# Does it matter if I use correlations or regression coefficients to suggest areas to focus on to improve overall customer satisfaction?

I am working with customer satisfaction data where the dependent variable is "Overall satisfaction" and the independent variables are satisfaction with various areas such as customer support, delivery etc.

I want to suggest areas where the company should focus on in order to improve overall satisfaction.

Option 1: I could look at correlations between the 'Overall satisfaction" and the independent variables and suggest that the company focus on the top 3 positive correlations as areas for improvement.

Option 2: I can use a linear regression and suggest that the company should focus on the areas associated with the 3 highest regression coefficients.

Are the two options equivalent? If not, which one is the better approach?

• This seems indirect to me. You (should) want to know what the most dissatisfied customers are dissatisfied about. That's not really a regression or correlation issue. For example, I can be very happy with quality of website, friendliness of staff, promptness of delivery but as mad as anything because the product just doesn't work. Multiply me, say, because that's a widespread problem, but regressions and correlations are not optimal to find us. That's descriptive statistics and basic exploration first: these mad people, what are they mad about? – Nick Cox Oct 25 '13 at 16:31
• I don't think this management speak, which no doubt is (a) correct enough (b) what you have to work with, bears on my point at all. I said descriptive statistics: I wasn't implying that you just look at means, but at the fine structure of who is most dissatisfied. There are other pitfalls too, e.g. high correlations go with high range of responses in practice. More crucially, you are asking which is better of two methods? My reply is that neither method sounds focused on the real problems. – Nick Cox Oct 25 '13 at 16:42
• Even within your limits: Candour compels me to say that your question seems a false choice to me. If you run the regressions, the correlations are there for free; indeed you need scatter plots too. Why commit yourself in advance to a particular criterion? Your context is what defines your problem. Why throw it away? – Nick Cox Oct 25 '13 at 17:17
• Just because one variable correlates more highly with overall satisfaction does not mean that changing it will have the greatest effect on overall satisfaction. I agree with @NickCox. If the problem you have stated is not your real problem, then please withdraw your post and re-post your real problem. If it is your real problem, listen to Nick. – Peter Flom - Reinstate Monica Oct 25 '13 at 17:39
• It means that there is a relationship in the data, but that doesn't mean that changing one thing will change the other. Basically, neither correlation nor regression solves your problem. – Peter Flom - Reinstate Monica Oct 25 '13 at 18:32

If I understand correctly, customers rate the company in various aspects of the transaction, and then, customers again give an overall score. This is the real-world structure. Making an assumption that customers are reasonably rational (i.e. consistent in their opinions), it means that somehow, they, in their minds, construct some sort of "weighted average" in order to go from the partial scores to the overall score.

Then you should use the regression approach, which reflects the above situation. Using partial correlation coefficients does not capture how one reasonably believes that the customers thought and acted when scoring the company.

This regression is in the spirit of "hedonic index regression", if we view "overall satisfaction" as the "price" of the "product" named "transacting with company", and the regressors as "features" of the product (that are provided in different levels for each customer, and hence their variability).

If the rankings are consistently coded (say, a higher number means a higher level of satisfaction for the partial scores and for the overall score), then a higher estimated regression coefficient on a partial score will indicate that this aspect of the transaction "bears more heavily" (has a higher marginal effect) on "overall satisfaction", and so indeed, focusing and improving on the areas with the higher regression coefficients, should yield larger benefits in overall satisfaction.

But also, in order to finally decide on the prioritization, one should also look how the various areas compare in average score. Say "customer support" has a higher regression coefficient than "delivery", but also, "customer support" is on average rated already very high by customers, compared to "delivery". Then the efforts to further improve "customer support" may be more costly and difficult, compared to improving "delivery". So while one unit of increase in customer support satisfaction may yield higher overall satisfaction increase compared to one unit increase in "delivery", this one unit increase may be more costly to achieve in customer satisfaction than in delivery, offsetting partially, or fully, the economic gains from the increase in "overall satisfaction".

Of course this last issue is not a statistical question, but I mentioned it so that any prioritization suggestion based on statistical analysis, at least mentions this aspect that must be taken into account for the final decisions.

• I agree on the thought process part. However, why would just looking at the correlations between "Overall Satisfaction" with the components be not enough? By the way, I do not mean partial correlation but the Pearson's product-moment correlation coefficient. – user31918 Oct 25 '13 at 16:39
• Pearson's coefficient totally ignores the fact that the dependent variable "Y" varies not just because of one explanatory variable, but because of all. Then it will give you the degree of covariance, but not the degree of effect of the explanatory variable on "Y", which is what you are after. – Alecos Papadopoulos Oct 25 '13 at 17:12