PLS bootstrap confidence interval
Zara Bergstrom
Posted on 03/10/08 10:26:18
Number of posts: 6
Dear PLS experts,
I would very much like your opinions on a discussion below, in order to clarify some things about how bootstrap testing works in PLS.
I'm using PLS on ERP data. My understanding of the bootstrap procedure in PLS is that the standard errors of the electrode saliences are estimated through bootstrap sampling, using
sampling with replacement and keeping the experimental conditions fixed
for all observations, and the electrode saliences are then recalculated
for each bootstrap sample. The ratio of the observed electrode salience to the bootstrap standard error
provides a standardised measure of the reliability of the electrode
salience, which is approximately equivalent to a z score (if the bootstrap distribution is normal).
We had in a manuscript used an absolute bootstrap ratio value of 1.96 as a cut-off for
determining whether the electrode salience was reliable, because values
above 1.96 and below -1.96 are significant at the P< 0.05 level (in line with prior research, although some people use a higher cut-off, e.g. 3 for P < 0.001).
A reviewer of our manuscript suggested that we should refer to the
ratio of 1.96 a corresponding to a 95% confidence interval, which
emphasizes the issue of reliability instead of referring to significance
values.
Am i correct in thinking that what they mean by this is that if the
bootstrap ratio is below -1.96 or above 1.96, the 95% CIs of the electrode
salience does not encompass zero, so it is hence reliably different from
zero with 95% confidence? Is this latter statement actually the case?
I asked these questions to a statistics expert in our department who responded with the following comments (I have shortened these a bit):
-----------------------------------------------------------------------
"In essence, the
issue is quite a subtle one. Significance testing is based on the null hypothesis ie assume there is no
effect and how big could the statistic be by chance. hence: "Oh look, z is
>2; that's too rare to be chance" (accepting a 5% error rate in that
statement).
bootstrapping is not a null hypothesis procedure as you randomise with
replacement: permutation tests are the resampling method appropriate to
null hypothesis significance testing. So the reviewer didn't like your use
of bootstrapping apparently to test significance.
so what does the bootstrap do? well in this case it is providing confidence
intervals for the electrode saliences -- and so asks are these electrode
saliences reliable / reliably different from zero?
The weird thing in the PLS approach is that it doesn't use the bootstrap in
a conventional way which probably antagonised/confused the reviewer a bit.
normally, the bootstrap distribution would directly give one the 95%
confidence intervals (ie take 10000 samples and compute the lowest 500 and
highest 500 to get the cutoffs). In PLS the bootstrap seems to be being used as a method of
estimating the standard error of the normal distribution numerically rather
than using a theory to derive a formula for it. This is a bit odd wrt the
standard bootstrapping approach.
I wonder if this odd usage of bootstrap in PLS is just a short-cut thing.
To use the bootstrap conventionally it would be typical to take 10000
samples in order to get a proper distribution of the electrode saliences
and so estimate the CIs. but with many electrodes to check that would take
ages even on a fastish computer. If you were just assuming normality and
using the bootstrap to calculate the standard error, you could probably get
away with a smaller sized bootstrap sample. What do you/they use in PLS for
this step?"
----------------------------------------------------------------------------
I also have two further questions:
First, whether an additional reason for reporting the bootstrap ratio rather than the CIs of the electrode saliences is that
the electrode saliences are not standardised? Since they are dependent on the
raw amplitudes of ERP data, directly reporting the CIs of the electrode
saliences themselves might be more confusing, because they would be
variable across time and location only on the basis of overall amplitude
differences?
Second, whether there is any reason to believe that a bootstrap distribution of electrode saliences would be non-normally distributed under certain cases? And if so, when?
I hope my questions make sense, and thank you very much in advance,
Zara