I don't comment on videos. However, I have read a lot of books and watched a lot of videos on DA this week. This video, sir, is the very best source on DA that I have ever come across (and I'm not statistical slouch, so I feel like my opinion is worth a tiny bot of something). Keep up the fantastic work!
@Mike Crowson. I really need help. My sample size is 3,000 respondents. I did DISCRIM and Box M is sig. at 0.00000 (Box's M=99,569, Approx = 3,547, df1=28, df2=1724964, Sig.=0.000). My object is compare the perception of product among respondents in East and West. Can I ignore the assumptions (IVs are normal distribution) because almost my IVs are non-normal? Thank you so much for your useful video. It is really helpful!
Hi there. Thanks for your question. First, I wanted to mention that I actually have a new discriminant analysis video (with dataset and Powerpoint) that you can access here: th-cam.com/video/WttCpaDLKBg/w-d-xo.html In that video and Powerpoint I talk a little about assumptions, which include multivariate normality and homogeneity of covariance matrices. Keep in mind that with 3,000 respondents, Box's test is likely to be overpowered, even with minor departure from homogeneity of covariance matrices. So I'm not sure I'd put too much weight on that test result. You should also keep in mind that Box's test is impacted by multivariate non-normality, so that's another consideration when evaluating that test result. MANOVA has the same assumptions regarding normality and homogeneity of covariance matrices as discriminant analysis - as it is mathematically equivalent (for more details, including evaluation of assumptions with MANOVA, see th-cam.com/video/hs8CA_kWkao/w-d-xo.html). MANOVA is fairly robust to a violation of this assumption. Personally, I have not spent much time examining the stability of the discriminant functions and discriminant loadings in the context of multivariate non-normality. However, Tabachnick and Fidell (2013; p. 384) state that discriminant analysis is "robust to failures of normality if violation is caused by skewness rather than outliers". So, you should probably spend some time investigating for possible outliers (see the MANOVA video referenced above). They also note that robustness is impacted by the degree of discrepancy in sample sizes between groups, and recommend larger sample sizes as differences in size between groups become more substantial. In general, you should read up more on pages 384-385 of Tabachnick and Fidell. As an aside, if you are less interested in identifying and describing discriminant functions and more interested in identifying predictors of group membership, you might consider logistic regression. You can use binary logistic regression (video here: th-cam.com/video/cpWSSJHuT2s/w-d-xo.html) if your outcome is dichotomous, ordinal logistic regression (video here: th-cam.com/video/rSCdwZD1DuM/w-d-xo.html) with ordered categorical outcome, or multinomial logistic regression (video here: th-cam.com/video/1BL5cL8_Cyc/w-d-xo.html) if your outcome is nominal. Hope this helps. Cheers!
DEAR VIEWERS: I HAVE JUST UPLOADED A MORE RECENT VIDEO ON DISCRIMINANT ANALYSIS AT th-cam.com/video/WttCpaDLKBg/w-d-xo.html . ADDITIONAL LEARNING MATERIALS ARE AVAILABLE UNDERNEATH THE VIDEO DESCRIPTION, SO PLEASE BE SURE TO CHECK THIS OUT!
I don't comment on videos. However, I have read a lot of books and watched a lot of videos on DA this week. This video, sir, is the very best source on DA that I have ever come across (and I'm not statistical slouch, so I feel like my opinion is worth a tiny bot of something). Keep up the fantastic work!
Thank you, great through explanation!
@Mike Crowson. I really need help. My sample size is 3,000 respondents. I did DISCRIM and Box M is sig. at 0.00000 (Box's M=99,569, Approx = 3,547, df1=28, df2=1724964, Sig.=0.000). My object is compare the perception of product among respondents in East and West. Can I ignore the assumptions (IVs are normal distribution) because almost my IVs are non-normal?
Thank you so much for your useful video. It is really helpful!
Hi there. Thanks for your question. First, I wanted to mention that I actually have a new discriminant analysis video (with dataset and Powerpoint) that you can access here: th-cam.com/video/WttCpaDLKBg/w-d-xo.html
In that video and Powerpoint I talk a little about assumptions, which include multivariate normality and homogeneity of covariance matrices. Keep in mind that with 3,000 respondents, Box's test is likely to be overpowered, even with minor departure from homogeneity of covariance matrices. So I'm not sure I'd put too much weight on that test result. You should also keep in mind that Box's test is impacted by multivariate non-normality, so that's another consideration when evaluating that test result. MANOVA has the same assumptions regarding normality and homogeneity of covariance matrices as discriminant analysis - as it is mathematically equivalent (for more details, including evaluation of assumptions with MANOVA, see th-cam.com/video/hs8CA_kWkao/w-d-xo.html). MANOVA is fairly robust to a violation of this assumption. Personally, I have not spent much time examining the stability of the discriminant functions and discriminant loadings in the context of multivariate non-normality. However, Tabachnick and Fidell (2013; p. 384) state that discriminant analysis is "robust to failures of normality if violation is caused by skewness rather than outliers". So, you should probably spend some time investigating for possible outliers (see the MANOVA video referenced above). They also note that robustness is impacted by the degree of discrepancy in sample sizes between groups, and recommend larger sample sizes as differences in size between groups become more substantial. In general, you should read up more on pages 384-385 of Tabachnick and Fidell.
As an aside, if you are less interested in identifying and describing discriminant functions and more interested in identifying predictors of group membership, you might consider logistic regression. You can use binary logistic regression (video here: th-cam.com/video/cpWSSJHuT2s/w-d-xo.html) if your outcome is dichotomous, ordinal logistic regression (video here: th-cam.com/video/rSCdwZD1DuM/w-d-xo.html) with ordered categorical outcome, or multinomial logistic regression (video here: th-cam.com/video/1BL5cL8_Cyc/w-d-xo.html) if your outcome is nominal. Hope this helps. Cheers!
Excellent presentation. Thank you so very much . . . you have been a great help to me both last year and this year.
thanks for your feedback. Much appreciated! Good luck with your research!
That is a good video on discriminant analysis
DEAR VIEWERS: I HAVE JUST UPLOADED A MORE RECENT VIDEO ON DISCRIMINANT ANALYSIS AT th-cam.com/video/WttCpaDLKBg/w-d-xo.html . ADDITIONAL LEARNING MATERIALS ARE AVAILABLE UNDERNEATH THE VIDEO DESCRIPTION, SO PLEASE BE SURE TO CHECK THIS OUT!