Conjoint Analysis is a way of assessing why people make the choices they make and of being able to predict what they will do if given an set of choices. Discrete Choice Modelling, DCM, is an alternative approach, to Conjoint type problems.

The table below shows a typical situation where Conjoint Analysis or DCM might be used.

A typical Conjoint problem

A manufacturer can make a laptop with the following options • 2 sizes of screen • 3 sizes of hard disk size • 2 different memory sizes • 5 prices to test Which product(s) should the manufacturer make, and at what price(s)?

This laptop manufacturer can make 2 * 3 * 2, 12 different combinations, and if they are combined with the 5 price points there are 60 options to test.
60 options is too many to test by dividing all the options into different concepts and testing each one. Conjoint Analysis allows the manufacturer work out the value of each attribute (for example size of screen versus size of hard disk), and each level of each attribute (for example 2MB of memory versus 4MB).

How Are the Values Calculated?

Conjoint Analysis recognises that people cannot accurately say how much each factor determines their final choice. Therefore, Conjoint Analysis values are created by giving people different sets of choices and asking them to make decisions.

The choices the respondent has to make are designed so that no one option is good in every way. The respondent has to trade-off something they want, in order to get something even better.

For example if I offer you a great meal at a really high price, or an OK meal at a lower price, there are two options available to you. If you pick the great meal, you have traded away the price; i.e. you are a less price sensitive person. But, if you pick the cheaper meal, you will have traded away the quality of the meal, and you are more price sensitive.

In a Conjoint study the respondent is presented with many choices and by processing these choices we can understand the value that a respondent places on each level of each attribute.

Once we know the attribute values for each individual we can conduct a wide range of analyse, including needs based segmentation, product optimisation, and what-if modelling.

What-if Modelling

Most market research only provides answers to the question that were specifically asked during the research. What-if modelling allows marketers to pose a wide range of questions and to see how customers would have reacted.

Choosing the right technique for the job

There are a wide variety different ways of doing Conjoint Analysis and Discrete Choice Analysis. At Virtual Surveys we have concentrated on two approaches, namely bespoke Discrete Choice Models and Adaptive Conjoint Analysis.

The two examples below show the relative advantages of these two techniques.

Example 1 – Adaptive Conjoint Analysis

Skiing Holiday

Attributes: This study had a large number of attributes, including flight time, transfer time, type of accommodation, type of accommodation, number of black runs, red runs, blue runs, snow board park, après ski, etc.

Because there were a large number of attributes, and different attributes were of different importance to different people, ACA was used.
In the first stage of the ACA the respondents put into order those attributes that did not have a ‘natural’ order.

For example, they told use whether they preferred self-cater, chalet, B&B, or hotel. But we did not ask them to tell us whether they preferred the holiday to be cheaper, or for the transfer time to be shorter – we can infer which of these is best.

The second stage of the interview asks the respondent to say how import the best and worst levels for each attribute.

For example, if two resorts were equal in every other way, how important would it be that one resort had a snowboard park and the other didn’t?
The first two sections of the ACA interview give us an idea of what is important to that respondent. This allows the software to create choices for the next stage that will ask questions that are relevant to each particular respondent.

The third and main stage of the interview creates 12 to 20 paired choices and asks the respondent how much they prefer one option to the other.
For example the screen below shows two options, each showing levels from five attributes.

Option A	Option B
French Alps Self-Cater 5 Black runs Quite night life £750 per person	Italian Dolomites Chalet 30 Black runs Great bars and disco £800 per person
Strongly prefer A Strongly Prefer B 1-----------2-----------3-----------4-----------5-----------6----------7-----------8-----------9

The computer creates a variety of choice tasks, using different attributes on different quesitions.

Option A	Option B
2 hour transfer Self-Cater English speaking ski school Ski right to accommodation £800 per person	4 hour transfer Chalet No ski school Bus to and from ski area £500 per person
Strongly prefer A Strongly Prefer B 1-----------2-----------3-----------4-----------5-----------6----------7-----------8-----------9

The results of the study were used to create a needs based segmentation, to identify how many types of skier there are. Segments varied from people who were skiing-focused, 8-to-an-apartment, looking to combine a low budget with extreme hills, to people for whom four afternoons of skiing, combined with good food and shopping was the ideal.

Example 2 – Discrete Choice Modelling

Petrol Station

In this study the number of attributes could be constrained to just four, these were: brand, price, convenient location, and shop (including co-location with superstore).

The study was designed to evaluate the future impact of petrol stations co-located with superstores. The client was aware that there were planning applications for many more petrol stations. A secondary aim of the research was to evaluate strategies for competing with the co-located petrol stations.

The study was conducted as a DCM project. The respondents were shown a set of tasks with 5 cards per task, set out as below:

Brand 1 £1.20 a litre On your daily route No shop

Brand 2 £1.25 a litre 15 minutes out of your way With a superstore

Brand 3 £1.15 a litre 5 minutes out of your way With a small shop

Brand 4 £1.30 a litre 30 minutes out of your way With a mini-supermarket

None of these

Compared with ACA, the DCM tasks are very simple, the respondent simply has to make a choice. This process is very close to the customers behaviour in the real market. People don’t rate choices in stores, they select items from a set of choices.

DSM also introduces the idea of None of These, which often creates a better context for questions.


Each respondent saw 16 of these screens, and made 16 choices. The data were processed using Hierarchical Bayes (see next sub-heading).
What-if modelling was used to assess:
• What would happen if co-location became more convenient (which is what happens when more stores open)?
• What price differential can the branded stations maintain over the unbranded options?
• What level of convenience is “convenient enough”?

Hierarchical Bayes

Hierarchical Bayes, or HB as it is more often referred to, is a powerful statistical technique to deal with sparse data sets. HB can be used for a number of purposes, but the most common use in market research is in the calculation of Conjoint and DCM utilities.

In many DCM studies, the number of choice tasks we ask each respondent is not sufficient to work at respondent level values for all the attributes and levels. HB allows us to use a small number of choices per respondent, typically in the range 12 to 20. It does this by using a large amount of computing power to estimate what respondents would have done if they had answered a larger number of choices.

What sort of products are Conjoint and DCM used for?

This section illustrates a few of the case where ACA or DCM has proved useful.

Laptop

Studies researching laptops are very suitable for both Conjoint and DCM approaches. Typically, studies look at: brand, size, weight, price, disk size, memory.

Care needs to be taken about interactions in the area of brand and styling. An Apple laptop that is very innovative may be seen as attractive, but a Dell with the same form may be seen as simply odd. DCM can be used to create different attribute sub-sets for different brands. For example a Dell or HP might be shown with XP or Vista operating systems, whereas Apple would be shown as having the Mac operating system.

Washing machine

White goods research typically looks at: brand, price, capacity, design*, functions, eg spin speed.

* Attributes like style and design are not always easy to include in a choice study. In a washing machine study one might include 4 photos of machines, showing them as basic, classic well built, stylish, and modernistic. The photos would not be branded, and would be sufficiently vague so as not to invalidate issues such as functionality.

Car

Within the area of car research, there are many areas that have been explored by choice modelling approaches. For example, some studies look at the drivers of choice (doors, brand, engine size, etc.). Other studies focus on details such as just in car entertainment or purchase options.

Hotel

Hospitality studies can be conducted with different types of groups, for example business travellers or holiday makers. These different samples produce very different questionnaires.

A common mistake in this type of study is to use attributes that have levels which are too precise for respondents to really appreciate. For example, check-in takes less than 2 minutes, 2-4 minutes, 5-7 minutes etc.

Another potential trouble-spot with hotel studies is to try to define non-tangible attributes such as atmosphere or style. Attributes such as traditional and minimalist convey some impression of the hotel, but only to a limited extent. Phrase such as boutique hotel are very fashionable, but if you find out 10% of the respondents are highly motivated towards boutique hotels, you still have no idea if they meant the same thing as each other, or how to achieve it.

Travel Insurance

Insurance in general is a popular field for choice modelling, and travel insurance is a good example. Attributes in travel insurance studies tend to include: annual premium, activities covered, whether family also covered, excesses, and items covered.

One issue with travel insurance is to be clear about when the respondent makes their typical purchase decision. Do they have an annual contract, in which case they only rarely make a decision? Do they buy it with their holiday purchase; if so this will also inhibit the chance of a preferred scheme being bought. Do they buy it at the airport? Or, do they look around, before going on holiday, and make a considered purchase. This last group are the most likely to be approximated by the choice or conjoint model.

FMCG products

Because Conjoint tends to deal with cases where there are tangible attributes, it is less often applied to the majority fast moving consumer goods research, such as soft drinks, detergents, or fish fingers. That is because most of the research in these areas concentrates on emotional scales, brand personalities, communications and advertising.

When conjoint and choice modelling is used with FMCG products it tends to be issues such as packaging (glass versus plastic versus can, size, with or without handle), elements of multi-packs, and pricing.

When to use Conjoint, and when not

Conjoint Analysis and DCM work best when the attributes can be clearly described and where attraction of the whole is a least approximated by the attraction of the parts. For example, you may like Tea and you may like vodka, but there is no reason to suspect you will like tea flavoured vodka. This is a case of where the attributes interact and where conjoint would not normally be helpful.

Similarly, attributes such as fun, happy, attentive staff, tend to be difficult to work with because respondents do not agree with each other about the definition. In a choice study we ideally want the respondents to agree about what the attributes mean, and then for them to have different importances for those attributes.

A simple way to assess whether the attributes are suitable for conjoint is to think of a matrix of attributes and products. How likely is it that most consumers would agree about which attributes apply to which products. The more they agree, the more likely it is that conjoint is appropriate.
For example, if we were conducting an instant coffee study, the attributes such as container (pouch, can, or jar), size (200g, 400g), with or without caffeine, brand, and price are all very clear.

By contrast attributes such as bitter, modern, strong, dark, and happy are less consistently applied, and therefore less suitable for choice modelling.
Choosing Between Conjoint and DCM

In general DCM is a more powerful solution than Conjoint Analysis. If a problem can be fitted to its limitations, then DCM should be chosen. The main limitation of DCM relates to the number of attributes it can handle. Most marketing scientists prefer to keep the number of attributes to seven or fewer.
In all but the most simple of cases, DCM is usually dependent on processing the data with Hierarchical Bayes.

If the number of attributes is greater than seven, and if the number really can’t be reduced, then Adaptive Conjoint Analysis is the most appropriate technique.

One limitation of the Adaptive Conjoint Analysis method is that it can only be administered via a computer, usually via online surveys, but CAPI (i.e. using laptops). DCM can be conducted via a computer, or via paper and pencil.