maxsum vs. maxsumt

[Michael Wibral]

Hi Eric,

thank you for pointing out the importnace of smalllestcluster for the
potential significance estimations of large cluster.

Best Regards,
Michael Wibral

Eric Maris schrieb:

> Hi Michael,
>
>
>> I just noticed that my version of clusterrandanalysis (20051127)
>> doesn't list the teststatistic 'maxsumt' in its header or in the
>> actual code as it is used in the tutorial (toolkit 2005). Instead
>> there's an option (the default) 'maxsum'. Are these two the same - or
>> is maxsum just the summed raw effect within a cluster ( and thus
>> maybe pretty sensitive to extreme outliers), whereas maxsumt refers
>> to summed t-scores from the initial tests performed for cluster finding?
>
>
> The change from 'maxsumt' to 'maxsum' occured when clusterrandanalysis
> also supported F-statistics at the level of
> (sensor,frequency,time)-triplets. Using 'cfg.clusterteststat=maxsum'
> in combination with 'cfg.statistic=indepsamplesT' gives the same
> results as the 'maxsumt'-option of the older version of
> clusterrandanalysis.
>
>
>> I also notice that maxsum 'always' yields more significants clusters
>> than orderedsums, is there an explanation for this behaviour (I
>> thought it might rather be the other way round...)?
>
>
> I am not surprised. The behavior of 'cfg.clusterteststat=orderedsums'
> depends on the value assigned to 'cfg.smallestcluster'. If you give
> 'cfg.smallestcluster' a small value, this will reduce the sensitivity
> for larger clusters (larger than this small value). The option
> 'cfg.clusterteststat=orderedsums' will result in more significant
> clusters than obtained with 'cfg.clusterteststat=maxsum' if the data
> contain one big cluster and a number of much smaller (but systematic)
> clusters. These smaller clusters may become significant with
> 'cfg.clusterteststat=orderedsums' if their size is larger than
> 'cfg.smallestcluster'. The value of 'cfg.smallestcluster' should be
> determined independently of the data. I guess there is no need to
> explain why this so.
>
>
> greetings,
>
> Eric Maris
>
> .
>