Alternatives to 3dttest when missing data

Hello AFNI team,

I'd like to run a paired t-test on my data set (n=24) where several subjects have less than 5 trials for one of the conditions. I'd like to treat those cases as missing data for that condition (cond=forgotten) while still including them in the other condition (cond=remembered).

Would 3dMEMA or 3dLME be most appropriate? Below is what I tried so far.

  1. For 3dMEMA, I created the GLT in my 3dDeconvolve modeling which directly contrasts the two conditions (remembered-forgotten). After looking at the missing_data flags and options I'm unsure how to best account for my situation. For these subjects, they would have data going into the forgotten condition but I'm worried about its reliability.

  2. Rather than doing the contrast within subject I also tried 3dLME to contrast the conditions at the group level. I ran the following model and created a GLT to contrast remembered-forgotten. To account for the missing data I didn't include the forgotten coefficients for those 5 subjects but still included them in the remembered condition in the data table. Does this model seem correct if 3dLME is the way to go?

3dLME -prefix PF \
-model 'Cond' \
-ranEff '~1+Cond' \
-num_glt 1 \
-gltLabel 1 'rem-forg' -gltCode 1 'Cond : 1*remembered -1*forgotten' \
-mask mask.nii.gz \
-dataTable \
Subj Cond InputFile \
s1 remembered input.nii.gz[1]
s2 remembered input.nii.gz[1]
s3 remembered input.nii.gz[1]
s24 remembered input.nii.gz[1]
s1 forgotten input.nii.gz[2]
s2 forgotten input.nii.gz[2]
s3 forgotten input.nii.gz[2]
s19 forgotten input.nii.gz[2]

The program 3dMEMA is designed to consider both the point estimate and its associated uncertainty, which is embedded in the t-statistic value. Essentially, it means that individuals with less reliable estimates due to a small number of trials aren't artificially removed from the analysis but rather receive less weight in the model.

If you prefer to use 3dLME and want to remove effect estimates with fewer trials, I recommend replacing:

-ranEff '~1+Cond' \


-ranEff '~1' \

As an additional note, you might find 3dLMEr more flexible, as discussed in this blog post. 3dLME will be gradually phased out.


great thank you gang! ill give those both a try and will start shifting to 3dLMEr. thanks for the heads up