ccao-data · Damonamajor · Sep 16, 2025 · Sep 3, 2025 · Sep 3, 2025 · Sep 3, 2025
@@ -178,7 +178,7 @@ prb <- function(assessed, sale_price, na.rm = FALSE) {
 # Calculate the Gini cofficients needed for KI and MKI
 calc_gini <- function(assessed, sale_price) {
   df <- data.frame(av = assessed, sp = sale_price)
-  df <- df[order(df$sp), ]
+  df <- df[order(df$sp, -df$av), ]
-  df <- df[order(df$sp, -df$av), ]
+  # This Gini coefficient algorithm is sensitive to the order of the input
+  # observations: If multiple observations share the same sale price but have
+  # different estimates, the output coefficients will be different depending
+  # on which of the sales with identical prices gets ordered first in the
+  # input dataframe. To ensure a stable sort order, Quintos recommends
+  # sorting by ascending sale price and then by descending estimate to break
+  # any ties. This produces "worst case" MKI/KI statistics, but ensures those
+  # statistics are deterministic. See this issue for more discussion:
+  # https://github.com/ccao-data/assesspy/issues/33#issuecomment-3180632954
+  df <- df[order(df$sp, -df$av), ]
-  df <- df[order(df$sp, -df$av), ]
+  # This Gini coefficient algorithm is sensitive to the order of the input
+  # observations: If multiple observations share the same sale price but have
+  # different estimates, the output coefficients will be different depending
+  # on which of the sales with identical prices gets ordered first in the
+  # input dataframe. To ensure a stable sort order, Quintos recommends
+  # sorting by ascending sale price and then by descending estimate to break
+  # any ties. This produces "worst case" MKI/KI statistics, but ensures those
+  # statistics are deterministic. See this issue for more discussion:
+  # https://github.com/ccao-data/assesspy/issues/33#issuecomment-3180632954
+  df <- df[order(df$sp, -df$av), ]
   assessed_price <- df$av
   sale_price <- df$sp
   n <- length(assessed_price)

@@ -0,0 +1,31 @@
+"Sale_Price","Assessed","Assessed_alt_sort_1","Assessed_alt_sort_2"
+32900,37299.37125,37299.37125,37299.37125
+36000,40165.89269,40165.89269,40165.89269
+54000,56317.4201,56317.4201,56317.4201
+64500,66183.77244,66183.77244,66183.77244
+68000,69486.97316,69486.97316,69486.97316
+70000,71514.52586,71514.52586,71514.52586
+74000,75338.28603,75338.28603,75338.28603
+80000,81035.95111,81035.95111,81035.95111
+84900,85672.85577,85672.85577,85672.85577
+89000,85021.0865,94088.93683,90046.33945
+89000,90046.33945,85021.0865,94088.93683
+89000,94088.93683,90046.33945,85021.0865
+105900,100227.0936,100227.0936,100227.0936
+109000,103156.7516,103156.7516,103156.7516
+115000,108290.1277,108290.1277,108290.1277
+124500,117098.7563,117098.7563,117098.7563
+129900,115346.9796,115346.9796,115346.9796
+135000,119678.4223,119678.4223,119678.4223
+149000,131630.9478,131630.9478,131630.9478
+155800,137321.2061,137321.2061,137321.2061
+163500,143973.5639,143973.5639,143973.5639
+175000,153571.8563,153571.8563,153571.8563
+179000,148456.8866,148456.8866,148456.8866
+185600,153488.3876,153488.3876,153488.3876
+199900,165039.8271,165039.8271,165039.8271
+215000,176939.5763,176939.5763,176939.5763
+235000,192959.3127,192959.3127,192959.3127
+250000,180046.1193,180046.1193,180046.1193
+279000,200240.2442,200240.2442,200240.2442
+295000,211445.4891,211445.4891,211445.4891
@@ -16,6 +16,15 @@ mki_ki_data <- read.csv(
 mki_ki_assessed <- mki_ki_data$Assessed
 mki_ki_sale_price <- mki_ki_data$Sale_Price
 
+mki_tiebreaks <- read.csv(
+  rprojroot::find_testthat_root_file("data/mki_ki_data_with_tiebreaks.csv")
+)
+
+tiebreaks_Sale_Price <- mki_tiebreaks$Sale_Price
+tiebreaks_Assessed <- mki_tiebreaks$Assessed
+tiebreaks_Assessed_alt_sort_1 <- mki_tiebreaks$Assessed_alt_sort_1
+tiebreaks_Assessed_alt_sort_2 <- mki_tiebreaks$Assessed_alt_sort_2
+
 
 
 ##### TEST COD #####
@@ -184,6 +193,22 @@ test_that("standard met function", {
   expect_false(mki_met(mki_out))
 })
 
+test_that("all sorting variations return the same MKI (tiebreak data)", {
+  mki_out_assessed <- mki(
+    tiebreaks_Assessed, tiebreaks_Sale_Price
+  )
+  mki_out_assessed_alt_sort1 <- mki(
+    tiebreaks_Assessed_alt_sort_1,
+    tiebreaks_Sale_Price
+  )
+  mki_out_assessed_alt_sort2 <- mki(
+    tiebreaks_Assessed_alt_sort_2,
+    tiebreaks_Sale_Price
+  )
+
+  expect_equal(mki_out_assessed, mki_out_assessed_alt_sort1)
+  expect_equal(mki_out_assessed, mki_out_assessed_alt_sort2)
+})
 
 
 ##### TEST KI #####
@@ -224,6 +249,24 @@ test_that("incomplete data returns NAs unless removed", {
   )
 })
 
+test_that("all sorting variations return return the same ki (tiebreak data)", {
+  ki_out_assessed <- ki(
+    tiebreaks_Assessed,
+    tiebreaks_Sale_Price
+  )
+  ki_out_assessed_alt_sort1 <- ki(
+    tiebreaks_Assessed_alt_sort_1,
+    tiebreaks_Sale_Price
+  )
+  ki_out_assessed_alt_sort2 <- ki(
+    tiebreaks_Assessed_alt_sort_2,
+    tiebreaks_Sale_Price
+  )
+
+  expect_equal(ki_out_assessed, ki_out_assessed_alt_sort1)
+  expect_equal(ki_out_assessed, ki_out_assessed_alt_sort2)
+})
+
 
 ##### TEST Median Ratio #####
 context("test median_ratio_met function")