Estimating the potential for UMTS related service

Estimating the potential for a UMTS related service. A researcher's Dre@m or nightm@re

by Marco Hoogerbrugge (Director planning, methods and techniques, SKIM Analytical), and Dirk Huisman (Chairman SKIM Group)

The third generation mobile communication will be 200 times faster than GSM. The increased bandwidth enables mobile communication options and services that were not possible before. As a consequence the new technology enables Networks, Service providers, and "device manufacturers" to create new services which will change the rules on the strategic marketing game board.

At the moment the markets are in turmoil and billions of dollars are spent to buy or create a link with as many individuals possible. The decision making process to spend these billions of dollars is based on the capitalised value of the future spending of the newly hooked-up individuals. What the individual will get in return for his/her spending or what he/she will spend her money on is not known yet but of vital importance for the return on investment. When the dust surrounding the mergers and acquisitions has settled down, the focus will be on anticipating specific behaviour of the millions of individuals. Insight in the drivers of behaviour of these individuals will be the basis to develop, position and price new services.

However, since these new services will be evaluated and selected by the individuals from a new and changing frame of reference, there is no guarantee that the capitalised value of the future spending will ever be realised. But there is no way back. New services will have to be introduced and it is the marketing intelligence's job to estimate the potential and to identify how the revenues can be realised and maximised.

In order to estimate the potential, scenario analysis, technological and demographic trend analysis and macro-economic analysis are essential starting points. With regard to the UMTS related services all these analyses show that the pie might be big, but which slice of the big pie will be ours?

The potential for a new UMTS related service or a new UMTS related product is the sum total of the number of users/choosers * the usage volume of each user/chooser * price. To estimate the usage volume of a totally new service is difficult when historic data regarding the usage pattern over time are missing. By way of analysing usage patterns of comparable behaviour the usage volume may be assessed, but in this presentation we will not discuss the methods to do so. We shall focus on the estimation of the number of users / choosers and on the triggers that make each potential user / chooser choose.

To estimate and realise the potential one has to anticipate the specific choices that millions of individuals will make. To anticipate we might ask them what they will do (=stated behaviour), but to build our estimation on stated future behaviour only, we can just as well throw the dice. Anticipating specific behaviour, it is necessary to understand the decision process and decision rules of the individual as well as to understand the drivers behind those decision rules. To do so we have to create hypothetical choices that mimic reality and observe the choices made.


Figure 1. Understanding and anticipating the choices made

This seems nothing new, as it reflects the core activity of marketing research: providing insight into the consumer and into the market. To provide the insight and to estimate the potential for new UMTS related services and products we are facing the following problems:

The changes in the telecommunication related markets are for a significant part technology driven. The speed of technological progress and of the technology driven changes is in conflict with the speed of adoption. As a consequence we have to face heterogeneity in the adoption and diffusion process. For instance, some users are heavily involved in WAP and anticipate the UMTS solutions; others have never heard of SMS.

The e-business revolution results in a value chain distortion. Infomediaries are a new development, they use the same techniques as market researchers have done, but they apply the techniques the other way around. The infomediaries advise the consumer which product fits his requirements or needs best rather than that the consumers find information and choose themselves.

An overload of variables or elements plays a role in the decision making or in the choice process. To provide insight in the choice of each individual we have to know the impact of each variable for that specific individual and not for the average individual.

People choose and react context specific and people often switch contexts. Besides, people react life-stage specific and the speed of adoption and learning, differ per life stage.

Estimating the potential of a new UMTS related product or service and coping with these problems is a dream because it regards life and changes in life. It is dynamic. It is a challenge. And it is an excellent opportunity to provide true insight into the consumer and the market.
At the same time it may turn out to be a nightmare because the positions on the strategic game board are changing in an unforeseen direction. The rules of the game change, but not as expected. The speed of adoption is over or under-estimated; an important variable was not included in the model; etc.

In this presentation we want to focus on the following problems one has to overcome when estimating the potential for a new UMTS related service or product:

One to one marketing (one of the new rules of the game) means that we have to anticipate and understand each individual, because in selling and providing the new service we have to tune to the individual. But the number of aspects or features (variables) that might play a role and determine choice, is big. How can we identify the sensitivity of each individual to each variable or element that plays a role. Or how can we make his fingerprint (DNA profile) while there is not much time to do and each individual is different?

To anticipate and forecast the choice behaviour one should observe the choice behaviour. Discrete choice techniques provide the solution to do so. But when a growing number of features play a role the choices become too complex and we don't know why the potential buyer made the choice he did. The traditional technique to cope with this problem is Adaptive Conjoint Analysis and that is the technique used amongst others by the more advanced intermediaries. But in the past we learned that the underlying preference model often overestimates the importance of features that are of lesser importance. That is why discrete choice models are used. So we have to combine both techniques.

Potential buyers are choosing products based on the features, price and image elements but that is not what they are looking for. They are looking for the best combination of solutions and benefits offered by the new services or products to fulfil their needs. To understand each individual we have to know how he or she links the features to benefits.

So in this presentation we shall focus on the black box of the choice process. To estimate the potential we should address the speed of adoption and the changes in the behaviour and life-stage related sensitivities as well, but for time reasons we will not do so.

The problem of heterogeneity or how to make a fingerprint
Aggregate forecasting leads to an under-estimation of importance of attributes that cause heterogeneity. For example, some consumers may prefer a small mobile phone so that it is easier to carry. Other consumers may prefer a large mobile phone so that it is easy to read the display and easy to touch the keys. These two opposite opinions can of course not be combined to one "average" opinion that the size of the phone does not matter. However, when forecasting choices at the aggregate level the mathematics do lead to such a wrong conclusion.

A practical problem is very often that at the individual level many attributes play a role but relatively few data are available to tell us something about an individual's choice behaviour. It is not uncommon that in a market research some 30 variables are tested in a study while each individual conducts only 10 or 20 choice tasks.

For this problem a variation of Hierarchical Bayes has been developed that copes with choice data. For those of you who have not heard of Hierarchical Bayes before, we will explain this in a nutshell.
In statistical science we have two fundamentally different approaches:

1. classic statistics, which draws conclusions based on observed data only
2. Hierarchical Bayes statistics, which combines observed data with subjective probabilities.

The latter sounds rather magical if you haven't heard of it before. An example will hopefully help to understand. Suppose recently somebody passed his exam to fly aeroplanes so he is an official pilot now. But the second time he flies as a pilot, he has an accident. Fortunately he and the passengers all survive. The statistical question is now: how big is the chance that he will have an accident the third time as well? Classical statistics says: he flew 2 times, got an accident once, so for the third flight the chance of another accident is 50%. Bayesian statistics says: let's talk to his examiner and ask the examiner to make a subjective estimate of this pilot's risk to get an accident, based on his experience during teaching. The final Bayesian chance of the pilot getting an accident in his third flight will be a weighted average of the examiner's subjective probability and the incidence rate of 50%. Because we have so few observations (only 2) the incidence rate will have a low weight and the examiner's opinion will have a high weight.

Now, the Bayesian approach in forecasting choices is a combinatory approach similar as described before:
We consider as subjective probability that an individual makes a certain choice: the aggregate probability that all individuals make that choice.

We consider as objective observations that an individual makes a certain choice: the choices of the individual in the conjoint interview. (That is what classic statistics would use entirely on its own. However, as told before, the problem for classic statistics applied to individual-level data, however, is that in most cases the number of variables to be estimated is far too large, relative to the number of observations.)

The final choice forecast for an individual may then considered to be a weighted average of the subjective probability (= aggregate probability) and the probability based on choices of the individual only. The weight of the individual choices is higher when the number of choice tasks increases and it is also higher when the consistency of the individual's responses is higher.

The question here is: how likely will it be that the individual respondent will choose for a specific UMTS related service? Let us assume there are 12 service attributes and one of these attributes was related to information services (like information from the stock market) that an individual is very sensitive to. In other words, as soon as this attribute is included in the choice task it triggers this individual's response. Then the utility value for this attribute will consequently be very high. But regarding another attribute, the E-mail option, the individual reacted less consistent. Twice he choses a concept with E-mail included and twice he rejects it (50%). But the latter may well have been caused because the E-mail option was accidentally shown in combination with a high price or in combination with "no automatic stock market information". Because the individual information available regarding E-mail is quite little and quite diffuse, we do not really know how sensitive the individual is to the E- mail option and consequently we "borrow" information from the total population. If on average E-mail is chosen in 80% of the tasks the "weight" of the email option for our individual will increase somewhat. Please bare in mind this description is an oversimplification of Hierarchical Bayes.

To summarise:
aggregate forecasts (only) lead to bias
pure individual forecasts are often impossible and always inaccurate, particularly the elements which are not dominating the individual' choices
a Bayesian approach combines the two and leads to very acceptable outcomes (in fact, we have experienced that even in a very complex study the choices of most respondents are explained by Bayesian mathematics for 70% to 90%).

The problem of many attributes, the overload of information
In practice it is not unusual to include a huge amount of attributes in a study that refer to service and cost characteristics of a provider or to product characteristics of a mobile phone. In the context of a full profile approach (all attributes are included in each concept) it is very questionable, though, whether respondents are able to give valid responses to concepts that consist of many attributes. Especially if the combination of these many attributes require that respondents should make complex calculations in order to determine the most profitable option.


Figure 2: Overload of information in complex choice tasks leads to a biased response

In the past this problem has been by-passed by the preference models and the adaptive conjoint system. Adaptive Conjoint Analysis is a partial profile approach, i.e. you do not have all attributes in a trade-off task but instead only two or three at the same time, while the attributes alternate across the trade-off tasks. By the way, this approach is now also used by infomediaries to help the end-users choose a product or service.
However, in the research practice it was generally felt that the importance of attributes was leveled off, in other words it was experienced that the most important attributes were underestimated, while the least important attributes were overestimated. There have been many papers on this subject, which have particularly focused on the underestimation of price in adaptive conjoint.

Since recently there is a similar solution available in Choice-Based Conjoint. It is possible to have partial profiles such that some attributes are always present in a concept while other attributes alternate across choice tasks. The potential buyer chooses from a number of products which are defined on the main characteristics or some key characteristics and the other characteristics are alternated across the choice task, but in a choice the products shown are defined on the same attributes.


Figure 3: Example of choice tasks with alternating attributes

At first sight this approach seems very promising. However, by having both fixed and alternating attributes there is a danger that you step into a vicious circle: you keep some attributes fixed because you think they are the most important attributes. Then from the data it appears that the attributes are the most important – but is this because they are indeed or because they were kept fixed?
Despite the methodological remark regarding the vicious circle we have experienced that the choices are better explained using this approach then following the traditional approaches in which attributes were left out or all attributes were included in the choice task.

The problem behind attribute sensitivity: Looking for benefits
As indicated, we are able to estimate sensitivity to attributes and concepts and we are able to simulate choices people will make. But the offers of UMTS providers or the functions that terminals offer will be quite similar, still x% chooses provider or terminal A and z% chooses B. This difference in choice will only partly be due to a difference in functionality between the providers. To get a clear picture of the drivers of choice we have to look for the benefits behind the attributes and for the needs explaining why the benefits are important.

The technique to analyse why attributes are important is laddering. It is a qualitative and very time-intensive technique in which the potential buyer has to specify why attributes and concepts are important to him/her or in other words what the meaning of the attribute, concept or product is.


Figure 4: The structure of a laddering interview

The result of a laddering survey is a map showing how features are linked to benefits and needs. So we get the picture, but it is an aggregated picture from a qualitative survey. This has to be validated and should enable us to link the individual choices to the ladders (links between concept-benefits-needs).


Figure 5: A map linking attributes to benefits and needs/values

In the Structured Open Association Pattern technique we start with the inventory of all benefits (phrased in consumer language) and needs. From the previous part we know which concepts the potential buyer will choose or which elements are driving his choice. Starting with the chosen concept or with an element driving choice, the potential buyers selects from a list of benefits the one benefit that - for him or her - explains the choice driving element best. This step is followed by the next step in which s/he has to specify why the benefit selected is important or which of the (higher) benefits or needs applies to him. This process is repeated several times. In the end we have a validated map which is aggregate and descriptive. But we have the individual ladders as well so we can explain the individual choices better, enabling us to better anticipate their choices and related benefits sought for.

For instance if we look to the potential buyers who indicate to be sensitive to the WAP functions of a terminal, we can see that the importance of this function is related to different benefits and needs:

being "immediately" or "automatically" informed so I can react directly;
I have an important position;
self esteem;

send an email any time any place;
keep my friends informed;
social responsibility;

#I can find what I look for;
you must be in control;

I can send a message to all my friends at the same time;
they all know immediately;
it is fun;
hedonism;

The difference in benefits correlates with the sensitivity to the terminals. This may explain why different terminals with WAP functions are appealing to different types of individuals and why x% of the WAP seekers choose terminal A and only z% terminal B. For instance, an individual being sensitive to WAP because he is driven by self esteem may be willing to pay more for the most advanced well-designed terminal because he is worth it or to show off.
Consequently, to estimate the potential one should not only take into consideration the impact of the attribute (WAP) on the choice. One should also take into consideration the benefit looked for behind the sensitivity for the attribute.

Benefits and needs change over a longer period of time and as such are partly life- stage specific. At the same time people learn and consequently the benefits which are of importance now for a specific life-stage group may be a non-issue for the same life- stage group in the future. The young singles and couples from today will be the young families tomorrow. The "fun benefits" may then be substitutes by more functional benefits (organising life) and care related benefits but what they take with them is their chat and sms experience and their related expectation of immediate response.

To estimate the potential for new UMTS related services and products, we base ourselves on choices of individuals at a certain moment in time. But we also have to take into consideration the changes over time. On the individual level this means we have to include past experiences and past behaviour and on an aggregated level this means we have to include the concept of moving cross sections in our forecasts.

Summing up
Changes in the value chain urge marketers to estimate the choices of the individual consumers. Understanding and anticipating the individual consumer will become more important - if not of essential importance.

This means that we have to be able to make each individual's fingerprint. To do this we are confronted with the problem of heterogeneity (each individual is different) combined with a large and growing number of elements that may play a role. There are too many variables that have to be measured in too short a period of time. Hierarchical Bayes is a new and innovative statistical technique making use of subjective probabilities, which helps us to cope with this problem and enable us to better forecast the choices of the individuals.

To forecast choices we have to observe and analyse the choices individuals make to identify the elements that determine the choice, as well as to analyse the decision rules of the individuals. Here also we are confronted with too many variables. Combining techniques to alternate attributes over the choice tasks with the discrete choice modelling or choice-based conjoint helps us to cope with this problem. However we have to realise that here we take some subjective decisions up front with the risk of the vicious circle.

The forecasts are based on product characteristics and product concepts. Consumers ask for product concepts and product characteristics, but they are looking for benefits to fulfil certain needs. To better forecast one should aim at integrating the link between product characteristics or product concepts and the benefits looked for. This enables us to understand the individual consumers better and consequently to better anticipate the choices they will make.
To estimate the potential of new UMTS related products and services the described approaches are of great help. However they are not sufficient because we still have to cope with the heterogeneity in speed of adoption and we have to cope with the learning effect or the impact of actual behaviour of future behaviour.