Continuous analysis

A finer-grained analysis of survey results

The standard Kano analysis is discrete: features are categorized in separate categories. There's something to say for the fact that a certain nuance is lost:

[Many people] typically feel that it seems like a loss of information (and in some cases is misleading) to reduce 25 possible combinations of answers from each respond[e]nt into just one of six possible categories, and then to reduce further the category derived for each respondent to just one of six possible categories for all respondents. (Berger et al., 1993)

Apart from adding more nuance to the numbers, continuous analysis also means you can make your analysis more visual and telling by rendering results in a graph. Berger 1993 contains two methods for continuous analysis that can help you better communicate the weight of a feature's category and ease decision making.

Better-worse analysis

Berger (1993) cites the engineer Mike Timko, who calculated the value of the presence of a feature and the cost of its absence, both expressed as a decimal between 0 and 1. He did so because he

wanted to calculate an average of some sort while also preserving some idea of the spread over Attractive, One-dimensional and Must-be features. This gave [him] the idea to reduce the data to two numbers: a positive number that is the relative value of meeting this customer requirement (…) and a negative number that is the relative cost of not meeting this customer requirement

Timko reduces the categories to a “cost of absence” (i.e. how much pain are we causing by not adding or improving a feature) and a “value of presence” (i.e. how satisfied will people be if we do add or improve the feature). He basically turns the M>O>A>I rule into a number.

The value of a feature's presence is calculated by adding the number of answers that put a feature in the Attractive or the One-Dimensional category and dividing it by the sum of the Attractive, One-Dimensional, Must-Be and Indifferent answers:

Value of presence = $A+O \over A+O+M+I$

The cost of a feature's absence is calculated by adding the One-Dimensional and Must-Be answers and dividing them by the sum of the Attractive, One-Dimensional, Must-Be and Indifferent answers. (To make it more obvious that this is a cost, add a minus sign in front of it).

Cost of absence = - $O + M \over A + O + M + I$

Let's look at an example where we've calculated three features' categories:

Feature	M	O	A	I
A	12	18	5	2
B	8	4	22	3
C	3	5	12	17

Calculating the value and cost turns this table into:

Feature	Value of presence	Cost of absence
A	0.62	-0.81
B	0.70	-0.32
C	0.46	-0.22

This can then be easily visualised in a graph, rendering a feature's category in a finer resolution. The graph is more clear if you remove the minus sign from the Cost of Absence totals.

You can now easily see how "pure" a feature's category is. For example, a purely Attractive feature's coordinates would be 1 on the Y-axis and 0 on the X-axis. In a survey with 20 surveyees where every answer categorizes the feature as Attractive, the calculation would be (20 + 0) / (20 + 0 + 0) = 1.0 for Value of Presence and (0 + 0) / (20 + 0 + 0) == 0.0 for Cost of Absence.

Another way of showing the same data (and one that makes it easier to prioritise features) is with a bar graph:

Reverse features are not part of this calculation. You should re-examine Reverse features to develop a positive counterpart.

Dumouchel's asymmetrical scale

The same article (Berger, 1993) contains a section by William Dumouchel who developed another way of increasing the resolution of survey results.

His calculation is based on the answers themselves, not on the resulting category. Dumouchel assigns a score to each functional and dysfunctional answer. The scale is asymmetrical as "Must-Be and One-Dimensional are stronger responses than Reverse or Questionable" (Berger, 1993):

	Functional	Dysfunctional
Dislike	-2	4
Live With	-1	2
Neutral	0	0
Must-Be	2	-1
Like	4	-2

The canvas where answers are plotted on then looks like this. (The graph is copied from Berger 1993, where they label Natural answers as Must-Be, which I don't agree with).

The answers are plotted on the graph by averaging them. So, if the answers of 20 surveyees for a particular feature are as follows:

Feature A	Functional	Dysfunctional
Dislike	2	9
Live With	3	8
Neutral	0	1
Natural	3	2
Like	12	0

then the averages would be:

• Functional: [(2 x -2) + (3 x -1) + (0 x 0) + (3 x 2) + (12 x 4)] / 20 = 2.34

• Dysfunctional: [(9 x 4) + (8 x 2) + (1 x 0) + (2 x -1) + (0 x -2)] / 20 = 2.5

This would categorize the feature as a weak One-Dimensional feature.

You can of course do this for all features, and for a three-feature survey the result could look like this:

You can even go a bit further, and include error bars that communicate the standard deviation in the answers per feature:

A warning about Dumouchel's continuous analysis

Notice how in the above example feature B and C seem close to each other. But if we look at the actual numbers, there is a big difference:

Feature B	Functional answer	Dysfunctional answer
Dislike	12	2
Live With	7	0
Neutral	1	1
Must-Be	0	2
Like	0	0

Feature C	Functional answer	Dysfunctional answer
Dislike	2	1
Live With	5	7
Neutral	10	12
Must-Be	3	0
Like	0	0

This is why I'm a bit wary about going too far with continuous analysis (especially on an asymmetrical scale like Dumouchel's).

The answers are not supposed to be turned into numbers (because they are not on a scale). The Kano method is a qualitative method, and I've seen researchers use continuous analyses to give the results of their surveys an air of complete scientific validity. It's as if some people can't cope with the certain vagueness that surrounds qualitative methods.

Treat the Kano model for what it is: a categorization of qualitative responses, not as a scientific measurement tool. To illustrate my point, have a look at this graph of qualitative categories:

This graph means most people's favourite colour is orange and a lot of people prefer green. It wouldn't make sense to treat the responses as a scale and calculate their mean. You'd arrive at the conclusion that people generally prefer light brown (which is what you get when you mix orange and green). The same goes for the Kano model.