[no subject]

BAD MSG:
estimates per subject using this method. I wonder what difference it
ould make using less reversals per threshold, or a different lag than
20. How did the five accepted tracks inter-correlate? You could probably
correct as Kollmeier et al did for the lack of independence, if you know
how much dependence there is and then add the (correlation * variance)
to the error term. Your approach sounds similar (in principle) to one by
Wetherill in the 1960s. He argued that if you took the average signal
level in pairs of peaks and valleys, you reduce the serial correlation
to the point where the SE estimate gets some validity. The problem with
Wetherill's approach is that it is highly sensitive to variability in
the psychometric function :-if the threshold or shape of the actual or
estimated function varies then an extra source of autocorrelation creeps
in which the pairs of up-downs correlate. I did some montecarlos to see
how well Wetherill's approach performs. It did poorly and underestimates
the variance almost as badly as no correction at all. I also found that
the variance within a track is an excellent predictor of the variance
across tracks provided the psychometric function is constant and you
know what the slope parameter of the PF is (you can also roughly
approximate on a 3-up 1-down staircase that the SE is approximately
twice the SD of the reversals). But ofcourse, rough approximations
aren't good enough, the PF is never constant and there are biases
involved in estimating the PF from a single staircase track, so it never
went anywhere...

Chris

--
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Chris Chambers
Department of Psychology
Monash University
Clayton, Victoria 3168
AUSTRALIA

Tel. +61 3 9905 3978
Fax. +61 3 9905 3948

EMAIL: chris.chambers@sci.monash.edu.au
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-