I am supposed to conduct a regression analysis on my multivariate data. But my data is not normally distributed, so I was considering doing a principal component regression. Therefore, I would like to confirm: what are the assumptions of principal component regression?

Is normality one of them?

  • 4
    $\begingroup$ The comments and answer to your previous (apparently related) question stats.stackexchange.com/questions/248168/… appear to address this, but to be perfectly clear no form of regression assumes normality of all the variables. $\endgroup$
    – whuber
    Commented Nov 28, 2016 at 19:01
  • $\begingroup$ To add to what whuber said, for simple linear regression the assumed model is $Y=aX+b+\epsilon$ where $\epsilon$ is a normally distributed noise and X can have any distribution. $\endgroup$
    – Hugh
    Commented Nov 28, 2016 at 19:34
  • $\begingroup$ My answer here stats.stackexchange.com/questions/169664 is very much related. $\endgroup$
    – amoeba
    Commented Nov 28, 2016 at 19:37
  • 1
    $\begingroup$ @Hugh, normality assumption is unnecessary even in regression models, except for the ability to use the standard critical values for $t$ and $F$ tests in small samples. For example, OLS estimator is consistent and efficient regardless of normality, and $t$ and $F$ tests work well in asymptotics regardless of normality as well. This has been discussed many times before. $\endgroup$ Commented Nov 28, 2016 at 19:55
  • $\begingroup$ Thanks @RichardHardy that hadn't occurred to me for applying a t test. But I thought that the F test was sensitive to non-normal residuals, if residuals are non-normal does it also work well asymptotically? $\endgroup$
    – Hugh
    Commented Nov 28, 2016 at 20:05


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.