help
help > RE: Multivariate regression - assumptions?
Oct 11, 2015 07:10 PM | Alfonso Nieto-Castanon - Boston University
RE: Multivariate regression - assumptions?
Hi Jenna,
Could you please clarify the specific form of your second-level analyses? There are at least two main ways in which you could be looking at the association between 4 clinical variables and FC, either using FC as a predictor variable and clinical scores as outcome variables (e.g. in Tools.Calculator; I am guessing you might be using this since you mention multivariate regression) or using clinical scores as predictors and FC as an outcome variable (e.g. in Tools.Calculator or in the second-level results tab; my guess is that you might not be using this since this would be just a multiple regression instead). In general, CONN uses the standard General Linear Model for all second-level analyses, so the typical normality/heteroscedasticity assumptions apply, but there is no explicit testing/checking in CONN of the validity of any of these assumptions in your data. If I am understanding correctly your analyses, the "equal variance" assumption does not really apply to these analyses (either because you are using a single FC outcome variable on a single group of subjects, or because you are using a multivariate test with multiple clinical score outcome variables -and in this case the clinical scores variance/covariances are explicitly modeled by the multivariate test). In any way, if you let me know more details about your case perhaps I can give you a more specific answer and/or suggest ways to test the standard normality/heteroscedasticity assumptions.
Hope this helps
Alfonso
Originally posted by Jenna Traynor:
Could you please clarify the specific form of your second-level analyses? There are at least two main ways in which you could be looking at the association between 4 clinical variables and FC, either using FC as a predictor variable and clinical scores as outcome variables (e.g. in Tools.Calculator; I am guessing you might be using this since you mention multivariate regression) or using clinical scores as predictors and FC as an outcome variable (e.g. in Tools.Calculator or in the second-level results tab; my guess is that you might not be using this since this would be just a multiple regression instead). In general, CONN uses the standard General Linear Model for all second-level analyses, so the typical normality/heteroscedasticity assumptions apply, but there is no explicit testing/checking in CONN of the validity of any of these assumptions in your data. If I am understanding correctly your analyses, the "equal variance" assumption does not really apply to these analyses (either because you are using a single FC outcome variable on a single group of subjects, or because you are using a multivariate test with multiple clinical score outcome variables -and in this case the clinical scores variance/covariances are explicitly modeled by the multivariate test). In any way, if you let me know more details about your case perhaps I can give you a more specific answer and/or suggest ways to test the standard normality/heteroscedasticity assumptions.
Hope this helps
Alfonso
Originally posted by Jenna Traynor:
Hello,
I am inquiring about the assumptions of normality and heteroscedasticity of residuals, specifically - when running a multivariate regression using the CONN toolbox, are these assumptions already controlled for in the second-level results?
I have used the toolbox to look at the linear association between 4 clinical variables and FC within a group of subjects. I am wondering if the significant results that I am seeing have normally distributed residuals and if the residuals have equal variances - is there a way for me to extract residuals to examine/plot them or has this already been taking into account?
Likewise - are there any papers that I can reference that explain how CONN toolbox deals with these assumptions?
Many thanks,
Jenna
I am inquiring about the assumptions of normality and heteroscedasticity of residuals, specifically - when running a multivariate regression using the CONN toolbox, are these assumptions already controlled for in the second-level results?
I have used the toolbox to look at the linear association between 4 clinical variables and FC within a group of subjects. I am wondering if the significant results that I am seeing have normally distributed residuals and if the residuals have equal variances - is there a way for me to extract residuals to examine/plot them or has this already been taking into account?
Likewise - are there any papers that I can reference that explain how CONN toolbox deals with these assumptions?
Many thanks,
Jenna
Threaded View
| Title | Author | Date |
|---|---|---|
| Jenna Traynor | Oct 11, 2015 | |
| Alfonso Nieto-Castanon | Oct 11, 2015 | |
| Jenna Traynor | Oct 12, 2015 | |
| Alfonso Nieto-Castanon | Oct 14, 2015 | |
| Shruti Kinger | 22 hours ago | |
| Jenna Traynor | Oct 14, 2015 | |
| Alfonso Nieto-Castanon | Oct 14, 2015 | |
| Jenna Traynor | Oct 26, 2015 | |
| Jenna Traynor | Oct 28, 2015 | |