First task: compare 2 conditions

RFX in SPM5
Floris de Lange
[email protected]
RFX Options
Compare 2 conditions
Conditions are MI LH (press left foot) and MI RH (press right foot); 8
subjects; threshold used is p<0.001 uncorrected, k>20 (arbitrary)
Methods used:
• One-sample T-test on difference images MI LH>MI RH
• Paired-samples T-test on MI LH and MI RH
•Measurements assumed independent
•Measurements assumed dependent
• Two-samples T-test on MI LH and MI RH
• Multiple regression analysis on MI LH and MI RH
• Full factorial
• Flexible factorial
The gold standard: one-sample T-test
Paired T-test: dependence/indepence
Independent: error covariance matrix = identity matrix (check SPM.xVi.V!)
Dependent: error covariance matrix will be estimated (check SPM.xVi.V!)
Paired-samples T-test dep.: same
Paired-samples T-test indep.: same
Dependence/inde
pendence doesn’t
make a difference
here, because
there’s only one
sample to
estimate
covariance from
Multiple regression analysis: same
= identical
Two-samples T test indep: worse
Degrees of freedom ↑
Variance term ↑
Two-samples T test dep: better
• the correlation
between the variance
of the subjects in the
first group and those in
the second group is
estimated
• this reduces the error
term
Two-samples T test: dep vs indep
Dependent measures
Independent measures
Two-samples T test: con images
Dependent measures
=
Independent measures
Two-samples T test: ResMS images
Dependent measures
<
Independent measures
Error terms is reduced for dependent
measures by modelling the
dependencies
Full factorial dep. = 2-sample T dep
Full factorial indep. = 2-sample T indep
Flex factorial dep. = 2-sample T dep
Flex factorial indep = 2-sample T indep
Summary
•There are two types of models:
• Models that specify the subject factor (e.g., one-sample,
paired-samples, MRA if you specify the factor yourself)
• Models that estimate the subject factor (e.g., two-samples Ttest, full factorial, flexible factorial; measurements are
dependent)
• If you don’t specify the subject factor, but also don’t estimate
the error covariance, you are likely to shoot yourself in the foot
because the errors will be assumed to be independent, and
simply added, leading to much higher estimates of the error
term
Is it valid to use 2-sample T test dep?
• It can be statistically beneficial to
specify the model as a “betweensubjects” model without modelling
subject, but instead estimating the
subject-induced regularities by
specifying that measures may be
dependent
• SPM5 manual suggests to do analyses
this way
• But is it valid? Aren’t df’s inflated?
SPM5 manual