Hello Uzma Zargham. I think you need to provide more info. Here are some questions to get things started.

Thanks for clarifying.

I'm sure Bruce Weaver could have continued his list (but he is supposed to be on holiday today), so I will add a few:

6. Do you have some theory to suggest a straight line would describe the relationships?

7. You say best predictor? How many predictors are you choosing among? How associated are they? Are you interested in interactions among these?

8. You have not defined what you mean as best, so at present it is not sensible to answer your question.

I'm supposed to be on holiday too, so I'll let someone continue.

Thanks for those additions, Daniel Wright, and happy Labo(u)r Day to you. ;-)

@Uzma Why don't you look for Smart-PLS? It doesn't have the assumption of normality. Moreover, you can also confirm the results of Smart-PLS with Predictive Analysis in SPSS Modeler or Artificial Neural Networks Analysis in SPSS.

Uzma Zargham let Theory and Evidence help you decide on statistical analysis options of your study. Data Normality is a rather complex assumption to follow considering "Sampling Distribution" and "Central Limit Theorem" from pure statistical point of view. Furthermore, bootstrapping procedure in many statistical packages can handle data normality issue well. Tq.

find attached materials under

Uzma -

It sounds like you have continuous data, and I suggest that you start with scatterplots. You may learn a great deal by first looking at scatterplots of one variable against another, even one 'independent' variable versus another 'independent' variable.

When you have candidate models, you can compare them using a "graphical residual analysis," which also may indicate heteroscedasticity, which is often a feature, not a bug. This would be for a given sample, so a "cross-validation" is often useful to avoid fitting too closely to a particular sample, and not fit so well for the population or subpopulation as a whole to which your model is to be applied. (However, you may discover that you need more than one model and need to split a population or subpopulation.) You could do a graphical residual analysis comparison of the candidate models on a second sample in the population or subpopulation of interest.

Heteroscedasticity, by the way, is often ignored when it should be modeled. You might check the following project and its updates - which are in reverse chronological order:

https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression

Cheers - Jim