This researcher reviewed methods for removal of outliers:

[email protected] or [email protected]

Just to give an update: I'm currently investigating the use of median absolute deviation (MAD) and it seems to work rather well. Using a MAD of 7 is "far out there" and catches the extreme values.

However, 7 MAD is an arbitrary choice and arguably a large value (assume normal distribution, multiply by 1.4826). I greatly appreciate any thoughts on this!

Thank you.

My previous answer to this question seems to be deleted. There is about hundreds of methods to manage outliers we were once taught. Review of differents methods are available on the net. One method to applys is the so called "trimmed mean" that is:

"A trimmed mean is calculated by discarding a certain percentage of the lowest and the highest scores and then computing the mean of the remaining scores. The trimmed mean has the advantage of being relatively resistant to outliers. When outliers are present in the data, trimmed means are robust estimators of the population mean that are relatively insensitive to the outlying values. After viewing the box plot, a potential outlier might be identified. If the upper and lower 5% of the data are removed, then it creates a 10% trimmed mean. An example of trimmed means is the recent change in the Olympic scoring system for ice skating, in which the highest and lowest scores are eliminated, and the mean of the remaining scores is used to assess skaters' scores. If a trimmed mean is presented, the untrimmed mean should be presented for comparison."

Dear Nic,

The simple descriptive method of outlier removel is Inter Quartile Range (Q3-Q1) or Inter Decile Range (D9-D1) whereby extreme values are removed from the data set. 

Outliers are, however the part of data set. It is better not to exclude them but follow some suitable transformation like log, sin, square, root, reciprocal etc. Transformations are allowed in econometrics/ research.

Hope it helps.

I am not familiar with your particular application, but if you have trouble with outliers in your data, you might consider employing LAD (Least Absolute Deviations) rather than OLS (Ordinary Least Squares). LAD is much more robust to outliers. It tries to find the median values experssed in the covariates, rather than mean values.

Dear Nic,

The simple descriptive method of outlier removal is Inter Quartile Range (Q3-Q1) or Inter Decile Range (D9-D1) that is how you can avoid extreme observations. Otherwise, you can also follow Median ± k(sd) in order to eliminate outliers.

Outliers are, however the part of data so its better not to eleminate them but follow some of the suitable transformations like log, sin, square, difference, root, reciprocal etc. Transformations are widely used in research to align the data.

Hope it helps. 

Hello everyone,

Thank you very much for your suggestions as they are all very helpful! Dealing with these outliers is the last piece to finalizing my analysis pipeline and so I'm rather grateful!

I am currently looking into median absolute deviation, will do so with LAD, and had thought about using quartiles as well. I'm not sure about transformations in this case. Trimmed mean is new to me and sounds interesting. Permutation testing of the values pre/post division was recommended on another message board and is a different approach, keeping outliers in the analysis :)

However, the extreme outliers are due to division with a very small denominator and are mostly a mathematical issue; they also don't make sense (i.e., are impossible) in context of my data. So removing them should not be as problematic from that point of view, I believe.

Thanks again!

Nic

PS: I also tried to post an update as an answer yesterday but it didn't show up.

I have had many challenges with head movement given that I study orgasm.  I found the FSL_motion_outliers did a very good job on my data and can be used in conjunction with other motion parameters. 

http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLMotionOutliers

I just saw your message - thank you!

Your research sounds very interesting in many ways and I am curious about the associated motion problems. Thank you for pointing me to your work!

This thread is extremely helpful, as I am also struggling with developing a principled way to remove outliers in fMRI data. Nic, were you by chance performing beta series analysis to obtain trial-wise estimates? I used "stim_times_IM" in AFNI's 3dDeconvolve, and have obtained a handful of trials per subject that are several orders of magnitude higher than the rest of the estimates across the sample. I had thought this had to do with overfitting, but would love some further insights into these extreme values if you have any further information about this. Also: does anyone here have experience with how to perform some of these approaches in AFNI? It looks like 3dToutcount might be the way to go if we're going to percentile rank the data.

Cheers, 

Jeff

Hi Jeffrey,

Great to hear that this thread is helpful beyond my problems :)

I did not perform beta series analysis (what exactly do you mean by that?) but voxel-wise divided the beta from task A by the beta from task B for each participant. Because some task B betas were really small, some of the results were (hilariously?!) large. There were a few things I tried at the preprocessing level (deconvolution, grey matter masks, ...) to address these small task B betas but still ended up with some participant's having way too large values.

I decided to use non-parametric tests at the group level (3dMannWhitney and 3dWilcoxon). Thus, I didn't cut voxels out. If you consider finding a rule to exclude voxels/trials/participants, you could use median absolute deviation instead of the standard deviation or a trimmed mean or quartile rule. These may be especially useful if you only have a few strong outliers (i.e., you don't bias/cut too much of the data). Perhaps bootstrapping helps? Feel free to send me an email if you'd like to further discuss: [email protected].

3dLocalstat may help you with percentile ranks, too.

Enjoy the process ;)

Nic

Hi Fernando,

Your approach sounds very interesting - could you tell a bit more?

Thanks a lot in advance!

Nic

Thanks Nic for your response! Turns out AFNI's 3dToutcount uses median absolute deviation, so I think we're on the right track. By beta series analysis, I am just referring to doing the GLM with a single predictor for each trial to get out trial-wise beta estimates. You then get out a "beta series" for each voxel, which can be used for functional connectivity analyses. In our analyses we are calculating the standard deviation of these beta estimates to get a metric of response stability; obviously though, the standard deviations are highly influenced by the presence of extreme values.

Perhaps computing a coefficient of variation to get a standardized measure of variability may also be worth investigating. I'm using MAD/median (instead of SD/mean) currently but there's a non-parametric quartile approach, too. 

What is your task? Is it event-related? You mentioned you are using stim_times_IM - have you compared that to other models? I've only ever used the "block" and the "tent" models.