Page 27 - Special Topic Session (STS) - Volume 1
P. 27
STS346 A.H.M. Rahmatullah Imon
focuses on both x and y dimensions at the same time, but also considers
classification techniques to identify outliers.
2. Methodology
Consider a standard regression model
= + (1)
where Y is an n-vector of observed responses, X is an × matrix
representing p explanatory variables including the constant, is a p-vector of
unknown finite parameters and is an n-vector of random disturbances with
2
() = 0 and () = . The traditionally used Ordinary Least Squares (OLS)
̂
̂
̂
−1
estimator of is = ( ) and the vector of fitted values is = =
. The matrix
−1
= ( ) (2)
is often referred to as weight or leverage matrix whose diagonal elements ℎ
̂
are termed leverages. The OLS residual vector ̂ is defined as ̂ = − .
Observations corresponding to exceptionally large ̂ values are termed
outliers. However, the question still remains how large is large? For this reason
we often consider the standardized version of residuals. One very popular
choice is deleted Studentized residuals (DSR) defined as
− ̂ (−)
= , = 1,2, … , (3)
̂ () √(1−ℎ )
Where ̂ (−) and ̂ (−) are the OLS estimates of and respectively with the
− ℎ observation deleted. We call an observation outlier when its
corresponding deleted Studentized residual value exceeds 3 in absolute value.
Observations corresponding to exceptionally large ℎ values are termed high
leverage points which are essentially outliers and high leverage points
simultaneously rather than separately. Gray (1986) proposed the Leverage-
Residual ( − ) plot where the leverage value ℎ for each observation is
plotted against the square of a normalised form of its corresponding residual.
The bulk of the cases will be associated with low leverage and small residuals
so that they cluster near the origin (0,0). The unusual cases will have either
high leverages or large residual components and so will tend to be separated
from the bulk of the data. High leverage cases will be located in the upper area
of the plot and observations with large residuals will be located in the area to
the right.
16 | I S I W S C 2 0 1 9