Solved – Whether to use Spearman’s rho or multiple regression to examine relationship between two Likert scales

likertregressionspearman-rho

I am doing my thesis and have absolutely no previous experience in statistics.

I have constructed several Likert scales by forming composites scores each based on 4-6 items which tests the level of agreement of my respondents.

Specifically, I have created two scales 'purchase behaviour' and 'website appeal'. I want to see whether 'website appeal' is correlated with 'purchase behavior'.

Now that I am doing my analysis, I am confused as to whether I should use Spearman rho or multiple regression.
The Spearman correlation and multiple regression have different p-values, so much so that one states that I need to reject my hypothesis and another accept.

So in this case should I use Spearman's rho or multiple regression?

Is there a theoretical rule that I must use, say, multiple regression because I am testing 4-6 items on the likert, although I have grouped them together and intend to 'view' them as two single variables.

Thanks, Jeromy. The article by Gelman and Stern (2006) is really interesting! Being the non-statistician me and trying to get on with my MBA thesis, I would be very tempted to find an analysis which gives me a simple method to analyse my data and ultimately, test my hypotheses. I know this shouldn't be the way, but stats aren't exactly fun nor interesting. I was talking to my supervisor and he suggested using regression when I had planned to use Spearman (cos Likert scale items are ordinal and if I want to test ordinal vs ordinal, I use Spearman – according to my research methods text)

Yes, I am planning to test only two variables (predictors?) at a time, so, Spearman can technically be used. But these two variables (both dependent and independent) are computed as a new variable from the different items I have on the Likert scale (does this make sense?). I am just concerned that my analysis would be deemed incorrect if I used the 'wrong' statistical analysis – or doesn't this matter?

Best Answer

  • The difference between significant and non-significant is not necessarily significant. In general, you shouldn't look at the p-value to decide which test to use (for article length treatment see Gelman and Stern (2006)).
  • If you are concerned with developing a predictive model, then regression would be suitable and correlation would not be. If you are concerned with summarising bivariate association, then some form of correlation coefficient would be suitable.
  • The p-value of a linear regression with a single predictor will be the same as the p-value for Pearson's correlation. Given that you have only one predictor, then this applies to you. Thus, your question could be rephrased in terms of whether to use Pearson or Spearman's correlation coefficient. This has been discussed elsewhere on this site such as here.
  • For typical Likert scales there is a limit on the degree to which extreme outliers can occur. As such Pearson and Spearman often give similar results. That said, a little movement in p-values is not surprising, and if this just happens to cross the magical .05 line then it may appear more substantial than it really is.
Related Question