I knew that I can produce that table of coordinate variables but what value should I choose on what basis
Thanks again
Hello everyone
While most of you is in vacation it came a cross my mind to post a question that I am trying to answer
How to identify the cutoff value of your predictor variable on the ROC curve
eg: id I am using the preoperative weight as a predictor for long term weight regain after surgery (dichotomous) and I need to estimate a value after which I can predict that the patient is most likely will not benefit from the proposed surgery at long term.
Hw would I do this
than you in advance
I knew that I can produce that table of coordinate variables but what value should I choose on what basis
Thanks again
I am starting to feel from readings that it is a matter of choice am I correct
the other possible option that I identify the prevalence of success and set the cut off value for a sensitivity related to 1-prevalence of success.
Am I on track or I am loosing it
Hi!
I am not an expert of Logistic Regression, and I have approached it just some months ago.
I found useful this article by Peng-So, Logistic Regression Analysis and Reporting. A primer that is available HERE.
May be that an optimal cut-off point can be derived from the ROC curve. But a useful (and more visually clear, at least to me) tool is the plot of specificity and sensitivity vs. verious cut-off points (see the article's figure 3). More info in the article, p. 48-49.
I do not know many software that can plot such graph; MedCalc (commercial) can.
I have recently discovered that in R (free) a number of packages are available for ROC analysis, but I have not tested them yet: ROCR and pROC seem interesting.
Hope this helps,
Regards
Gm
http://cainarchaeology.weebly.com/
UpDate:
I have just made some experiments with the R packages I was referring to in my earlier reply.
I must confess that I found them a little tricky....Instead, I found the 'Epi' package quite user friendly, at least as far as plotting the ROC curve and "an" optimal cutoff value is concerned. it comes also with a detailed manual (LINK).
If you are familiar with R, you could try this, just to give an example (using the dataset from the pROC package):
I found the output plot (attached) very useful and clean. It also summarizes the Logistic Regression parameters.Code:library(pROC) #load the library data(aSAH) #load the dataset to be used with the Epi package library(Epi) #load the library ROC(form=outcome~s100b, data=aSAH) #compute and plot the ROC curve, using the variable outcome as y, s100b as x
Hope this helps
Regards
Gm
http://cainarchaeology.weebly.com/
You have multiple options for the cutoff. Their are two typical prudent approaches. The first is Youden's Index - this maximizes the area under the curve. However, usually more appropriate is the first approach you listed, which would use your own selected cutoff based of the trade off of false negative and false positives. The importance of this latter more subjective method is that you as an educated clinician can decided if it is more important to have a lower false negative or false positive rate.
An example I give in a lecture, is that a if the outcome may have dire consequences you may want to select the cutoff that errors on the side of more false positives than false negatives. In particular, a screening should ideally tell more people incorrectly that they may be positive for HIV than incorrectly telling them they are negative. Under this rationale, you have follow-up tests for the positives to confirm findings in lieu of telling people they may not be positive when they are and potentially introducing greater risk for virus spread in the population.
Stop cowardice, ban guns!
GM,
Can you confirm what the "ir.eta" represents.
Sidenote, adding the n-value and prevalence to the graph can help the reader perform all superfluous calculations.
Stop cowardice, ban guns!
@hlsmith:Can you confirm what the "ir.eta" represents.
As it appears digging through the 'Epi' package .pdf (p. 95), that value represents the
It can be swithed on/off with the parameter MX in the ROC() command. The same holds for the report of the model summary, shown in the plot when MI=TRUE.“optimal cutpoint” (i.e. where sens+spec is maximal)
See for example:
Regards,Code:ROC(form=outcome~s100b, data=aSAH, MX=TRUE, MI=TRUE)
Gm
http://cainarchaeology.weebly.com/
Tweet |