Receiver Operating Characteristic ( ROC ) Curves

Assessment of predictive accuracy is a critical aspect of evaluating and comparing models, algorithms or technologies that produce the predictions. In the field of medical diagnosis, receiver operating characteristic (ROC) curves have become the standard tool for this purpose and its use is becoming increasingly common in other fields such as finance, atmospheric science and machine learning. There are surprisingly few built-in options in SAS for ROC curves, but several procedures in SAS/STAT can be tailored with little effort to produce a wide variety of ROC analyses. This talk will focus on the use of SAS/STAT procedures FREQ, LOGISTIC, MIXED and NLMIXED to perform ROC analyses, including estimation of sensitivity and specificity, estimation of an ROC curve and computing the area under the ROC curve. In addition, several macros will be introduced to facilitate graphical presentation and complement existing statistical capabilities of SAS with regard to ROC curves. Real data from clinical applications will be used to demonstrate the methods. INTRODUCTION Receiver operating characteristic (ROC) curves are useful for assessing the accuracy of predictions. Making predictions has become an essential part of every business enterprise and scientific field of inquiry. A simple example that has irreversibly penetrated daily life is the weather forecast. Almost all news sources, including daily newspapers, radio and television news, provide detailed weather forecasts. There is even a dedicated television channel for weather forecasts in United States. Of course, the influence of a weather forecast goes beyond a city dweller’s decision as to pack an umbrella or not. Inclement weather has negative effects on many vital activities such as transportation, agriculture and construction. For this reason collecting data that helps forecast weather conditions and building statistical models to produce forecasts from these data have become major industries. I will give examples from other areas where prediction plays a major role to motivate statisticians from diverse fields of application. Credit scoring is an excellent example: when a potential debtor asks for credit, creditors assess the likelihood of default to decide whether to loan the funds and at what interest rate. An accurate assessment of chance of default for a given debtor plays a crucial role for creditors to stay competitive. For this reason, the prediction models behind credit scoring systems remain proprietary but their predictive power needs to be continuously assessed for the creditors to remain competitive. The final example is concerned with the field of medical diagnostics. The word "prediction" rarely appears in this literature, but a diagnosis is a prediction of what might be wrong with a patient producing the symptoms and the complaints. Most disease processes elicit a response that is manifested in the form of increased levels of a substance in the blood or urine. There might be other reasons for such elevated levels, and blood or urine levels mis-diagnose a condition because of this. The kind of analysis one would perform for weather forecasts is similarly valid for these blood or urine "markers." ROC curves provide a comprehensive and visually attractive way to summarize the accuracy of predictions. They are widely applicable, regardless of the source of predictions. The field of ROC curves is by and large ignored during statistics education and training. Most statisticians learn of ROC curves on the jog, as needed, and struggle through some of the unusual features. To make matters worse for SAS users, very few direct methods are available for performing an ROC analysis although many procedures can be tailored with little attempt to produce ROC curves. There is also a macro available from the SAS Institute for this purpose. The goal of this paper is to summarize the available features in SAS for ROC curves and expand on using other procedures for further analyses. BASIC CONCEPTS: BINARY PREDICTOR One of the simplest scenarios for prediction is the case of a binary predictor. It is important, not only pedagogically because it contains the most important building blocks of an ROC curve, but also practically because it is often encountered practice. I will use an example from weather forecasting to illustrate the concepts and at the end of the section mention some situations from other prominent fields. The article by Thornes and Stephenson (2001) reviews the concepts of assessment of predictive accuracy from the perspective of weather forecast products. Their opening example is very simple and accessible to all data analysts regardless of their training in meteorological sciences. The example relates to frost forecasts produced for M62 1 Statistics and Data Analysis SUGI 31