophisticated methods, for example, try not simply to maximize customer satisfaction given a budget constraint, but also introduce additional constraints, for example a minimum degree to which each engineering characteristic has to be met. Some methods also try to differentiate between different types of resources instead as just overall costs. Finally, a number of approaches use fuzzy models to deal with impreciseness and uncertainty. Sophisticated as they approaches might be, they neither address the more fundamental methodological problems that are due to Arrow’s Impossibility Theorem nor the more specific problems with respect to target setting I have sketched. Some in fact seem to increase the methodological problems by making more or stronger assumptions than conventional QFD approaches do.
5.2 Alternative selection procedures
QFD is usually not understood as a method for choosing between different design concepts, but as a method for setting engineering targets. Nevertheless, the outcomes of QFD can be used to choose between different designs. It is therefore interesting to see if alternative selection procedures exist that can help to overcome the methodological problems of QFD. In the literature, a number of approaches has been proposed that claim, sometimes implicitly, to overcome the methodological problems that arise due to Arrow’s Impossibility Theorem.
Franceschini and Rossetto (
1995) have proposed multi criteria analysis, in combination with outranking, as an alternative to the relationship matrix in QFD for setting the relative importance of the engineering characteristics. However, this does not overcome Arrow’s Impossibility Theorem, even if Scott and Antonsson (
1999) claim that multi criteria analysis is not plagued by Arrow’s Impossibility Theorem. Franssen (
2005), however, has shown that their arguments beg the question because they presuppose that an aggregate order among the multiple criteria exists, while that is just what is at stake.
Dym et al. (
2002) have proposed pairwise comparison charts for comparing alternative designs, a method that could also be used to rank design criteria or engineering characteristics by importance. As they show, their approach is equivalent to the Borda count, i.e. it gives the same outcome. The Borda count is known to violate the condition “independence of irrelevant alternatives” of Arrow. Saari has argued that the Borda count is nevertheless superior to others methods of aggregation because it uses all relevant available information (Dym et al.
2002). It is, however, contestable what information exactly is relevant and available (cf. Franssen
2005, p. 48).
5.3 Market segmentation
A third category of alternatives focuses on market segmentation. To understand the importance of market segments, it is useful to look at an example presented by Hazelrigg (
1996) to show how Arrow’s theorem can result in erroneous QFD results. Suppose a product has three attributes: colour, size and shape, and suppose that each attribute has two possible options: red or green (colour), large or small (size) and flat or bumpy (shape). Suppose that there are three groups of customers, whose preferences are represented in Table
2.
Table 2
Preferences of three groups of customers with respect to colour, size and shape
Customer 1 | Red | Large | Flat |
Customer 2 | Red | Small | Bumpy |
Customer 3 | Green | Large | Bumpy |
Collective | Red? | Large? | Bumpy? |
On the basis of this table, one might be tempted to think that the group preference is a red, large, bumpy product. It might be the case, however, that customer 1 dislikes a bumpy product so much that is has no value to him, while customer 2 dislikes large products so much that they have no value to her; for customer 3, finally, red products may have no value at all. What seems to be the most preferred product is actually disliked by all customers.
Lowe and Ridgway (
2000) present two possible solutions to the example presented by Hazelrigg. The first has to do with how the preferences of the three individuals are aggregated. Hazelrigg presupposes a kind of majority voting on each attribute separately. However, we might also ask each of the customers to rate the importance of each attribute on a scale from 0 to 1 and then calculate the weighted average importance of each attribute. Even if this procedure gives a better solution in this particular case, as Lowe and Ridgway argue, it does obviously not solve the fundamental issues that arise due to Arrow’s Impossibility Theorem.
16 In other situations, it might be Lowe and Ridgway’s aggregation method instead of Hazelrigg’s one that leads to “erroneous” results.
The second solution presented by Lowe and Ridgway is to supply not a single product but a number of products, each of which would satisfy some customer group; in this case, this would imply the supply of three different products. In terms of Arrow’s Impossibility Theorem, the choice of a specific customer segment can be seen as way to introduce domain restrictions, so weakening the second condition (unrestricted domain) on which the theorem rests: we only count certain customers as members of a market segment if their preferences meet certain domain restrictions.
It is known that under certain domain restrictions Arrow’s Impossibility Theorem does not apply. One such a restriction is single-peakedness. This condition implies that there is one underlying criterion alongside which all the individuals order the options. An example is the left–right distinction in politics. The idea is that while individuals will prefer different options on the left–right axis, their preferences will fall monotonically to both the left and the right side of their most-preferred option on the left–right axis. A similar restriction does not seem reasonable in QFD, however. If customer preferences are eventually determined by only one criterion, what is the point in distinguishing different customer demands in QFD? This seems to presuppose that there is in fact not one underlying criterion, and there are, I think, good reasons for this presupposition.
Nevertheless, the use of market segments may introduce domain restrictions that even if they do not avoid Arrow’s Impossibility Theorem, at least alleviate the consequences of it. It can be shown, for example, that under reasonable domain restrictions for market segments, QFD results in a Pareto improvement among the customers in that market segment, so avoiding the erroneous results suggested by Hazelrigg.
To show this, I start with supposing that the preferences of each customer in a market segment can be represented by an ordinal value function. This is usually possible if the preferences of each customer over the options form a weak order. Note that in this case, the options are formed by some combinations of the values of the customer demands
s
i
as discussed in Sect.
3.2. In contrast to there, I do not presuppose that the value function takes the form of a linear additive value function. This supposition can be written as: Now also suppose that the value function
v
x
of each customer x in market segment
m has the following two properties: These conditions could be used to define market segments, so that these conditions are by definition met in each market segment. (Note that in different market segments
s
i
,…,
s
m
can be different). In this situation, an improved meeting of the customer demands will result in a Pareto improvement, in the sense that no customer in the market segment is worse off and at least one customer is better off.
Now suppose that it is also possible to find a set of engineering characteristics meeting the following conditions:
1.
For each customer x, a value function v
x
of customer demands s
i
,…,s
m
exist so that \( v_{x} (a_{i} , \ldots ,a_{m} ) \ge v_{x} (b_{i} , \ldots ,b_{m} ) \Leftrightarrow {\mathbf{a}}(a_{i} , \ldots ,a_{m} )\succsim {\mathbf{b}}(b_{i} , \ldots ,b_{m} ) \) where a
i
,…,a
m
and b
i
,…,b
m
are different combinations of values for s
i
,…,s
m
and \({\mathbf{a}} \succsim {\mathbf{b}}\) means that customer x weakly prefers option a over option b.
2.
\( {{dv_{x} }} /{{ds_{i} }} \ge 0 \) for all customer demands s
i
of each customer x.
3.
\( {{dv_{x} }} /{{ds_{i} }} > 0 \) for at least one combination of i and x.
4.
For each customer demand si an ordinal value function γi of the engineering characteristics e
j
,…,e
n
exist so that \( \gamma _{i} (x_{j} ,...x_{n} ) \ge \gamma _{i} (y_{j} ,...,y_{n} ) \Leftrightarrow x(x_{j} ,...x_{n} )\underset{\raise0.3em\hbox{$\smash{\scriptscriptstyle\thicksim}$}}{ \succ } _{i} y(y_{j} ,...,y_{n} ) \) where x
j
,…,x
n
and y
j
,…,y
n
are different combinations of values for e
j
,…,e
n
and \({\mathbf{x}}\,\succsim_{i}\,{\mathbf{y}} \) means that option x meets customer demand s
i
at least as good as option y.
5.
For each customer demand s
i
: \( \frac{{d\gamma _{i} }} {{de_{j} }} \ge 0 \) for all engineering characteristics e
j
.
6.
\( \frac{{d\gamma _{i} }} {{de_{j} }} > 0 \) for at least one combination of i and j.
If conditions 4, 5 and 6 are met, better meeting one of the engineering characteristics—without doing worse on any of the other engineering characteristics—automatically implies a Pareto improvement among the customer demands. If conditions 1, 2 and 3 are also met, this also implies a Pareto improvement for the customers.
Note that in traditional QFD, it is presupposed that γ
i
can be written as
$$ \gamma _{i} = {\sum\limits_{j = 1}^n {a_{{ij}} {\text{ }}e_{j} } }{\text{ with }}a_{{ij}} \ge 0. $$
Under this presupposition, conditions 4, 5 and 6 are indeed met. Note, however, that conditions 4, 5 and 6 are much weaker than what is usually presupposed in QFD. It is, for example, not presupposed that each γ
i
can be written as a linear additive value function of
e
j
, so avoiding a number of the methodological problems discussed in Sect.
4.1. (Note, however, that condition 5 and 6 are a kind of reformulation of the presupposition in QFD that
a
ij
is always non-negative.) It should also be noted that condition 4 is not plagued by Arrow’s Impossibility Theorem. The reason is that condition 4 requires the solution of a single criterion instead of a multiple criteria problem.
17 Condition 4 requires that it is possible to weakly order different combinations of engineering characteristics
e
j
,…,
e
n
with respect to
one specific customer demand
s
i
; it does not require weakly ordering combinations of engineering characteristics
e
j
,…,
e
n
with respect to
combinations of customer demands
s
i
,…,
s
m
. So, we only have one “voter”, i.e. customer demand
i, while Arrow’s Impossibility Theorem only applies to two or more “voters”.
Market segmentation can thus be used to introduce certain domain restrictions. A minimal result that can be achieved by market segmentation is to avoid that some customers are actually less satisfied with the new product than with the current product. It might be possible to introduce even stricter, but still plausible, domain restrictions than those suggested above in defining market segments, so that Arrow’s theorem can be avoided.
18 This is, however, beyond the scope of this article.
5.4 Demand modelling
Another interesting development is demand modelling to predict the demand for products with certain features. On the basis of such predictions, the desirable characteristics of products can be chosen. Some authors have also attempted to introduce such considerations into QFD. I will discuss at some length a proposal developed by Cook and Wu (
2001).
Cook and Wu use the so-called S-model to predict user demand. This is a phenomenological model for predicting demand, expressed in terms of the values and prices of products. The demand is taken to be equal to the total amount of a product sold over a period of time, assuming that there is no scarcity of supply. The value of a product is a measure for the amount of money people are willing to spend on that product. The assumption is that people buy a product if its value is higher than its price. If demands and prices are known for a range of competing products, the value of these products can be calculated.
Using the S-model, predictions of future demand can be made if prices and values of a new product are known. The difficult part, of course, is to predict the value of a new product. Cook and Wu propose the direct value (DV) method to this end. In this method, customers are asked to compare a baseline product with an imaginary alternative product in which one or more of the values of the product attributes have been modified. The customers are asked to choose between the baseline and the alternative product for a series of prices of the alternative product. Next, the fraction of respondents choosing the alternative is plotting against the price of the alternative. On basis of this plot, a so-called neutral price
P
N is determined; this is the price at which half of the respondents chooses the alternative product and half the baseline product. On basis of the S-model, it can now be shown that:
$$ V - V_{{\text{0}}} {\text{ = }}P_{{\text{N}}} - P_{{\text{0}}} $$
In this formula V
0 is the value of the baseline product and P
0 is the price of the baseline product. V is the value of the new product. If V
0 and P
0 are known, and P
N has been determined, V can easily be calculated.
The DV method is usually used for one attribute change at the time. Note that this supposes that the attributes are preferentially independent: the change in value due to changes in one attribute does not depend on the values of the other attributes. As discussed earlier, this might be a problematic assumption (Sect.
3.2). The DV method also presupposes that people can compare non-existing products with current ones, which might be problematic (cf. Sect.
3.1).
Another methodological issue with respect to the DV method is that, from earlier research, it is known that there is a gap between the maximum price for which someone is willing to buy a product and the minimum price for which that person is willing to sell it. Most people want a higher minimum price for selling a product than they are prepared to pay for buying the same product. Usually this phenomenon is phrased in terms of Willingness to Pay (WTP) versus Willingness to Accept (WTA). Cook and Wu interpret this phenomenon in terms of uncertainty. Even if this would be a right interpretation, an implication seems to be that the DV method will probably yield different values for the same product if different baselines are chosen. In general, it might seems reasonable to use the current product as baseline; however, in reality consumers will not choose between the current product and the new product but between a number of—new—products of competitors and the newly developed product.
Cook and Wu integrate the S-model in QFD in order to increase the profit of the company. Their proposed QFD approach proceeds as follows. Customer demands are listed and related to engineering characteristics. The engineering characteristics for the current (baseline) product are measured. A range of alternative products is devised with other engineering characteristics. With the DV method, the changes in customer value for these alternative products are measured. Using the S-model, the expected additional demand given a certain price for the alternatives is calculated. By also estimating the expected additional costs for developing and producing each alternative, the added profit for each alternative can easily be calculated. The option with the highest additional profits is chosen.
Some people would probably argue that the approach proposed by Cook and Wu is no longer a QFD approach. It does, for example, not set target values for the engineering characteristics. Although, they use a kind of House of Quality, they do not make any of the calculations mentioned in Table
1. However, by not making these calculations they avoid most of the methodological problems that were discussed in Sect.
3.3 and in Sect.
4. This is not to say that their approach is completely without methodological problems; some of these have been indicated above.
What is perhaps most important is that Cook and Wu’s approach avoids Arrow’s Impossibility Theorem because it does not try to aggregate individual preferences into collective ones, but just models how many people would probably buy a product with certain features. This seems more generally true for demand modelling approaches. Wassenaar et al. (
2005), for example, write about their demand modelling approach that it “aggregates the customer choices (not preferences) by summing the choice probabilities across individual decision makers (customers), thus avoiding the paradox associated with aggregating the utility or preference of a group of customers” (Wassenaar et al.
2005, p. 522).