Update

Predicted changed into predictive,
confusion matrix inverted,
disparate impact is pp and not dp,
missing = error for eo

R/fairness_selection.R

@@ -232,7 +232,7 @@ fairness_selection <- function(q1 = NULL,
     }
   } else if (q1 == 2) {
     name <- "Disparate Impact"
-    measure <- "dp"
+    measure <- "pp"
     q1_name <- "No"
   }
   output <- list()

R/model_fairness.R

@@ -24,7 +24,7 @@
 #' quantify potential fairness or discrimination in the algorithms predictions.
 #' Available parity metrics include predictive rate parity, proportional parity,
 #' accuracy parity, false negative rate parity, false positive rate parity, true
-#' positive rate parity, negative predicted value parity, specificity parity,
+#' positive rate parity, negative predictive value parity, specificity parity,
 #' and demographic parity. The function returns an object of class
 #' \code{jfaFairness} that can be used with associated \code{summary()} and
 #' \code{plot()} methods.
@@ -98,11 +98,11 @@
 #'     \item{False positive rate parity (\code{fprp}): calculated as FP / (TN
 #'       + FP), quantifies whether the false positive rate is the same across
 #'       groups.}
-#'     \item{True positive rate parity (\code{tprp}): calculated as TP / (TP +
+#'     \item{True positive rate parity (\code{tprp}, also known as Equal opportunity): calculated as TP / (TP +
 #'       FN), quantifies whether the true positive rate is the same across
 #'       groups.}
-#'     \item{Negative predicted value parity (\code{npvp}): calculated as TN /
-#'       (TN + FN), quantifies whether the negative predicted value is equal
+#'     \item{Negative predictive value parity (\code{npvp}): calculated as TN /
+#'       (TN + FN), quantifies whether the negative predictive value is equal
 #'       across groups.}
 #'     \item{Specificity parity (\code{sp}): calculated as TN / (TN + FP),
 #'       quantifies whether the true positive rate is the same across groups.}
@@ -253,11 +253,11 @@ model_fairness <- function(data,
     group <- levels(data[, protected])[i]
     groupDat <- data[data[, protected] == group, ]
     # Confusion matrices for each group
-    confmat[[group]][["matrix"]] <- table("Actual" = groupDat[, target], "Predicted" = groupDat[, predictions])
+    confmat[[group]][["matrix"]] <- table("Predicted" = groupDat[, predictions], "Actual" = groupDat[, target])
     confmat[[group]][["tp"]] <- tp <- confmat[[group]][["matrix"]][positive, positive]
-    confmat[[group]][["fp"]] <- fp <- sum(confmat[[group]][["matrix"]][negative, positive])
+    confmat[[group]][["fp"]] <- fp <- sum(confmat[[group]][["matrix"]][positive, negative])
     confmat[[group]][["tn"]] <- tn <- sum(confmat[[group]][["matrix"]][negative, negative])
-    confmat[[group]][["fn"]] <- fn <- sum(confmat[[group]][["matrix"]][positive, negative])
+    confmat[[group]][["fn"]] <- fn <- sum(confmat[[group]][["matrix"]][negative, positive])
     confmat[[group]][["n"]] <- sum(confmat[[group]][["matrix"]])
     # Performance measures for each group
     performance[[group]][["support"]] <- performance[["all"]][i, 1] <- sum(confmat[[group]][["matrix"]])

vignettes/articles/model-fairness.Rmd

@@ -28,7 +28,7 @@ algorithmic decision-making systems. It computes various model-agnostic metrics
 based on the observed and predicted labels in a dataset. The fairness metrics
 that can be calculated include demographic parity, proportional parity,
 predictive rate parity, accuracy parity, false negative rate parity, false
-positive rate parity, true positive rate parity, negative predicted value
+positive rate parity, true positive rate parity, negative predictive value
 parity, and specificity parity [@calders_2010; @chouldechova_2017;
 @feldman_2015; @zafar_2017; @friedler_2019]. Furthermore, the metrics are tested
 for equality between protected groups in the data.
@@ -346,12 +346,12 @@ model_fairness(
       privileged group (Caucasians). This indicates relatively fair treatment
       in the algorithm's true positive predictions.
 
-8. **Negative predicted value parity**: Compares the negative predicted value
+8. **Negative predictive value parity**: Compares the negative predictive value
   (e.g., for non-reoffenders) of each unprivileged group with that of the
   privileged group.
 
-    The formula for the negative predicted value is $NPV = \frac{TN}{TN + FN}$,
-      and the negative predicted value parity for unprivileged group $i$ is
+    The formula for the negative predictive value is $NPV = \frac{TN}{TN + FN}$,
+      and the negative predictive value parity for unprivileged group $i$ is
       given by $NPVP = \frac{NPV_{i}}{NPV_{privileged}}$.
 
     ```{r}
@@ -367,17 +367,17 @@ model_fairness(
     ```
 
     ***Interpretation:***
-    - *African American*: The negative predicted value parity for African
+    - *African American*: The negative predictive value parity for African
       Americans is 0.98013. A value close to 1 indicates that the negative
       predicted value for African Americans is very similar to the privileged
       group (Caucasians). This suggests fair treatment in predicting
       non-reoffenders among African Americans.
-    - *Asian*: The negative predicted value parity for Asians is 1.1163,
-      indicating that their negative predicted value is slightly higher than for
+    - *Asian*: The negative predictive value parity for Asians is 1.1163,
+      indicating that their negative predictive value is slightly higher than for
       Caucasians. This could suggest potential favoritism towards Asians in
       predicting non-reoffenders.
-    - *Hispanic*: The negative predicted value parity for Hispanics is 1.0326,
-      suggesting that their negative predicted value is slightly higher than for
+    - *Hispanic*: The negative predictive value parity for Hispanics is 1.0326,
+      suggesting that their negative predictive value is slightly higher than for
       Caucasians. This indicates potential favoritism towards Hispanics in
       predicting non-reoffenders.
 
@@ -472,12 +472,12 @@ The `fairness_selection()` function offers a method to select a fairness measure
 tailored to a specific context and dataset by answering the questions in the 
 developed decision-making workflow. The fairness measure that can be selected 
 include disparate impact, equalized odds, false positive rate parity, false 
-negative rate parity, predictive rate parity, equal opportunity, specificity 
-parity, negative predictive value parity, accuracy parity [@castelnovo_2022;
-@feldman_2015; @friedler_2019; @hardt_2016, @verma_2018]. After answering the 
-questions in the decision-making workflow and selecting the fairness measure to 
-apply, a graphical representation of the followed path can be created based 
-on the responses.
+negative rate parity, predictive rate parity, equal opportunity (also known as true positive 
+rate parity), specificity parity, negative predictive rate parity, accuracy parity 
+[@castelnovo_2022; @feldman_2015; @friedler_2019; @hardt_2016, @verma_2018]. 
+After answering the questions in the decision-making workflow and selecting the fairness 
+measure to apply, a graphical representation of the followed path can be 
+created based on the responses.
 
 As mentioned earlier, not all fairness measures are equally suitable for every 
 audit situation. For this reason, the `fairness_selection()` function we propose