I have some problems regarding the interpretation of statistics after the 2nd level:
First, GLMs for all participants were created (using scrf1 and z-normalization). Contrasts were defined with the reconstruction-option. Some contrasts looked at single activation, eg [0 1 ..], some at comparisons, eg [-1 1...]. On the 2nd level, one-sample t-tests were defined across the GLMs from all participants.
When I now look at the results with "Report Statistics", how can the significant t-tests for contrasts like [0 1],[1 0 1],.. (without a comparison) be interpreted? Does it mean, that for the specified condition/regressor, the SCR amplitude is significantly increased compared to a baseline? Or that the parameter estimate is larger than 0? Or is that equivalent?
Many thanks,
Natalie
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Contrasts that do not add up to zero are essentially tests of the intercept for the specified conditions.
In other words, you test whether in the specified conditions the SA (sympathetic arousal) estimates are significantly different from zero. Physiologically, they should be greater than zero. But as they contain some error, they can also be negative.
If SA estimtes are significantly negative, this often indicates a problem in the model specification. Your model makes assumptions about when to expect phasic SA (for example, after experimental events). If SA systematically (i.e. non-randomly) occurs at other points in time, the model can be invalid, and one may see systematically negative SA estimates.
Hope this helps
Best
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Great, thank you that helps a lot!
Indeed I get some significantly negative estimates for one event... But I can still draw conclusions about all the significantly positive estimates, right?
And what about contrasts that add up to zero that include one of such negative estimates and therefore are significant? I guess it would be not valid to interpret them?
Best,
Natalie
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
given significantly negative SA estimates, I personally would in any case try to find out whether the model is misspecified. For example, are the onsets stated correctly, are perhaps some events longer (> 1 s) and responses occur with a longer latency, or are some events that elicit very large SA (e.g. start of session, announcement of a break) not modelled at all.
Best
Dominik
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Dear Dominik,
I have some problems regarding the interpretation of statistics after the 2nd level:
First, GLMs for all participants were created (using scrf1 and z-normalization). Contrasts were defined with the reconstruction-option. Some contrasts looked at single activation, eg [0 1 ..], some at comparisons, eg [-1 1...]. On the 2nd level, one-sample t-tests were defined across the GLMs from all participants.
When I now look at the results with "Report Statistics", how can the significant t-tests for contrasts like [0 1],[1 0 1],.. (without a comparison) be interpreted? Does it mean, that for the specified condition/regressor, the SCR amplitude is significantly increased compared to a baseline? Or that the parameter estimate is larger than 0? Or is that equivalent?
Many thanks,
Natalie
Hi Natalie
Contrasts that do not add up to zero are essentially tests of the intercept for the specified conditions.
In other words, you test whether in the specified conditions the SA (sympathetic arousal) estimates are significantly different from zero. Physiologically, they should be greater than zero. But as they contain some error, they can also be negative.
If SA estimtes are significantly negative, this often indicates a problem in the model specification. Your model makes assumptions about when to expect phasic SA (for example, after experimental events). If SA systematically (i.e. non-randomly) occurs at other points in time, the model can be invalid, and one may see systematically negative SA estimates.
Hope this helps
Best
Dominik
Great, thank you that helps a lot!
Indeed I get some significantly negative estimates for one event... But I can still draw conclusions about all the significantly positive estimates, right?
And what about contrasts that add up to zero that include one of such negative estimates and therefore are significant? I guess it would be not valid to interpret them?
Best,
Natalie
Hi Natalie
given significantly negative SA estimates, I personally would in any case try to find out whether the model is misspecified. For example, are the onsets stated correctly, are perhaps some events longer (> 1 s) and responses occur with a longer latency, or are some events that elicit very large SA (e.g. start of session, announcement of a break) not modelled at all.
Best
Dominik
Hi Dominik,
okay, thanks a lot for your advice!
Best,
Natalie