robust statistic

ml

what is robust statistic
good or bad?
how to deal with outlier
any model is robust or non-robust?
performance?

• link: intro
• Robust statistics provides an alternative procedure, which provides a model describing the ‘good’ part of the data,
  but does not require us to identify specific observations as outliers or exclude them

example:

• the median method: take the central value of the
  ordered data (the median) as the estimate of the mean. outliers do not affect
• Huber’s method

analysis

• Using robust estimates of mean and standard deviation to predict future values from a normal distribution may mislead the unwary because the presence or probability of outliers is not predicted.
• Obtaining a robust statistical model of a data set provides probably the best method for identifying suspect values for further investigation.
• Robust methods assume that the underlying distribution is roughly normal (and therefore unimodal and symmetrical) but contaminated with outliers and heavy tails. The methods will give misleading results if they are applied to data sets that are markedly skewed or multimodal, or if a large proportion of the data are identical in value