Multidimensional Reliable Outlier Detection Using R

Question

What is the preferred method (in your opinion) for automatically detecting multidimensional reliable detection of outliers in R, i.e. without manual inspection and scheduling?

I found the dprep package, but it seems to be terminated. However, since discovery detection is a frequent and important task, a common default method should be available, for example. MCD score (Rousseeuw and Van Driesen, 1999).

Answer 1

Try covMcd in robustbase package.

Answer 2

Use Cook Distance You could use the distance from cooking. The exposure time is calculated based on the linear regression model. This means that you can include several X variables to calculate outlier (observations with high influence, more precisely). This effectively gives you the ability to add or omit the variables by which you want to determine deviations. A way to compute it for each observation in R would look something like this:

mod <- lm(Y ~ X1 + X2 + X3, data=inputData) cooksd <- cooks.distance(mod) 

In general, these observations with a cooking distance> 4 * mean (cook distance) are considered outliers. For more information on the formula and interpretation of the roll distance, see this example.

Disclaimer: I am the author.