Dear Michael
You asked about confidence intervals in plotting PSTHs - here's my 2p's
worth:
The spm_graph error bars are, I believe, 90% confidence intervals for a
given subject (based on the standard errors of the relevant contrasts of
the betas eg for each FIR time bin). If your inference is based on a
multi-subject fixed effects FIR model you could use the standard error of
the contrast across all subjects but this is not the same as the average of
the within-subject standard errors so I can't see a reason why one would
want to average the .pcis (I’m sure someone will correct me if I’m wrong!)
Generally, what you want to plot in terms of confidence intervals/ variance
measures will depend on your statistical inference - fixed or random
effects, within- or between-groups. Russ & Karl’s earlier emails explain
this general point - see
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0304&L=spm&P=R21469&I=-1
It's also worth considering the role of FIR-PSTH plots in particular in
this context (see Rik's message
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0311&L=spm&P=R154&I=-1). If
your statistical tests are based on a model that uses a non-FIR basis set,
variability in the FIR estimates does not relate in any clear way to the
inference you are making. Some people include plots as a check that the
basis set is comprehensive enough, for example in between-group
comparisons. But you may find that some referees prefer plots of the
parameter estimates on which one's statistics are based.
Hope this helps
Alexa
> > PS -- The detail you added to the documentation about pci conversion to
> > SEM helps a lot. I was wondering about that myself. I imagine you could
> > average the .pci across subjects the same way I'm trying to average
> > .psth? Or, is it more appropriate statistically to make a gargantuan
> > fixed-effects model and simply extract a single psth and pci for each
> > regressor of interest?
>
> I don't know the correct statistical answer to this. I think there are
> tradeoffs doing it either way. We have tried both methods including
> multi-session runs and then averaging the session psth results, vs single
> session runs. A problem with multi-session runs occurs if some sessions
> have very few events of a particular type, then the estimation of the
> psth can go horribly wrong, for example if there is only a single event.
> Single session runs can cause a problem if some of your events bridge
> across the run barrier, so you have to be careful to try not to estimate
> events too close to the end of one run and the start of the next.
>
> Darren
|