Mike Lawrence | 23 Feb 23:54 2009

Univariate Location estimator for skewed distributions?

Hi all,

Looking for a little advice on which robust estimator to use. Gist: I
need a univariate estimator of location that performs well when the
true distribution is positively skewed.

Context: I study human manual response time (RT), which yields
notoriously positively skewed data. What's more, this skew can be
systematic such that a given experimental manipulation may affect both
skew and location of the RT distribution. Much work has been done to
independently quantify these effects, relying on such procedures as
maximum likelihood estimation of ex-gaussian parameters, etc. However,
such procedures necessitate >50 RT observations per cell of interest
and many common experiments collect as few as 20 observations per
cell. In such circumstances I'm relatively resigned to focusing on
location only, but this is a field where outliers, manifestations of
the human failing to pay attention to the response stimulus, are
common, yielding occasional large positive data points. These outliers
could possibly be represented as a uniform distribution.

With this context in mind, I'm hoping the list could direct me to R
code for an appropriate robust estimator of location. That is, an
estimator that is OK if the true source distribution is skewed but
that will be robust against outliers as described above.




(Continue reading)