Hi,

a common practice in connectivity analysis is to identifiy significant connections by computing a null distribution with surrogates. Is this also a standard procedure when using NBS or might there be issues? E.g. because sparse matrices could lead to type I error.

thanks

NBS does indeed estimate an empirical null distribution using permutation testing.

It is unclear why spare matrices per se would be particularly prone to type 1 error. 

Originally posted by jumac:

Hi,

a common practice in connectivity analysis is to identifiy significant connections by computing a null distribution with surrogates. Is this also a standard procedure when using NBS or might there be issues? E.g. because sparse matrices could lead to type I error.

thanks

Hi Andrew

Thank you for your response.

I'd like to clarify my previous query. My concern isn't about sparse matrices per se, but rather the process of pre-thresholding connectivity matrices using surrogate data (for instance, derived from an AR model) before employing NBS. Specifically, if I threshold my matrix based on surrogate data, setting connections that fall within the 95% interval to zero, and then proceed with NBS, could this preprocessing step potentially lead to false predictions or biases in the results? My worry is that, while the surrogate thresholding aims to enhance robustness, it might inadvertently introduce errors or biases.

Best regards

Originally posted by Andrew Zalesky:

NBS does indeed estimate an empirical null distribution using permutation testing.

It is unclear why spare matrices per se would be particularly prone to type 1 error. 

Originally posted by jumac:

Hi,

a common practice in connectivity analysis is to identifiy significant connections by computing a null distribution with surrogates. Is this also a standard procedure when using NBS or might there be issues? E.g. because sparse matrices could lead to type I error.

thanks

 

Hi Jumanc,

I can't see how thresholding a before NBS would introduce biases or errors. Unless the thresholding is performed is dependent on the contrast or effect that you aim to assess using the NBS. For example, if you are testing a difference between two group and you perform thresholding based on a t-test assessing that between-group difference, this would not make sense due to cicularity. 

Thresholding away connections that are close to zero sounds reasonables. In fact the NBS will aoutcomatically exclude any connections that are exactly zero across all subjects/samples. 

I hope that helps, 

Andrew

Originally posted by jumac:

Hi Andrew

Thank you for your response.

I'd like to clarify my previous query. My concern isn't about sparse matrices per se, but rather the process of pre-thresholding connectivity matrices using surrogate data (for instance, derived from an AR model) before employing NBS. Specifically, if I threshold my matrix based on surrogate data, setting connections that fall within the 95% interval to zero, and then proceed with NBS, could this preprocessing step potentially lead to false predictions or biases in the results? My worry is that, while the surrogate thresholding aims to enhance robustness, it might inadvertently introduce errors or biases.

Best regards

 

Originally posted by Andrew Zalesky:

NBS does indeed estimate an empirical null distribution using permutation testing.

It is unclear why spare matrices per se would be particularly prone to type 1 error. 

Originally posted by jumac:

Hi,

a common practice in connectivity analysis is to identifiy significant connections by computing a null distribution with surrogates. Is this also a standard procedure when using NBS or might there be issues? E.g. because sparse matrices could lead to type I error.

thanks