如何将命中率与误报率数据导入pROC包的roc对象？

Absolutely! The pROC package is fully capable of creating a roc object from your cumulative empirical ROC point data—perfect for fitting curves and calculating partial AUC values in your recognition memory task. Here's a step-by-step guide tailored to your data:

Step 1: Prepare Your Data #

First, store your cumulative false alarm rates (FPR = 1 - Specificity) and hit rates (TPR = Sensitivity) as vectors. Make sure they're ordered by increasing confidence threshold (i.e., FPR and TPR both rise as you lower the confidence cutoff, which matches your provided data):

library(pROC)

# Your cumulative empirical ROC points
fpr <- c(0.05, 0.11, 0.20, 0.28, 0.45)
tpr <- c(0.45, 0.52, 0.57, 0.59, 0.62)

Step 2: Create a ROC Object #

Since you don't have raw participant-level data, you can construct a valid roc object by initializing a dummy object and replacing its core sensitivity/specificity values with your empirical points. This is a clean workaround that works seamlessly with pROC's downstream functions:

# Initialize a dummy ROC object (we'll overwrite its core data next)
roc_obj <- roc(response = c(1, 0), predictor = c(1, 0))

# Replace with your empirical data
roc_obj$sensitivities <- tpr
roc_obj$specificities <- 1 - fpr  # Convert FPR to Specificity
roc_obj$thresholds <- 1:length(tpr)  # Use confidence interval indices as thresholds

# Verify the object (optional)
roc_obj

Step 3: Fit Curves & Calculate Metrics #

Now you can use standard pROC functions to visualize the curve, fit a smoothed version, and compute AUC/partial AUC:

Plot the Empirical ROC Curve & Smoothed Fit #

# Plot empirical points
plot(roc_obj, 
     main = "Empirical ROC Curve (Recognition Memory Task)",
     col = "#2c3e50", 
     lwd = 2,
     print.auc = TRUE)

# Add a smoothed fitted curve (optional, for better visualization)
smooth_roc <- smooth(roc_obj)
lines(smooth_roc, 
      col = "#e74c3c", 
      lwd = 2, 
      lty = 2)

# Add legend
legend("bottomright",
       legend = c("Empirical Points", "Smoothed Fit"),
       col = c("#2c3e50", "#e74c3c"),
       lwd = 2,
       lty = c(1, 2))

Calculate Full & Partial AUC #

# Full AUC
full_auc <- auc(roc_obj)
cat("Full AUC:", round(full_auc, 3), "\n")

# Partial AUC (e.g., FPR range from 0 to 0.3)
partial_auc <- auc(roc_obj, partial.auc = c(0, 0.3))
cat("Partial AUC (FPR 0-0.3):", round(partial_auc, 3), "\n")

Important Notes #

• Order Matters: Ensure your FPR/TPR points are ordered by decreasing confidence threshold (i.e., FPR and TPR increase as you relax the confidence cutoff). Your provided data already follows this pattern, which is correct.
• Raw Data Option: If you have the total number of target words (positive samples) and lure words (negative samples), you can construct a more precise roc object by simulating raw predictor values based on your cumulative rates. For example, if you had 100 targets and 100 lures, you'd create a predictor vector where 45 targets have the highest confidence, 5 lures have the highest confidence, etc. This allows pROC to calculate confidence intervals via bootstrapping.

内容的提问来源于stack exchange，提问作者Jason Finley