Heteroscedasticity - Correction

[@MrHugosky1]

Hello Professor Pat Obi. I hope you are very well. I have the following question: If I am working with an ARDL, having a serious case of heteroskedasticity, would you recommend to try the transformations you mentioned in this video, in order to remove it? Thank you very much in advance for your help.

[@PatObi]

Yes.

[@MrHugosky1]

Professor @@PatObi Thank you very much in deed 😊!. By the way, is there any online course or specialization you recommend to me to gain more knowledge about multivariate models?

[@ResilientFighter]

Great channel thanks for all the videos. I am rather curious though, could quantitie regression be another mediator for heteroscedasticity?

[@PatObi]

I'm not familiar with that. But I suppose if it helps reduce heteroscedasticy, you'd have more efficient coefficient estimates.

[@Guilopes99]

Great video once again! I have this issue: I want to select the best 10 out of 100 X variables to predict Y. To have the best model, would I have to perform all the possible combinations of variable sets and chose the best model based on F test and adjusted R-square? How could I lookout for heteroscedasticity in such loopings? Also, how do you do the quadratic model if you had all these independent variables? Would you include another panel of the same variables but squared to this looping selection + model test?

[@PatObi]

I don't recommend that. You could use a stepwise regression to decide on the most parsimonious model. Look up my video on Stepwise.

[@Guilopes99]

@@PatObi hey, thanks for your quick response. I've seen the video, so I would need to run the model 95 times, to keep the best 5 variables, taking one out of the time? Isn't there a better way to select the best variables?

[@PatObi]

@@Guilopes99 Better to use a software, not spreadsheet. But you really shouldn't work with so many variables.

[@Guilopes99]

I would ask the question in other way: to find the best possible (more effective and less biased) model, what's the workflow you recommend? My current one is, based on your videos and prior (not much ahah) knowledge: - input missing values for all variables (mean, for example) - deal with outliers, case by case, which usually I input max or min, ir eliminate the observation if it's too extreme - check pairwise correlations between variables and eliminate highly correlated variables - with the remaining values, run models forever with the possible combinations of variable sets - select the best model based on adj r sq and f test - evaluate heteroscedasticity and transform the Y variable to ln(Y) if needed Is this the recommended workflow? And I'm missing steps?

[@raymeester7883]

I just found your channel. I am amazed and grateful for the work you have done. Please could you reach out to other KZbinrs to shares your ideas and concepts. It will help promote what you do and contribute to the knowledge of people. KZbinrs such as Oshay Duke Jackson, Jordan B. Peterson, Fort Bend Tutoring and Yvette Carnell. It will open your work and enlighten a whole new audience. Please consider this.

[@PatObi]

Raymeester great idea

[@PatObi]

Raymeester thanks.