I realize you have rebranded and are using the PsPM name now, but I thought since this is related to an older version, I should post it here. This is for analysis done using the developer's version of SCRalyze around December. If you'd like me to repost to the PsPM forum, let me know and I'm happy to.
I have 3 questions for you today!
1: Latency: I have always understood "latency" to mean the time from the stimulus to the onset of the resulting SCR. However, I'm not sure how to interpret latency in the context of DCM flexible responses. I understand that it is the latency not necessarily of the SCR following the stimulus, but of the estimated neural response. Is there a formal definition?
2: Dispersion: I would also love a formal definition of dispersion. I understand it intuitively as the "width" of the SCR. But, I'm not sure how to interpret it in the context of neural responses. Is there also a formal definition of dispersion?
3: Normalization: In the documentation you mention that the data is z-transformed, which I have always understood to mean that the mean is subtracted and the resulting value is divided by the standard deviation. However, it appears the transformation is mutliplicative by the number called zfactor. In a previous thread you explain that the resulting values aren't z-values, and that they are in arbitrary units. Can you explain the way that the amplitudes are normalized when the "normalize" section is checked?
Thank you,
Mallory
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
no problem - your last question is specific to the developer's version of SCRalyze so I leave it here.
(1-2) The central neural response is modelled as Gaussian function. Estimated latency is from start of your event window with respect to the centre of that Gaussian. Dispersion is the standard deviation of that Gaussian. There is no easy relation to latency and dispersion of the ensuing SCR but you could send a Gaussian through the ODE and measure the latency and dispersion of the output. For a formal definition you could look at the PsPM manual which has a long and detailed introduction dealing with these model details.
(3) The data should be z-transfomred and the minimum value per session subtracted. But I know there was a weird bug in the developer's version for some time by which the data was multiplied with, rather than divided by, the standard deviation. To check this, you could look at the field dcm.sn{session-number}.y which contains the data that go through the inversion. SD of these data should be 1 - if it isn't, you have unfortunately used the spoilt version of the code. However, that isn't a problem as you can easily adjust your amplitude estimates.
Hope this helps
Best
Dom
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The new PsPM manual is great, especially the tutorial sections. I still have a couple of questions though:
In DCM, we look at intervals of a few seconds, rather than responses to immediate events. I understand that PsPM estimates amplitude, duration, and latency within this window. But aren't there likely several SCRs (and thus SN bursts) within a several second window? Does PsPM (and SCRalyze) report the first response once the window opens? I believe this is correct, but I want to be sure.
Also, regarding the z-transformation. The standard deviation is indeed 1. But I'm still unclear about the procedure that was applied. What was subtracted in order to z-transform these data - do you mean the minimum was subtracted instead of the mean?
Thanks,
Mallory
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
(1) the algorithm tries to fit ONE neural response within the specified window. If there are several, it won't select the first but just fit the overall waveform with ONE response. This is desired to suppress noise in the neural responses. However, if you expect more than one TRUE response (e. g in longer intervals), you can also specify several (even overlapping) windows, and then several responses will be fit.
(2) The data are z-transformed and then the minimum is substracted from the data (because only positive SCR values make sense and the DCM algorithm expects meaningful numbers).
Best
Dom
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
(1) How would you guess how many true responses could occur within the period? Is there a guideline/rule of thumb based on the length of the interval?
(2) Sorry, maybe I'm not explaining my question well. The data can't be z-transformed in the traditional sense of mean subtracted and divided by the standard deviation, because the mean of these isn't 0, right? So I'm wondering what exactly is done. Is a constant added back in after the data are z-transformed?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
ad (1): depends what you're doing. For fear conditioning, a literature search on FIR/SIR/TIR may be helpful - these are the abbreviations for first-second-third interval response. In any case, with a window < ~8 seconds you will probably not be able to distinguish two SCRs so you cannot estimate them at the same time anyway, whatever method you are using. In that case, modelling just one is more plausible.
(2) The data are z-transformed and THEN the minimum is substracted from the data. This is exactly what is done, nothing else. Does that not make sense? The mean will not be 0 after the minimum is subtracted. Btw, the reason for subtracting the minimum is that the model cannot deal with negative values of the SCR.
best
Dom
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi Dom,
I realize you have rebranded and are using the PsPM name now, but I thought since this is related to an older version, I should post it here. This is for analysis done using the developer's version of SCRalyze around December. If you'd like me to repost to the PsPM forum, let me know and I'm happy to.
I have 3 questions for you today!
1: Latency: I have always understood "latency" to mean the time from the stimulus to the onset of the resulting SCR. However, I'm not sure how to interpret latency in the context of DCM flexible responses. I understand that it is the latency not necessarily of the SCR following the stimulus, but of the estimated neural response. Is there a formal definition?
2: Dispersion: I would also love a formal definition of dispersion. I understand it intuitively as the "width" of the SCR. But, I'm not sure how to interpret it in the context of neural responses. Is there also a formal definition of dispersion?
3: Normalization: In the documentation you mention that the data is z-transformed, which I have always understood to mean that the mean is subtracted and the resulting value is divided by the standard deviation. However, it appears the transformation is mutliplicative by the number called zfactor. In a previous thread you explain that the resulting values aren't z-values, and that they are in arbitrary units. Can you explain the way that the amplitudes are normalized when the "normalize" section is checked?
Thank you,
Mallory
Hi Mallory
no problem - your last question is specific to the developer's version of SCRalyze so I leave it here.
(1-2) The central neural response is modelled as Gaussian function. Estimated latency is from start of your event window with respect to the centre of that Gaussian. Dispersion is the standard deviation of that Gaussian. There is no easy relation to latency and dispersion of the ensuing SCR but you could send a Gaussian through the ODE and measure the latency and dispersion of the output. For a formal definition you could look at the PsPM manual which has a long and detailed introduction dealing with these model details.
(3) The data should be z-transfomred and the minimum value per session subtracted. But I know there was a weird bug in the developer's version for some time by which the data was multiplied with, rather than divided by, the standard deviation. To check this, you could look at the field dcm.sn{session-number}.y which contains the data that go through the inversion. SD of these data should be 1 - if it isn't, you have unfortunately used the spoilt version of the code. However, that isn't a problem as you can easily adjust your amplitude estimates.
Hope this helps
Best
Dom
Hi Dom,
The new PsPM manual is great, especially the tutorial sections. I still have a couple of questions though:
In DCM, we look at intervals of a few seconds, rather than responses to immediate events. I understand that PsPM estimates amplitude, duration, and latency within this window. But aren't there likely several SCRs (and thus SN bursts) within a several second window? Does PsPM (and SCRalyze) report the first response once the window opens? I believe this is correct, but I want to be sure.
Also, regarding the z-transformation. The standard deviation is indeed 1. But I'm still unclear about the procedure that was applied. What was subtracted in order to z-transform these data - do you mean the minimum was subtracted instead of the mean?
Thanks,
Mallory
Hi Mallory
(1) the algorithm tries to fit ONE neural response within the specified window. If there are several, it won't select the first but just fit the overall waveform with ONE response. This is desired to suppress noise in the neural responses. However, if you expect more than one TRUE response (e. g in longer intervals), you can also specify several (even overlapping) windows, and then several responses will be fit.
(2) The data are z-transformed and then the minimum is substracted from the data (because only positive SCR values make sense and the DCM algorithm expects meaningful numbers).
Best
Dom
(1) How would you guess how many true responses could occur within the period? Is there a guideline/rule of thumb based on the length of the interval?
(2) Sorry, maybe I'm not explaining my question well. The data can't be z-transformed in the traditional sense of mean subtracted and divided by the standard deviation, because the mean of these isn't 0, right? So I'm wondering what exactly is done. Is a constant added back in after the data are z-transformed?
Hi Mallory
ad (1): depends what you're doing. For fear conditioning, a literature search on FIR/SIR/TIR may be helpful - these are the abbreviations for first-second-third interval response. In any case, with a window < ~8 seconds you will probably not be able to distinguish two SCRs so you cannot estimate them at the same time anyway, whatever method you are using. In that case, modelling just one is more plausible.
(2) The data are z-transformed and THEN the minimum is substracted from the data. This is exactly what is done, nothing else. Does that not make sense? The mean will not be 0 after the minimum is subtracted. Btw, the reason for subtracting the minimum is that the model cannot deal with negative values of the SCR.
best
Dom
Ohhhh ok, I understand now. Sorry for my confusion! Thanks for your answers.
Mallory