For the equal amounts of trimming in each tail of the
distribution, the Winsorized sum of squared deviations is
defined as
SSDt j = (gj +1)(X(gj+1) j −Xt j)2+(X(gj+2) j −Xt j)2+: : :
+(X(nj−gj−1) j −Xt j)2+(X(nj−gj) j −Xt j)2
When allowing different amounts of trimming in each
tail of the distribution, the Winsorized sum of squared
deviations is then defined as,
SSDt j = (g1 j +1)(X(g1 j+1) j −Xt j)2+(X(g1 j+2) j −Xt j)2+: : :
+(X(nj−g2 j−1) j −Xt j)2+(g2 j +1)(X(nj−g2 j+1) j −Xt j)2
−
{(g1 j)[X(g1 j+1) j −Xt j]+(g2 j)(X(nj−g2 j) j −Xt j)}2
nj
Note that we used trimmed means in the SSDt j formula
instead of Winsorized means.
Hence the trimmed F is defined as
Ft( j) =
Jå
j=1
(Xt j −X j)2
(J−1)
Jå
j=1
SSDt j
(H − j)
where J = number of groups,
hj = nj −g1 jg2 j
H =
Jå
j=1
hj
and Xt =
Jå
j=1
hjHt j
H
Ft(g) will follow approximately an F distribution with
(J−1;H −J) degree of freedom.
2.2 Scale Estimator
Let X = (x1;x2; : : : ;xn) be a random sample from any
distribution and let the sample median be denoted by
medixi.
2.2.1 MADn
MADn is median absolute deviation about the median.
Given by
MADn = bmed|xi−medxi|
with b as a constant, this scale estimator is very robust
with best possible breakdown point and bounded
influence function. MADn is identified as the single most
useful ancillary estimate of scale due to its high
breakdown property [5]. This scale estimator is simple
and easy to compute.
⃝c 2013