I am wondering how to interpret the output of a discriminant analysis in SYSTAT? I have previously used SPSS and relied on the structure matrix to indicate which variables related to each function. In SYSTAT there is no structure matrix in the output - is there a way to generate this or another way to interpret the output that would provide the same information? Any advice gratefully received!
[SYSTAT] - help - discriminant analysis
- Thread starter krisfire
- Start date
You could use the standardized discriminant function coefficients to analyze the contribution each variable makes to the discriminant function. In fact this is a better option, since the structure matrix measures correlations between the variables and the functions, so those are used mainly for assigning meaningful labels to the discriminant functions.
I don't remember the output for a Discriminant Analysis in SYSTAT, but I'd like to think that they include this coefficients (MINITAB does not and I don't understand why:shakehead). Still, if you are working with only two groups you could obtain the coefficients, but with more than two groups it will be a little harder.
Hope that helps
I don't remember the output for a Discriminant Analysis in SYSTAT, but I'd like to think that they include this coefficients (MINITAB does not and I don't understand why:shakehead). Still, if you are working with only two groups you could obtain the coefficients, but with more than two groups it will be a little harder.
Hope that helps
Thanks for the advice, I was under the impression that in fact the structure matrix was meant to be more robust in terms of interpreting the variables on each function. In essence what I am aiming to do is understand which variables are responsible for the successful discrimination of the groups. SYSTAT does provide the coefficients, so I can look at these.
If you wouldn't mind a follow-up question...?
I have a strange situation that I am unable to understand regarding the structure matrix output. All of my 20 variables are most highly correlated with the 3rd function - all of the functions are significant, but the third is responsible for only perhaps 10% of the variance. In addition, the "loadings" are all fairly similar on each function (ie. all variables have a loading of .01 to.06 on function 1, .1 to .25 on function 2, .29 to .39 on function 3, .01 to .05 on function 4 etc).
I am not clear on what this means in terms of which variables are "driving" the analysis, as every other DFA I've ever seen resulted in the loadings spread more widely on each function and the highest correlations spread across across the functions.
Does this result make any sense to you?
If you wouldn't mind a follow-up question...?
I have a strange situation that I am unable to understand regarding the structure matrix output. All of my 20 variables are most highly correlated with the 3rd function - all of the functions are significant, but the third is responsible for only perhaps 10% of the variance. In addition, the "loadings" are all fairly similar on each function (ie. all variables have a loading of .01 to.06 on function 1, .1 to .25 on function 2, .29 to .39 on function 3, .01 to .05 on function 4 etc).
I am not clear on what this means in terms of which variables are "driving" the analysis, as every other DFA I've ever seen resulted in the loadings spread more widely on each function and the highest correlations spread across across the functions.
Does this result make any sense to you?
Both structure matrix and the coefficients should be used to fully understand the nature of your classification, still, many statisticians usually give a major importance to the coefficients.
Now, regarding your results, well, there are many reasons why you might not see much difference in your coefficients. First, make sure you are analyzing the standardized coefficients, which are more informative (specially if you have different scales). Try also looking at the classification table. Is your model good for classifying? Is it better than random? Look for the Jacknife's Leave-One-Out Cross Validation , that way you can see whether your predictors are useful or not. In fact, considering the number of variables you have, you could be over-fitting the model (that is, you have more predictors than you need); try reducing your variables and analyze how the miss-classification changes.
Finally, make sure your models met the assumptions of linear discriminant analysis, such as multivariate normality and equal covariance matrix among groups. Besides It seems you have many groups, so you would need a good sample size to provide power to the analysis and the discriminant analysis can be sensitive to big differences in sample sizes for groups.
I hope that's not too confusing
Good luck
Now, regarding your results, well, there are many reasons why you might not see much difference in your coefficients. First, make sure you are analyzing the standardized coefficients, which are more informative (specially if you have different scales). Try also looking at the classification table. Is your model good for classifying? Is it better than random? Look for the Jacknife's Leave-One-Out Cross Validation , that way you can see whether your predictors are useful or not. In fact, considering the number of variables you have, you could be over-fitting the model (that is, you have more predictors than you need); try reducing your variables and analyze how the miss-classification changes.
Finally, make sure your models met the assumptions of linear discriminant analysis, such as multivariate normality and equal covariance matrix among groups. Besides It seems you have many groups, so you would need a good sample size to provide power to the analysis and the discriminant analysis can be sensitive to big differences in sample sizes for groups.
I hope that's not too confusing
Good luckRe: SYSTAT help - discriminant analysis
Dear all
I think all of you can use and understand well on the Discriminant of SYSTAT.
Now in Vietnam, I am trying to use the Discriminant model to predict climate (Monthly temperature) for Vietnam. So I don't understand well about this model.
Could you please tell me step by step to do this?
Could you please reply to me by the email mau.imhen@gmail.com
Best regards and hope to hear good information from you soon
Mau
Dear all
I think all of you can use and understand well on the Discriminant of SYSTAT.
Now in Vietnam, I am trying to use the Discriminant model to predict climate (Monthly temperature) for Vietnam. So I don't understand well about this model.
Could you please tell me step by step to do this?
Could you please reply to me by the email mau.imhen@gmail.com
Best regards and hope to hear good information from you soon
Mau