Hi,
I wonder if there is any problem in using percetages to make a discriminant analysis (DA).
I have percentages representing the surface area ocuppy by a specific feature of a bone relative to the whole surface area of that bone (an multiply by 100). I have 5 percentages per individual (from 5 different bones) and I measured three different species. I would like to use a DA on those variables. When I make the DA, the first axis explain 100% percent of the variance which make me think that maybe there is a problem whith what I am doing; 100% is too much! no?. Any thoughts?
Hello Ana,
First response: Being able to account for 100% of variance is definitely out of the ordinary, so something is likely amiss.
Second response: It is unlikely that the use of percent as your metric is the problem.
How many cases are there (total, and per species)? It is possible to overfit in linear models when the number of predictors (in your case, the five measurements) is large relative to the number of cases. As well, Ette's question of what the respective variables are is relevant. Finally, be sure that you are eliciting the DA correctly.
Good luck with your work!
Hello Ana,
As David Morse David has said, having the first axis accounting for 100% of the variance is extremely odd. Also, as Ette Etuk's said, what are the variables? I will suggest looking again to your set of data, i.e., representing correctly your data, having enough cases to run the analysis and finally but not less if the type of variables (continuous or categorical) fits the type of DA that you are trying to use. In my opinion, this is what usually happens when you are having odd results.
Good luck with it!
No you can't. The distributions are different for percentages than means hence your p-values will be wrong. The relationships among variances won't be the same either. I don't know if anyone has worked out the case for %s or even if it's possible. In fact, I can't find anything with Google. Can you use means from the original data?. I have attached a review that might be helpful. Best, D.Booth
Thank you all for your responses.
My number of predictors are always three (3 different species), and the number of cases ranges from 25 to 30 per phalanx. The variables I am analyzing are the insertion site of a muscle in 5 phalanges (the insertion area occupied, for instance, the 20% of the second phalanx) of chimpanzees, gorillas, and humans. For instance, I have 25 percentages for chimpanzees, 28 for gorillas and 34 for humans for the second phalanx, a similar number of cases for the third, fourth and fifth phalanges. So, I don't think the sample size is the problem.
David, I can`t use means because they the information I am looking for. But means and most of the percentages are normally distributed and there is homogeneity of variances between the variables as well. Is this what you mean?
Is it possible that the DA is explaining the 100% of variance because all the variables I have are just representing size and nothing else (the size of a bonny feature relative to the size of the whole bone, in five different bones)?
No I meant that formulae for variances etc will not yield comparable values if % is substituted for the means. This is like you MUST use counts in a Chi-Sq test not %. Remember that from Genetics. Best, David
Dear David, The theoretical formulation of Discriminant Analysis requires that the dependent variable or variables being grouped or classified should be categorical variables while the independent variables which are predictors should be interval variables( or percentages). Therefore, before you can correctly do your Discriminant Analysis, you have to recode your dependent variable data from percentages to categorical data.
Harold Chike Thank you for agreeing that she can't use percentages in a discriminant analysis. Please however note that that the DVs must be continuous otherwise where is the multivariate normal error distribution. See development of the classification rule in the attached notes. Best, D. Booth
According to SPSS, the variable being classified needs to be re-coded to categorical form if it is in a continuous form. This makes sense. This is the first step of the classification process. Classification is a major purpose of Discriminant Analysis apart from its prediction capability. Please refer to the SPSS manual.
Regards.
@Harold Chike-I apologize for having just come across your last post to this ancient question. My comments: 1. If your source is an SPSS manual you are often on a very slippery slope. 2. Please see this classic text(BTW it is well worth having on your shelf):https://www.pearson.com/us/higher-education/product/Johnson-Applied-Multivariate-Statistical-Analysis-6th-Edition/9780131877153.html. BTW you can usually find a used copy for considerably less. May the Force be with you, David Booth
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics