-
Notifications
You must be signed in to change notification settings - Fork 31
FAQ Frequently asked questions
This is not a bug!
Most statistical tools cannot plot specificity from 1 to 0, so instead they plot 1-specificity from 0 to 1. This results in exactly the same plot, just with a different labelling of the axis.
R doesn't have this limitation, and pROC
plots specificity from 1 to 0, saving you from doing a few subtractions when you look at the plot. If you are using pROC in S+, you only have the choice to plot 1-specificity.
If you are concerned that your reviewers will be confused and really wish to have the "old-style" 1-Specificity from 0 to 1, or if you plan to change the axis label to "False Positive Rate" (which is really 1-specificity), you can use the legacy.axis
argument to the plot function. See ?plot.roc
for more details.
You need to use the predict()
function to get predictions that can be passed to pROC (as the predictor
argument).
Not directly, but you can! You need to use a modeling function (such as glm()
) or package (such as lme4); fit and check the model, then use the predict()
function to get predictions that can be passed to pROC (as the predictor
argument).
A large effort has been made to render pROC more efficient with large ROC curves. An algorithm with complexity of O(n) has been rolled out with pROC 1.6 and can be used with algorithm=2
argument to roc
system.time(roc(round(runif(1E6)), rnorm(1E6), algorithm=2))
This only takes about 20 seconds on my laptop machine. The DeLong test has also been made more efficient and the memory footprint was removed.
If your dataset is much larger, it might help to use the old algorithm that goes in O(n**t*), where t is the number of thresholds that can be measured in your data. You can reduce the number of thresholds by rounding your predictor so that if you keep about a 1000 unique predictor values you still get a good approximation of your ROC curve. Use algorithm=3
for a faster C++ implementation with Rcpp.
system.time(roc(round(runif(1E6)), round(rnorm(1E6), digits=1), algorithm=3))
You can play around with algorithm=0
that runs a microbenchmark and chooses the fastest algorithm for your data.
Yes you can since pROC 1.6 and its ci.coords
function. For instance:
> ci.coords(aSAH$outcome, aSAH$s100b, x="best", input = "threshold", ret="threshold")
95% CI (2000 stratified bootstrap replicates):
2.5% 50% 97.5%
threshold 0.115 0.205 0.4854
There is 95% chance that the optimal threshold lies between 0.115 and 0.4854.
Several packages on CRAN provide alternative roc
or auc
functions. These packages can interfere with pROC, especially if they are loaded later in the session (and hence appear earlier in the search path).
For instance, here are a few messages you may see if you have the AUC
package loaded:
Not enough distinct predictions to compute area under the ROC curve.
Error in roc(outcome ~ ndka, data = aSAH) : unused argument (data = aSAH)
If that happens, you should unload the package by typing detach("package:AUC")
. Alternatively you can refer to the pROC version of the function specifically through their namespace:
pROC::auc(pROC::roc(aSAH$outcome, aSAH$ndka))
Here is a list of packages providing roc
and auc
functions, possibly conflicting with pROC:
library(sos)
auc.search = findFn("auc")
auc.search[auc.search$Function == "auc", c("Package", "Function", "Description", "Link")]
roc.search = findFn("roc")
roc.search[roc.search$Function == "roc", c("Package", "Function", "Description", "Link")]
Package | Function | Description and Link |
---|---|---|
analogue | roc | ROC curve analysis |
MAMSE | roc | Receiver Operating Characteristic (ROC) Curves |
spatstat | roc | Receiver Operating Characteristic |
AUC | roc | Compute the receiver operating characteristic (ROC) curve. |
sdm | roc | plot ROC curves |
NEArender | roc | ROC for NEA benchmarks |
fmsb | roc | Calculate Receiver Operating Characteristic (ROC) curve |
epiDisplay | roc | ROC curve |
aucm | roc | ROC and AUC |