Skip to content
Prev 258 / 523 Next

[RsR] Singular covariance in plot.lmrob

Hi Christian, Matias,

With this example covMcd() fails before starting the actual algorithm - the
first to do is to standardize the data by (for each variable) subtracting
med and dividing by mad (actually not mad but the h-quantile) . Of course
this is not possible for h= (n+p+1)/2=199 since 204 observations in variable
'cylinders' (unfortunately most of the cars are with 4 cylinders). You could
avoid this by increasing alpha (reducing the breakdown point).

Then comes the __false__ error message that 14 observations lie on a
hyperplane - they are 204 out of 392, I am going to investigate the FORTRAN
code to find out what is happening.

But even in case of of more than h observations lying on a hyperplane
covMcd, apart from issuing an warning message, returns the MCD location and
scatter matrix (of course singular).

In ltsReg() you can specify mcd=FALSE if you do not want to investigate the
leverage information (and you do not want in this case, don't you?) and this
will save you from the failure in the plot function.

Best regards,
Valentin





On Fri, Nov 28, 2008 at 9:31 PM, Matias Salibian-Barrera <matias at stat.ubc.ca