Implicit Methods of Driver Analysis

In our previous post, we examined explicit methods of driver analysis, in which respondents rate the importance of brand or category drivers in an overt and rational manner.

New to driver analysis? In short, driver analysis involves researching how people choose one brand/option in a category over another, by measuring the relative importance of brand/option attributes (drivers) in making choice. Learn more in our full introduction to driver analysis.

These methods assume that when we make decisions, it is conscious and deliberate, slower, and requires effort. If that was how we always make decisions, then stated methods would probably do a pretty good job of measuring, categorising and sizing different driver attributes.

The reality, though, is that we don’t always make decisions this way. Our brains and minds are inherently lazy and will take the path of least resistance to reduce cognitive load by using heuristics and cognitive biases (shortcuts) to get to where we need to be (literally – think about how we drive on “autopilot”).

The same is true for many category-based brand decisions. We have our favourite and not-so-favourite brands which can be based on things other than rational choice – such as how the brands make us feel about ourselves (image) when we use or are seen using them. Thus, to really get to the bottom of how choosers choose, we must also find ways to measure the more implicit or less spoken motivators.

Implicit methods include questioning techniques that effectively tap into and bring to light the more automatic or heuristic and passive ways that we make decisions about things, sometimes referred to as System 1 thinking (see Daniel Kahneman’s Thinking, Fast and Slow). Let’s have a look at these.

Derived importance

One of our favourite techniques to measure the implicit drivers in how choosers choose is the derived importance method.

Derived importance links the appeal or propensity (likelihood to consider or purchase) of brands with features or attributes associated with that brand. It doesn’t care which brands are associated with which attributes, just that if a brand (any brand) is associated with an attribute (X), then it has an average associated appeal or propensity score of Y. Put another way, this is the associated level of appeal for brands with a specific attribute (X) – thus a derived metric. This can be done for multiple features using a range of brands in a category and involves asking the following two questions:

A derived measure of importance is constructed using both of these questions to understand less articulated motivations.

So, in the fictitious example above, makes of cars that are associated with the attribute “low emissions” (Tesla, Honda and Toyota) have a corresponding associated propensity score (likelihood to consider) of 6.0 out of 7 (calculated from the mean of those makes’ scores in Q1), making this an important derived driver.

This score can be compared to a stated score to understand whether this attribute is overstated or understated, revealing even more about the type of driver it is, and the realm of driver classification rather than hierarchy – more on this later. Similarly, we can work out the associated or derived scores for features such as “safety rating” (3.5) and “extended warranty period” (3.3).

Of course, we need to be careful how we interpret these scores as it may be that only car makes that are struggling to get share in the market offer something like an extended warranty, so even though it appears to be undesirable on a derived metric, it might not be and just happens to be associated with poor performing car makes at a certain time in the market – again, the example here is completely fictitious.

Regression analysis

Another popular and really useful derived method of importance is regression analysis. Regression is a modelling tool traditionally used for prediction by measuring the relationship between a set of independent variables (drivers or influencers) on a dependent variable (the thing you are trying to understand). In layman’s terms, this means sorting out which and to what extent various factors have an impact on the things of interest. Staying with our car example, it would mean, “which of the features have the most impact on a car being chosen?”

To create the regression model, you will need to have all the data for all of the dependent and independent variables. Regression works better when these are non-binary metrics (e.g. instead of a binary yes/no choice for the importance of “has an extended warranty”, the choices would span extended warranty options of 1 year, 2 years, 3 years etc.), allowing for more variation in the data to do the prediction.

Once you have the model, you don’t need to have the dependent variable – you can predict it as an outcome. So with our car example, you might use car make features (independent variables) and past sales data (the dependent variable) to create your regression model and then use the model to predict future sales just by having an inventory of the features of stock you do have.

The fact that regression can do predictions means that it can, by default, also do driver analysis because those things which have a predictive value on an outcome can be viewed as drivers of that outcome.

The downside of regression is that this is all predicated on regression finding a model that can explain the dependent variable, and this isn’t always the case. Without getting too far under the bonnet, regression will give you an estimate of how much of the variation in the dependent variable it can explain (adjusted R-squared). Anything over 0.6, which represents 60% of the variation, is considered good – although you need to then also accept that the model doesn’t explain 40% of the variation and therefore other unknown things are driving the outcome.

Regardless, let’s give a positive example. We can use this method to understand the drivers of Net Promoter Scores (NPS) – a metric which is simply a likelihood to recommend a make of car that you have owned to someone else – effectively a proxy for propensity.

Let’s say we collected NPSs for a range of car manufacturers (we’ll take the previously mentioned makes).

1. Tesla
2. Volkswagen
3. Ford
4. Hyundai
5. Honda
6. Toyota

At the same time, we also measure each of the makes’ performance on a set of similar but modified attributes (below) using a 5-point Likert scale of “poor”, “fair”, “good”, “very good”, “excellent”.

1. Safety rating
2. Range of colours
3. Warranty period
4. Servicing costs
5. Styling
6. Emissions

We now have data for both the dependent variable of NPS (the thing we are interested in), and some independent variables (the performance scores on each of the driver attributes) and can run a regression analysis independent of car maker (we could also do it by car maker if desired), so it will be a driver analysis for the category of those cars only.

Let’s say that regression analysis returns an adjusted R-squared value of 0.675. We could conclude that the model including those 6 attributes can explain 67.5% of the variation in the dependent variable – i.e. the NPS score. This would be considered a pretty good model. It doesn’t explain everything, but it does explain the majority of the variation.

The model would then report an importance or contribution weight for each of the attributes (called beta co-efficients) that can then be used to size the importance of each attribute in the model and therefore how important each driver is to the dependent variable. Some packages will express this as a percentage to make the interpretation easier to understand, so let’s say in this case:

NPS = Safety rating (23%) + Emissions (15%) + Warranty period (12%) + Servicing costs (8%) + Range of colours (6%) + Styling (4%) + Other unknown (32%)

Note: It could be the case that some of the attributes don’t make it into the model because they don’t impact the dependent variable, which is likely. For this example, we have assumed that they all have an impact.

From this analysis, we would conclude that “safety rating” is the most important driver of NPS in motor vehicles, followed at a distance by “emissions” and “warranty period” (noting that, again, this is a fictitious example).

Regression is a useful method when you have data for both the drivers (independent variables) and the thing of interest (dependent variable) and should be considered as an effective method of driver analysis when that is the case, and you can achieve a good model. Like all things, there are no guarantees and there is a need to experiment with different inputs and model variations.

Choice modelling – MaxDiff

MaxDiff (maximum difference scaling) – also known as best-worse scaling – is another popular way of undertaking driver analysis by creating a hierarchy of preferences relating to a specific choice such as buying a car. MaxDiff does this by forcing respondents to pick the most and least important item from a list of attributes, helping to identify which they truly value.

How it’s done

It is a discrete choice model where survey respondents are shown a set of the possible items and are asked to indicate the single most and least important items. MaxDiff is an antidote to stated importance scales discussed previously; respondents find these very easy but they can deliver results which indicate that everything is “very important”, making the data not especially actionable. MaxDiff, on the other hand, forces respondents to make choices between options, whilst still delivering rankings showing the relative importance of the items being rated. Here are some additional details of the MaxDiff approach.

• The technique can accommodate larger numbers of potential attributes and put them into a choice scenario where respondents are asked to choose the most and least preferred/important attributes from a smaller list e.g. five.
• Depending on the number of attributes, each respondent is allocated to one of X groups of attributes (designs) and will see Y different attribute combinations (scenarios). This way, respondents do not have to make all possible comparisons, which would be unwieldy and take too long.
• The analysis infers comparisons that are not directly made and distinguishes the most important using a point of parity.

Here is an example.

Results are then expressed in terms of an importance index where “1” equals the point of parity. The point of parity represents the number of times the attribute would have been chosen as the most important if all things were equal. Attributes with scores over 1, therefore, can be declared as more important with attributes displayed in order of importance magnitude. The chart below is just for demonstrative purposes and is not based on real data.

MaxDiff is a fantastic way to create a hierarchy of drivers with a known probability-based importance weight attached to each, which can be used to make decisions about what to invest your product and marketing dollars in. We love this approach and highly recommend it as a very reliable and valid driver analysis approach.

Click here to read case studies involving our use of the MaxDiff approach.

In our next and final blog post on driver analysis, we’ll look at hybrid approaches which use a combination of stated (explicit) and derived (implicit) methods to classify drivers rather than create hierarchies. Stay tuned!

For a detailed guide to the methods used in driver analysis – and the pros and cons of each – download our FREE printable Driver Analysis Guide by completing the form below!