@Navya Kumar You adopt PLS approach for testing your causal relationship model as PLS approach doesn't require normality assumption. For PLS based modeling, you can use Smart-PLS (one month trial version).

@Navya You can refer to below attached paper for reference regarding use of Smart PLS for non normal data.

Imran Anwar thank you for your response! Much appreciated.

However, I am still looking for some inputs on the use of bootstrapping.

Because am hoping for the possibility of using multiple regression and logistic regression etc. as well.

Navya Kumar -

If you want to do multiple regression, why not leave the data as skewed? At an agency for official statistics, I dealt with a great deal of very skewed establishment survey data for many years. Skewed data in regression are not a problem. It is nice to have nearly Gaussian estimated residuals (the estimated residuals being close to the distributional form of the actual residuals), but the unconditional data can have any distributional form. And if you are looking at linear regression (which could include polynomial linear regression), with continuous data, then you could consider examples referenced in the following:

Invited presentation for mathematical statisticians at the US Energy Information Administration:

https://www.researchgate.net/publication/319914742_Quasi-Cutoff_Sampling_and_the_Classical_Ratio_Estimator_-_Application_to_Establishment_Surveys_for_Official_Statistics_at_the_US_Energy_Information_Administration_-_Historical_Development, 2017, using prediction.

Skewed data may highlight the fact that larger predictions are generally associated with larger variance. Thus we note heteroscedasticity in regression, which is a naturally occurring feature.

The following two projects with updates in reverse chronological order might be of some help:

https://www.researchgate.net/project/Cutoff-and-quasi-cutoff-sampling-with-prediction-for-Official-Statistics

and

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

As far as bootstrapping goes, it may be a good way to handle a problem when it gets messy, or when there is no closed form solution, or even as a check, but I don't think you need it here. Also, transformations should not be used if they damage interpretation, unless really needed. Often people use a log transformation just to try to reduce heteroscedasticity, which may often be either too little or too much adjustment to be satisfactory, when weighted least squares (WLS) regression is what is needed. OLS regression is just a special case of WLS regression, which is often not a good choice.

If you did need bootstrapping and you don't have access to an expert, you'd need to be careful of any software. One needs to know what the software is doing to be certain it is really doing what you want. I can't help you with bootstrapping.

I think you need WLS regression. If you are looking at something else, say count data where you need Poisson regression, that's different.

Cheers - Jim

James R Knaub Thank you so much for your response! A very interesting perspective about letting data remain as is. Will certainly look at WLS.