Model Performance Assessment Statistics – Concordance: Steps to Calculate

Binary Decision Outcome is one of most common analytics problem being solved in the industry. Some examples where binary predictive models are used..

Some of the common Model Performance Statistics which are used assessing performance of a Binary Predictive Models are:

  • Confusion Matrix
  • ROC Chart
  • Concordance %
  • Gini
  • KS
  • Lift Chart
  • Gain Table and Chart

In this blog, we will show steps to calculate concordance %.  In a binary predictive modeling scenarios, we have two important variables are:  Observed Target Variable and Predicted Outcome Variable.  For example, in a credit risk default scenario, we have created target variable which takes value 1 (if a customer has defaulted) and 0. This is our observed target variable.

Based on the predictive model especially when a logistic regression technique is used, we get predicted probabilities which is Predicted Outcome Variable.

Compare Observed Target Variable values and Predicted Outcome Variable values to assess efficacy of the predictive model.

In concordance, predicted probabilities of Observed Target variable value 1 is compared against the predicted probabilities of Observed Target Variable value 0. If predicted probabilities value for an observation with Observed Target variable 1 is higher than the predicted probabilities value for  an observation with Observed Target variable 0, then it is concordant otherwise discordance. When the predicted probabilities for observed target variable value 1 and 0 is equal, it is called tied.

Steps:

  1. Columns – ID variable, Observed Target variable and Predicted Probabilities (based on a model)
  2. Find pairs – for comparison, we need to find all probable pairs. Candidates with Observed Target Variable 1 are compared with the candidates with Observed Target variable value 0. In this example, Target variable 0 has 9 observation and target variable 1 has 10 observations. So, all comparison pairs are 9*10=90.
  3. We can create a comparison matrix structured as follow.
  4. Categorize pair into Concordant, Discordant or Tied based on Predicted Probabilities.Concordant: If Predicted Probability for Observed Target variable value 1 is more than that of target variable value 0Discordant: If Predicted Probability for Observed Target variable value 0 is more than that of target variable value 1Tied: If Predicted Probability for Observed Target variable value 0 is equal to that of target variable value 1
  5. Calculate % Concordance
  1. # of Pairs: 9*10=90% Concordance:  56/90 => 62%% Discordance:    34/90 => 38%% Tied: 0/90 => 0%

 

Leave a Comment