turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- Proc GLM

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-01-2015 10:16 AM

I was experimenting a MANCOVA analysis as below.

My Sample data as follows.

Venue | Bowling | Overs | Runs | Wkts |

Home | Spin | 594.4 | 2583 | 71 |

Away | Fast | 396.1 | 1897 | 68 |

My code as follows.

proc glm data= cricket;

class bowling ;

model Wkts runs = overs bowling /ss3;

manova h=_all_ ;

run;

Result shows that P value is significant for both overs (covariate) and bowling (Indepedent).

My questions are,

1. Is my analysis (mainly selection of covariate) is right to conduct a study on MANCOVA?

2. What other procs can be used for MANCOVA?

3. What else we should look for in result regardless of P value?

4. How can we get the sample data to further expand this study as I don't seem to have the data for this analysis?

I appreciate for any help you offer for me. Thanks.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-01-2015 01:58 PM

- If you think the term belongs in the model, then this is fine
- As far as I know, PROC GLM is the only one.
- MANOVA and MANCOVA can be analyzed via bi-plots to produce graphical representations of the results, sometimes resulting in greater understanding of the results, if that's of interest to you. This (and a lot of other relevant information) is discussed in many books, for example "Applied Multivariate Statistics with SAS Software" by Khattree and Naik
- I don't understand the question.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-02-2015 08:39 AM

Just noticed that there should be moderate correlation between dependent variables and no correlation between covariate and independent variable whereas in my data there was no correlation between dependent variables (Wkts and runs).

Result which produced from proc GLM seems to be valid (P value is significant) although was no correlation between dependent variables. I wanted to know whether I can stick with this analysis which I did before (as shown in my initial post) or I should tweak the analysis until I find the moderate correlation between dependent variables

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-02-2015 10:12 AM

I wanted to know whether I can stick with this analysis which I did before (as shown in my initial post) or I should tweak the analysis until I find the moderate correlation between dependent variables

I really don't know what this means. The data that you have will either show correlation between dependent variables, or not. Tweaking the analysis doesn't change the existing correlations.

I am also somewhat skeptical of the claim that there is no correlation between dependent variables, as I have never seen this happen with two real-world variables, there is always non-zero correlation. (I'm not talking about some textbook problem where you might get the correlation between dependent variables to be exactly zero). Do you mean the correlation is close to zero? And why is that a concern?

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-02-2015 10:23 AM

Yes, correlation is close to zero (<0.001) for dependent variables.

According to the documentation http://www.statisticssolutions.com/multivariate-analysis-of-covariance-mancova/, there should be a moderate correlation (values between 0.2 to 0.7) between dependent variables to conduct this study. It is my belief that our claim is skeptical if we violate this assumption.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-02-2015 10:34 AM

If the correlation is really <0.001, I see no harm in doing a MANOVA or MANCOVA, although I don't really see a benefit either. The multivariate aspect of the analysis will simply separate the two dependent variables into two "dimensions", where each "dimension" is essentially equivalent to the univariate analysis.

But there is no "tweak" to the analysis that will get you there. The correlations don't change no matter what you do to analyze them.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-02-2015 10:47 AM

Thanks for the quick reply.

Can I assume that my initial analysis is correct although the correlation **two dependent variables** is close to zero? What about the correlation between independent variable and covariate?

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-02-2015 11:35 AM

Can I assume that my initial analysis is correct although the correlation

two dependent variablesis close to zero?

My personal opinion is that there is nothing wrong with your multivariate analysis in the presence of nearly zero correlation between dependent variables, and that there also isn't any real benefit as well.

When the independent variables are correlated, this causes difficulties in Ordinary Least Squares (ANOVA or ANCOVA or Regression), specifically the variance of the estimates of the model terms are inflated because of the correlation between the independent variables; and similarly the variance of the predictions in inflated as well. In this case, I would recommend Partial Least Squares (PROC PLS in SAS) instead of Ordinary Least Squares, where studies have shown that the mean squared error of the predictions and the mean square error of the model coefficients fit by Partial Least Squares is lower than the same model fit using Ordinary Least Squares (but the coefficients and predictions are biased using Partial Least Squares). Reference: Frank,Ildiko E., and Jerome H. Friedman. "A statistical view of some chemometrics regression tools."....

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

04-02-2015 10:45 AM

there should be a moderate correlation (values between 0.2 to 0.7) between dependent variables to conduct this study

I don't see this advice at the link you gave. I also see at that link that "The dependent variables cannot be too correlated to each other. Tabachnick & Fidell (2012) suggest that no correlation should be above r = .90." I would disagree with this statement.