I have 400 samples chemically analysed for 48 elements. Some elements are correlated. My goal is to identify outliers in a multivariate space, samples that are unique in their composition. I compute Mahalanobis distance after auto-scaling and normalizing the data to find anomalous samples. Now, I like to know what variables (elements) or variable combinations contribute most to extreme Mahalanobis distance. I like to display the results in a bar diagram.

I use factor analysis with oblique rotation to determine factor loadings and sample scoring on the components. I retain all 48 components. Now, I investigate high scores on the components and high residuals ( sample being off the line of components). I expect high scores on the components contribute most to MAHA-distance. The same should apply to high residuals but I don't know how to describe it mathematically. So far I haven't considered the scaling along the component vectors (shrinking or enlarging) to account for correlation. It has to do with the eigenvalue of the component. Do I take care of the scaling by multiplying the sample score on the component by 1/eigenvalue? I'm not sure if all of it is a viable approach or if I should rather look into multivariate regression to solve my goal?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.