Linear regression - violation of assumptions
Hello. If one encounters a non normal or heteroscedastic data in multiple linear regression, what alternatives can one do in JASP? Are there any nonparametric or modern robust alternatives using point and click interface?
Thank you in advanced.
Kind regards.
Ivan
Comments
which alternatives did you have in mind, specifically?
EJ
Classically, you'd consider a generalized linear model with the appropriate family and link, depending on the distribution associated with your non-normally and heteroskedasticity. You could also, in the event of heteroskedasticity, use the bootstrapping function in the linear regression module to resample your data (may I recommend 10000 iterations and making a cup of tea while you wait?) to "fix" your p-values/standard errors/confidence intervals. If you wanted to go the data transformation route for heteroskedasticity, you could also transform the offending predictor using a log of square root transformation using the compute column function.
This all depends on the severity of the non-normality and heteroskedasticity, though. The linear regression is fairly robust to non-normality, so in many cases you can simply bootstrap in the event you have mild to moderate non-normality and heteroskedasticity and still end up with a decent model.
Mr. EJ,
I didn't have any specific alternative in mind. I was just wondering what were my options :-)
Mr. Dmartin427,
I didn't know that bootstrap could be used in events of mild to moderate heteroskedasticity. I thought it was only used for non-normality to get robust p-values/confidence intervals. Thank you very much for detailed descriptions :-)
Best regards
Ivan