Missing data in 3dLME versus including betas with low trial count

Hello AFNI team,

I'm using 3dLME for my group analysis (n=32) and each subject has 8 conditions. Within each condition there is a maximum of 4 trials. For some of the subjects, there are only 1-2 trials that would create the beta coefficient used in the model. In that case, 1) would the data reliable enough to include in the model with 1-2 trials? Or would it be preferable to exclude those conditions? Approximately 23 out of the 736 conditions would be excluded if that's the way to go.

Thank you,
Catherine

Is this an event-related or a block-designed study? Given the potential for varying uncertainty across conditions, I recommend considering the use of 3dMEMA for each contrast.

Gang

Hello Gang,

Thank you for the response. It's a within-subjects design event-related study. I'm interested in testing for brain regions changing according to a specific function (e.g., power law) across my task conditions. I chose 3dLME for that reason - is 3dMEMA capable of doing the same? I've included an example of my 3dLME model. Currently I don't run any specific GLTs at the group level - I cluster threshold the F map.

3dLME -prefix PF
-jobs 8
-model 'PF'
-ranEff '~1+PF'
-qVars 'PF'
-mask mask.nii.gz
-dataTable
Subj Cond PF InputFile
sub-001 t2017to2015 0.333333333 3dREML_output+tlrc.BRIK[591]
sub-001 t2014to2012 0.083333333 3dREML_output+tlrc.BRIK[594]
sub-001 t2011to2009 0.047619048 3dREML_output+tlrc.BRIK[597]
sub-001 t2008to2006 0.033333333 3dREML_output+tlrc.BRIK[600]
sub-001 t2005to2003 0.025641026 3dREML_output+tlrc.BRIK[603]
sub-001 t2002to1998 0.020833333 3dREML_output+tlrc.BRIK[606]
sub-001 t1997to1993 0.015873016 3dREML_output+tlrc.BRIK[609]
sub-001 t1992to1988 0.012820513 3dREML_output+tlrc.BRIK[612]
sub-002 t2017to2015 0.333333333 3dREML_output+tlrc.BRIK[591]
sub-002 t2014to2012 0.083333333 3dREML_output+tlrc.BRIK[594]
sub-002 t2011to2009 0.047619048 3dREML_output+tlrc.BRIK[597]
sub-002 t2008to2006 0.033333333 3dREML_output+tlrc.BRIK[600]
sub-002 t2005to2003 0.025641026 3dREML_output+tlrc.BRIK[603]
sub-002 t2002to1998 0.020833333 3dREML_output+tlrc.BRIK[606]
sub-002 t1997to1993 0.015873016 3dREML_output+tlrc.BRIK[609]
sub-002 t1992to1988 0.012820513 3dREML_output+tlrc.BRIK[612]
.....
sub-032 t2017to2015 0.333333333 3dREML_output+tlrc.BRIK[591]
sub-032 t2014to2012 0.083333333 3dREML_output+tlrc.BRIK[594]
sub-032 t2011to2009 0.047619048 3dREML_output+tlrc.BRIK[597]
sub-032 t2008to2006 0.033333333 3dREML_output+tlrc.BRIK[600]
sub-032 t2005to2003 0.025641026 3dREML_output+tlrc.BRIK[603]
sub-032 t2002to1998 0.020833333 3dREML_output+tlrc.BRIK[606]
sub-032 t1997to1993 0.015873016 3dREML_output+tlrc.BRIK[609]
sub-032 t1992to1988 0.012820513 3dREML_output+tlrc.BRIK[612] \

In that case, it might be fine to include all the data, regardless of their sample sizes. The alternative of removing some data can introduce biases. If this is a genuine concern, you could try both approaches and compare the differences.

Is the PF column generated based on a customized function? An alternative could be to consider an adaptive approach, as discussed in this blog post.

Gang