Can SPSS be used to identify and exclude outliers in a multivariate data system? If not can someone suggest a software?
Yes it can be used and it is used by the researchers for identifying multivariate outliers using mahalanobis distance and some further calculations. If one will do a Google search for identification of multivariate outliers one will find a rich literature to deal with this issue
SPSS is fine for identifying outliers. i have used it. It's in the section 'Descriptives', and I click on Box-plots (if I remember correctly). I don't have my lap-top with me and I don't have SPSS in my present desktop.
Since R is one of the topics, I suggest R-library mvoutlier. My experience is from seismic travel time data, though.
Thank you Rakesh, Miranda and Timo
I have found SPSS useful for the purpose.
I shall try to explore into the possibility of using R-library mvoutlier, which Iam not familiar with, though.
SPSS's Box plots are more easy and convenient in interpretations...
Boxplots are not how you identify multivariate outliers. Consider a simple regression with variables x and y and a near perfectly fitting line. Now, imagine the points, while all very close to, or on, the line, are distributed in both x, and y, all very close to the beginning of the line except for one which is far out near the end. You have a good fitting model, haven't violated any assumptions, and you fit all of your data. Is the single point far out an outlier? In boxplots of either x or y it may appear to be. This gets even more complicated when you consider complex multivariate data where the definition of an outlier would have to be in the context of the multivariate space.
I don't know how to do it in SPSS but the mvoutlier might be a good start in R.
The bigger problem answering this question is that you haven't defined what a multivariate outlier is beforehand. You need to do that first. Until you do that no amount of software you throw at it is going to solve the problem.
I also suggest besides boxplots, there two more methods in SPSS, one is mahalanobis distances and cookes distances. Cookes reports where availability of outliers.
Rapid miner is a good one. It has a "Detect Outlier" Operator which will add a binary attribute to your data set for determining whether each record is an outlier or not.
You may also download "Anomaly detection" extension for RapidMiner, which includes 13 algorithms for outlier detection. Downloads from here: http://madm.dfki.de/rapidminer/anomalydetection
Just a little hint here that was not pointed out: With unsupervised algorithms that are included in "Anomaly detection" extension for RapidMiner you do not need to specify multivariate outliers beforehand (a priori). So, unlabelled data is NOT a computational problem in outlier detection
Outliers detection, normality tests and a wide range of facilities are also provided by Gretl which is a friendly free open source software.
Gretl is available in different languages (English, French, Italian, Spanish, Polish, German, Portuguese, Russian, Turkish, Czech, Traditional Chinese, Albanian, Bulgarian and Greek)
http://gretl.sourceforge.net/
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS