Hi,
I'm familiarizing myself with the R and JAVA GUI of iMRMC, and I have a few questions.
The data is binary with only 1 modality,
-
I generated data with a different number of readers, different number of cases and increasing true probability to get a feel for the behavior (which is as expected). I generated data under assumption of independence for simplicity, but I then analyze the data as if it came from a MRMC set-up with its dependencies. I have noticed that in some cases there are warning messages (e.g. that MLE should be used or messages about the degrees of freedom). However, in some situations, neither the GUI nor R provide any error/warning and it seems like it is stuck in a loop. In R, I had to kill the R session. I attach an example of such a case. I can’t understand what in this situation, creates the problem. Any idea?
Moreover, in R I did a simulation when data was generated under same conditions/assumptions and repeating 1000 times in a loop. At some point, it seems that the same problem occurs and R gets stuck in a loop. Since I don’t understand when or why it happens I don’t know how to check the data before I call the function.
-
With regard to the trick of adding fake data in order to use the tool for binary data: For one scenario I added 3 fake subjects and then 5 fake subjects. The results are identical. I see why the point estimate is not impacted by adding fake subjects. However, I don’t have an intuition why it does not impact the SE and hence the CI or p-value. Can you provide an intuition or refer me to a paper/presentation that discuss this?
-
In the GUI tool, the bottom part refers to the sizing of the study:
a. Is paired referred to the case of 2 modalities? If not, then what it is? If yes, what should I use if having only 1 modality?
b. How is the effect size defined for binary data? The default in the GUI is HO of AUC=0.5 or equivalent for P=0.5. I was not sure if the effect size is defined as the difference between P under the null and P under HO or differently. If this is how it is defined, then why it does not needed to give also the null? I see why this is not needed for AUC, but for binary, the sizing is not the same if the difference is 10% and P0 is 1% or 50%.
CO.Ind.P0.9.2R.40C.xlsx
Thanks,
Anat
Hi,
I'm familiarizing myself with the R and JAVA GUI of iMRMC, and I have a few questions.
The data is binary with only 1 modality,
I generated data with a different number of readers, different number of cases and increasing true probability to get a feel for the behavior (which is as expected). I generated data under assumption of independence for simplicity, but I then analyze the data as if it came from a MRMC set-up with its dependencies. I have noticed that in some cases there are warning messages (e.g. that MLE should be used or messages about the degrees of freedom). However, in some situations, neither the GUI nor R provide any error/warning and it seems like it is stuck in a loop. In R, I had to kill the R session. I attach an example of such a case. I can’t understand what in this situation, creates the problem. Any idea?
Moreover, in R I did a simulation when data was generated under same conditions/assumptions and repeating 1000 times in a loop. At some point, it seems that the same problem occurs and R gets stuck in a loop. Since I don’t understand when or why it happens I don’t know how to check the data before I call the function.
With regard to the trick of adding fake data in order to use the tool for binary data: For one scenario I added 3 fake subjects and then 5 fake subjects. The results are identical. I see why the point estimate is not impacted by adding fake subjects. However, I don’t have an intuition why it does not impact the SE and hence the CI or p-value. Can you provide an intuition or refer me to a paper/presentation that discuss this?
In the GUI tool, the bottom part refers to the sizing of the study:
a. Is paired referred to the case of 2 modalities? If not, then what it is? If yes, what should I use if having only 1 modality?
b. How is the effect size defined for binary data? The default in the GUI is HO of AUC=0.5 or equivalent for P=0.5. I was not sure if the effect size is defined as the difference between P under the null and P under HO or differently. If this is how it is defined, then why it does not needed to give also the null? I see why this is not needed for AUC, but for binary, the sizing is not the same if the difference is 10% and P0 is 1% or 50%.
CO.Ind.P0.9.2R.40C.xlsx
Thanks,
Anat