In my corner of the world, I encounter this situation. A student will work on his/her dissertation/thesis and have a set of variables. To address the hypothesis, the student conducts a correlation matrix. I often encounter faculty who state "well, if you have several variables that you are correlating, why not conduct a regression?"

In general, I find that my colleagues response is misplaced. It's as if they are applying a rule that whenever you have multiple variables, do a multiple regression. It's as if the faculty want to see (a) which variables are more important than the others based on the beta coefficients, (b) which variables cancel each other out when they are examined simultaneously, i.e., entered as predictors in regression, or (c) examine the changes from the zero-order correlation to a partial correlation for each variable in the matrix.My colleagues also maintain that we should conduct a partial correlation matrix and observe associations between all the variables while partialing out the covariates, such as demographics. I get the feeling that among my colleagues, it seems more sophisticated to use a regression analysis, or conduct partial correlations, instead of examining a correlation matrix.

My counter arguments against perfunctorily conducting a regression analysis whenever you have more than two variables are the following: (a) you have to have a theory and a research question that dictates which variables are the predictors or outcome variables, and which ones are presumed to have stronger predictive power than the other variables, (b) this view of "why not conduct the regression and see what happens?" means that the results are not meant to address any established hypothesis and research question, (c) running these analyses starts to border on "fishing" and creating more work that does not address the original research question. Conducting a correlation matrix is simpler and if examining the set of associations between the variables is sufficient for answering the research question, then conducting the correlation matrix is sufficient.

if my colleagues are correct in that all associations should be analyzed using multiple regression or partial correlations, then it would seem that there is no need to use a correlation matrix for anything except an initial observation of the associations among variables. I don't agree with this assertion but I am running out of ways to support my view that conducting a correlation matrix, as long as it suits the study's objectives, is just fine. Any ideas on how else I can best support my view?

marked as duplicate by kjetil b halvorsen, Michael Chernick, gung regression Sep 16 '17 at 1:32

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

  • 2
    Welcome to Stats.SE! When you get a chance,please consider taking our tour. In relation to your question: "you have to have a theory and a research question that dictates which variables are the predictors or outcome variables, and which ones are presumed to have stronger predictive power than the other variables" seems to be the whole point of writing a thesis or dissertation. Correlation information without this assumption also seems rather meaningless. There are lots of things which correlate but have absolutely no cause and effect relationship. – Tavrock Jan 26 '17 at 9:56