**Which Features Are Important And Which Not So Much?**

When we develop a product or a service, we face many trade-offs. We have to decide which characteristics we will focus on and which we will put aside. Typically, we can not afford giving full attention to all of the features. Also, we have to select the features that we are going to communicate as trademarks or differentiating points of our product or service. The ultimate goal is to improve the overall impression we leave on our potential and current customers. If you want to know which features are the most related to the overall impression (e.g., likeability or purchase intention), what you need is **driver analysis**. Besides helping you identify the important features, driver analysis can also help you approximate the magnitude of change in the overall impression if the performance on a particular feature is changed. Pretty powerful, right?

**Key metrics:** What do I need?

**Likeability:**„*Overall, how do you like this product/service? Express your evaluation on a scale from 1 to 5, where 1 is "Dislike it extremely", and 5 is "Like it extremely ".***Purchase intention:**„*How likely would you be to buy this product/service? Express your evaluation on a scale from 1 to 5, where 1 is "Definitely would not buy "and 5 is "Definitely would buy ".***Features evaluation:***„How do you evaluate the following features of this product? – Taste. Express your evaluation on a scale from 1 to 5, where one is "Dislike it extremely, "and five is "Like it extremely ".*

**The main concepts: **What should I know?

Behind the scenes, driver analysis is actually linear regression. Regression is a procedure of estimating a size of a relationship between one outcome variable and one or more predictor variables. In driver analysis, the outcome variable is typically some measure of overall impression (e.g., likeability or purchase intention), while the predictor variables are usually evaluations of individual features. There are many forms of regression. One of the most commonly used is ordinary least squares multiple linear regression. If you are not familiar with these terms, they might sound intimidating, don’t give up yet! By the end of the text, you will feel confident explaining them to your clients.

**What is the ordinary least squares multiple linear regression?**

Let’s start with "ordinary least squares". Imagine that, in a market research study, we evaluated the taste and purchase intention of a product and that we plotted our results, as in *Picture 1*. Purchase intention is on the Y-axis, evaluation of taste is on the X-axis, and individual dots are our respondents.

Focus your attention on the dot **A**. It represents a response of a person A who disliked the taste but still wants to purchase the product (maybe they would buy it for a family member). On the other hand, dot **B **is a person who gave the lowest mark for both the taste and the purchase intention. That makes sense - B dislikes the product and doesn't want to buy it. Person **C **really likes the product but is not completely sure about the purchase (maybe the price is too high). To summarize all those individual stories, we plot a regression line, also visible in *Picture 1*. A regression line is an approximation of the relationship between outcome (purchase intention) and predictor (taste evaluation). How did we make the regression line? The goal was to find the line that is the best possible descriptor of the relationship of the measures – that has the smallest distance to the actual data points. We determine the overall distance from our line to the actual data by summing the distances (shown in vertical dotted lines) of the individual points from the line. But, before we sum them, we need to square them first, simply in order to avoid positive and negative values cancelling out. By doing that, we are trying to find the position of the line that will produce the smallest sum of the squared distances. And that’s why we call the method “least squares”.

Next, why do we call the regression "linear"? Simply because we used a straight line to model the relationship between the variables. Maybe a curved line (e.g., S-shaped) would have done a better job. But for the sake of simplicity, we commonly stay with a straight line.

Finally, why do we call the regression "multiple"? In the simplified example shown in Picture 1 we see the relationship of the purchase intention with only one feature, taste, in practice, we typically want to examine more than one feature, in which case we call the analysis multiple regression.

**The interpretation of the outcome: **How should we read it?

We said that we want to plot a regression line through the center of our data to summarize the relationship between the predictor variables and outcome variable. While this is true, the ultimate goal is to making predictions. We want to know how the purchase intention score would change if we changed the taste likeability.

Let's think again about our regression line. The slope of the regression line describes the amount of the change in Y (purchase intention) for the given change in X (taste evaluation). The coefficient that defines this relationship is called *β* (beta), and those are the values we are looking at in order to interpret results of the driver analysis. They tell us about the amount of change we would observe in the outcome variable if we increased the predictor variable for one measurement point (in our example, that would be one point on a 5pt scale that we used for taste evaluation), and all the other predictors were being held constant. *β* (beta) can also take negative values, which simply means that an increase in predictor’s value would be associated with a decrease in the value of the outcome variable.

**Key take-aways:** How can I use it?

The information about the *β* (beta) score for any particular predictor variable is the most useful when we compare it with the scores of other predictor variables, because this comparison allows us to assess the relative importance of various features. In our example, we can compare *β* (beta) scores for taste with those of price affordability, packaging likeability, etc. This information can guide our choices when deciding which attribute to prioritize when developing or communicating our product or services.

**Further considerations**

Sometimes, things can get more complicated. As we already mentioned, for some attributes we could image a non-linear relationship of a product feature and its likeability. If we considered „spiciness“ of a food product, or in the case of many product and services - their price, we could image that the likeability would grow with the increase in these attributes (some spiciness is appealing, and people are usually weary of choosing the cheapest option), but that, at some point, further increasing these attributes would lead to a decline in likeability (it would start to become too spicy, or too expensive).

It’s always important to visually inspect the relationship of the predictors and the outcomes, and pay attention to the model diagnostics in order to determine whether the linear model is a good fit (model diagnostic tools are available in every statistical software offering regression analysis).

## Comments