Hello,

I would definitely use regression instead of correlation. Regression gives you more info: slope, intercept, and you also predicting dependent variable with an independent predictor(s). 

You may build some research questions and use regression to test them. In a nutshell regression is a natural extension of correlation. 

Regards, Witold

Strictly speaking it doesn't matter if your model is of the form y=a + bx +eps and the scatterplot is roughly linear. For multiple regression, it won't work. R = +/-sqrt(R^2) is not a simple correlation coefficient anymore. Witold's comments apply in this case. Best wishes.

Thank you Witold and David, I wanted to be more clear on this issue. lets' assume we have four independent variables as X1, X2, X3, X4 ( X4 is categorical),  and Y as dependent variable. You're willing to test  some hypotheses if there is any positive or negative relation between X1, X2, X3, X4 and Y (irrespective to the coefficients values, per se, the type of relation matters to you only). 

In regression analysis you test the relationships between Y and Xi adjusted for the other Xj (partial correlation); Does the results (type of relation and its significance level) are always in line with correlation analysis? If so why?

Thanks.

Regression coefficients are semi-partial coefficients because you partial out other predictor effects on Y only. In the simplest case of a single predictor the correlation between X and Y is equal to R, but this is not generally true. Depending on the pattern of correlations between predictors and between each predictor and Y the coefficients and semi-partial correlations computed from them might be different in magnitude or sign from the simple correlations between X and Y.

Another situation in which the simple correlations and semi-partial correlations will be near-identical is when all X variables are uncorrelated with each other.

Amir, I think you are mixing two things here. The first test for linear regression tests if the linear model is significant or not while the correlation test checks if a particular independent variable is significant or not. Though H0 in first will signify H0 in second test and H0 in all Xs (X2,X2,..X4) will indicate H0 for the first test. 

Thank Subodh, basically if the correlational structure is robust  you can use regression analyses to test if there is a significant relation between dependent and independent variables and the correlation coefficients signs are identical with the regression model coefficients signs.