baseline norm via Z transform. std across trials or time or both?

[Demetris Roumis]

Thank you for writing such an accessible and helpful book, I constantly refer back to it and recommend it to my colleagues.

When doing a baseline normalization via a Z transform, the mean and std of the baseline period is needed.

To get the std of the baseline period, should I compute the mean over trials (for every time bin) and then the std over time?

Chapter 18 code makes it look like this is the case because the std is only performed once.

I'm guessing this might be ok because the post event period is also averaged across trials. Could you confirm?

Also, I would like to use the median instead of the mean to reduce the influence of large outliers/noise, but I don't see anything in the book or code that does a Z transform using the median.

If I wanted to do this, I should use the median of the baseline period instead of the mean, correct? Is this still statistically sound?

Thank you so much!

-Demetris

[Mike X Cohen]

Hi Demetris. I generally don't recommend using z-score for normalization of time-frequency power. You can run into trouble if the pre-stim variability is very low, which can produce uninterpretably large z-values. Also, z-score is very rarely done in the literature. The direct answer to your question is that I suppose you could use either. It's so rarely done that I don't think there is an established convention.

That said, if you want to use the median instead of the mean, you should probably use the median absolute difference instead of the standard deviation. I didn't write about that in the book, but there is a wikipedia page about it.

Mike

--
You received this message because you are subscribed to the Google Groups "AnalyzingNeuralTimeSeriesData" group.
To unsubscribe from this group and stop receiving emails from it, send an email to analyzingneuraltimes...@googlegroups.com.
Visit this group at https://groups.google.com/group/analyzingneuraltimeseriesdata.
To view this discussion on the web visit https://groups.google.com/d/msgid/analyzingneuraltimeseriesdata/500044fb-5100-4c11-8bd5-fca5ad89a1c6%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Mike X Cohen, PhD
Fresh look: mikexcohen.com

[Demetris Roumis]

Thanks for the quick and helpful response. I'll go with something more convention (decibel conversion + time shift permutation testing with pixel based mc correction) since I don't have a good reason to do the z score aside from simplicity.

Thanks!

-Demetris

On Thursday, July 4, 2019 at 1:36:56 PM UTC-7, Mike X Cohen wrote:

Hi Demetris. I generally don't recommend using z-score for normalization of time-frequency power. You can run into trouble if the pre-stim variability is very low, which can produce uninterpretably large z-values. Also, z-score is very rarely done in the literature. The direct answer to your question is that I suppose you could use either. It's so rarely done that I don't think there is an established convention.

That said, if you want to use the median instead of the mean, you should probably use the median absolute difference instead of the standard deviation. I didn't write about that in the book, but there is a wikipedia page about it.

Mike

On Thu, Jul 4, 2019 at 9:33 PM Demetris Roumis <rou...@gmail.com> wrote:

Thank you for writing such an accessible and helpful book, I constantly refer back to it and recommend it to my colleagues.

When doing a baseline normalization via a Z transform, the mean and std of the baseline period is needed.

To get the std of the baseline period, should I compute the mean over trials (for every time bin) and then the std over time?

Chapter 18 code makes it look like this is the case because the std is only performed once.

I'm guessing this might be ok because the post event period is also averaged across trials. Could you confirm?

Also, I would like to use the median instead of the mean to reduce the influence of large outliers/noise, but I don't see anything in the book or code that does a Z transform using the median.

If I wanted to do this, I should use the median of the baseline period instead of the mean, correct? Is this still statistically sound?

Thank you so much!

-Demetris

--
You received this message because you are subscribed to the Google Groups "AnalyzingNeuralTimeSeriesData" group.

To unsubscribe from this group and stop receiving emails from it, send an email to analyzingneuraltimeseriesdata+unsub...@googlegroups.com.

Visit this group at https://groups.google.com/group/analyzingneuraltimeseriesdata.
To view this discussion on the web visit https://groups.google.com/d/msgid/analyzingneuraltimeseriesdata/500044fb-5100-4c11-8bd5-fca5ad89a1c6%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.