myspace views counter
Close

Search Waypoint Resources

Search phrase

Filter by resource type

-- Article --

Correlation? Regression? Neither?!?

Posted on May 16, 2014 , by Steve Bernstein

I’m often asked, “For Net Promoter / Customer Feedback key driver analysis which is better, correlation or regression?”  Correlation and regression each tell you different things.  The answer might be “neither is right.”

key driver analysis

Especially in B2B, statistics can confuse and delay action. Improving customer (account) relationships through the front-line account teams can be a straighter path to financial gains.

Correlation shows how 2 variables are related, along with the strength of that relationship. The core output of the correlation analysis (aka Pearson’s) is the correlation coefficient (“r”), which will tell you how strong the relationship is.  The values will range from a value of 1 (there’s a perfect relationship between the 2 variables) to -1 (there’s a perfect inverse relationship – as one goes up the other goes down in the exact opposite direction – something we don’t usually see in customer feedback work).
Statisticians will also say that by squaring the result you can determine how much of the relationship is explained.  For example, if you find an r value of .5 for the relationship between overall recommend and “ease of doing business” then stats folks might say that 25% (r2) of the variation is related to these 2 variables.  That is, as ease of business changes, or as overall recommend scores change, 25% of that change is related to the other.
There are issues with correlation, including:

  • Correlation assumes that the differences between the numbers are the same.  In survey work that might not be valid.  Is the difference between a score of 0 and 1 the same as a difference between a 7 and 8?  There may be degrees of strength that can come into play, making the relationships non-linear.
  • There’s no causality – we only know that the 2 variables are related, and we don’t know which is dependent on the other.  But in NPS work we generally assume that the Recommend score is the dependent variable (it depends on other factors) and that there is a one-way relationship.

Regression is used to predict the impact that a variable has on another, and creates an equation that shows the best fit of a line between the variables.  Say you want to know what the Recommend score is likely to be if the “Ease of Business” score improves from a 5 to 6 – the regression equation could predict the answer.  Generally (in NPS work) the most meaningful part of this is the slope of the line, or the beta coefficient.  If the line has a steep slope then a small change in an attribute’s score could have a large impact on the Recommend score.
Regression can also handle multiple variables at once, which is important since there are usually many factors that need to be combined in order to improve the Recommend score.
All this depends on assumptions.  As with correlation analysis, straight linear regression assumes a relationship. But in NPS-land that may not be accurate and logistic regression could prove beneficial.  You’ll also want to be careful in modeling your data –often more art than science!  You’ll need to deal with missing values (skipped questions or “don’t know” responses), be sure the data is in a normal distribution, define standard error values, understand your r-squared and adjusted r-squared values, etc.
Critical B2B consideration, where the front-line account teams have great influence on the customer’s experience:  much of this experiential work may be handled more accurately by the account team, not statistics.  The front-line can best understand, set, and manage customer expectations appropriately.  There are often far too many variables to try to get this accurately defined by aggregate statistical approaches.  And since go-to-market strategy and execution should be driving your analysis (since the customer experience is generally different for different go-to-market tactics), you’ll have an even bigger problem when the data get too thin to conduct an accurate analysis for a segment.
IMHO, aligning customer expectations with delivery is fundamentally what “Net Promoter” is all about, and companies have to decide if they want to invest in strengthening customer relationships or if things are working fine as-is.  In all my years of doing this kind of work I’ve rarely found the silver bullet “fix.”  Since B2B purchases are high-consideration, high-research, and a generally via a group of people with different expectations, can alignment happen without effective customer dialogs?
I hope more people will contribute here.  Statistics and B2B-NPS often collide!


Interested in more analysis guidance?

Check out our 10 Commandments of Voice of Customer.

10 best practices for deployment and analysis of customer feedback in B2B.

commandments scroll transparent