An Introduction to ROC curves. Mark Whitehorn. Mark Whitehorn

Size: px
Start display at page:

Download "An Introduction to ROC curves. Mark Whitehorn. Mark Whitehorn"

Transcription

1 An Introduction to ROC curves Mark Whitehorn Mark Whitehorn

2 It s all about me Prof. Mark Whitehorn Emeritus Professor of Analytics Computing University of Dundee Consultant Writer (author) m.a.f.whitehorn@dundee.ac.uk Mark Whitehorn

3 It s all about me Teach Masters in: Data Science Part time Distance learning - aimed at existing data professionals Data Engineering Mark Whitehorn

4 Scope Note the word introduction in the title. We are simply going to do a little gentle roc climbing to try to flatten out the learning curve.

5 ROC Curves Receiver Operating Characteristic (ROC) Curves ROC curves have a long and glorious history and are very broadly applicable. Their history explains the name and we will come to that but since this conference is about Machine Learning (ML) we will start there and then go forward to the past.

6 Definitions Augusta Ada King-Noel, Countess of Lovelace Image: Alfred Edward Chalon [Public domain], via Wikimedia Commons

7 Definitions Ada Lovelace The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform. It can follow analysis; but it has no power of anticipating any analytical relations or truths. Its province is to assist us to making available what we are already acquainted with.

8 Definitions Arthur Samuel* (1959) built a computer program that could play draughts better than he could. He defined ML as a Field of study that gives computers the ability to learn without being explicitly programmed. *Worked with Donald Knuth.

9 Where do ROC curves fit in ML? ML is often split into: Unsupervised Learning which doesn t concern us Supervised Learning which does Mark Whitehorn

10 A very brief overview of Machine Learning Supervised learning uses data which can be classified: Good/Bad customer Male/Female ROC curves are used with classification algorithms. Not all classification is binary of course but ROC curves are typically used with binary classifications.* (You can use one vs all classifier to reduce many classifications down to two.) * Volume Under the ROC Surface for Multi-class Problems. Exact Computation and Evaluation of Approximations Ferri. et. al 2003 Mark Whitehorn

11 Data and classification Features, Predictors, Attributes, Dimensions Output, Response Cases, Examples, Instances, Observations CaseID Date of Incident Time of Incident PostCode Value of Items Number of Items Etc Fraud 1 23/01/ :30 DD1 4HN No 2 23/01/ :50 HR2 5ES No 3 23/01/ :45 RD2 5VG 1,230 2 Yes 4 23/01/ :40 DF4 2WS No 5 23/01/ :50 TH7 4RD No 6 23/01/ :20 WE3 5Rf No etc etc etc etc etc etc etc etc

12 Data modelling Training data + algorithm = data model Mark Whitehorn

13 Data modelling Training data + algorithm = data model Data model + test data = evaluation ROC curves are associated with this testing/evaluation phase of this process Mark Whitehorn

14 Data modelling Training data + algorithm = data model Data model + test data = evaluation New data + data model = new information Mark Whitehorn

15 Classifiers The model that we build in this instance is a classifier, so we could say that: Classifiers consistently classify cases into classes and it would certainly ROC if we did so. So, for example, classifying new insurance claims as fraudulent. Note that if a claim is not classified as fraudulent then the assumption is that it is not fraudulent. (This being the nature of binary classification.) Mark Whitehorn

16 Classifiers We usually try many different classification algorithms, end up with multiple models and need to choose the best one. Deciding which is best is harder than it sounds, this is where we use ROC curves. (As an aside, we may also want to combine several models into an ensemble model; but either way we want to be able to estimate the efficiency of each of our models.) Mark Whitehorn

17 Classifier evaluation We are performing binary classification (can be male/female, good/bad) so we need generic terms: Positive Negative In turn this means that we have to make a decision We are trying to find Good customers. Having made that decision, a Good customer is positive; anything that is not a good customer is negative. Mark Whitehorn

18 Classifier evaluation Each incoming case is either Positive or Negative and our ML algorithm will apply a Positive or Negative classification. The classification can be right or wrong; so we can have four states: Classification Mark Whitehorn Actual Female Male Female Not Female

19 Classifier evaluation True Positive (A bit like the True True in Cloud Atlas. But not true much.) Classification Female Not Female Mark Whitehorn Actual Female Male True Positive

20 Classifier evaluation False Positive Classification Mark Whitehorn Actual Female Male Female False Positive Not Female

21 Classifier evaluation False Negative Mark Whitehorn Actual Female Male Female Classification Not Female False Negative

22 Classifier evaluation True Negative Mark Whitehorn Actual Female Male Female Classification Not Female True Negative

23 Classifier evaluation Four possible states: Blue is good Classification Classified Positive CP Classified Negative CN Actual Actual Positive P True Positive TP False Negative FN Actual Negative N False Positive FP True Negative TN Mark Whitehorn

24 Classification models This kind of grid (Confusion Matrix) is used very frequently, for example: Classified Actual Female Not Female Total Female P 1276 TP 1231 FN Male N 738 FP 34 TN Total

25 Model evaluation So, you have a model and it will classify with four results. How good is it? - You going to have to evaluate it. In fact, you may have several so you need to evaluate them all to find out which is the best. Mark Whitehorn

26 Model evaluation You have one or several models so you need to evaluate them all to find out which is the best. One option. Is each model: A. slightly good? B. very good? C. Whizzo? We need a more precise measure.

27 Accuracy Classified Positive CP Classified Negative CN Accuracy (ACC) can be estimated as: Actual Positive P Actual Negative N True Positive TP False Positive FP False Negative FN True Negative TN TP + TN = Total ( )/2014 = 0.96 Classified Female Not Female Total Actual Female P 1276 TP 1231 FN Male N 738 FP 34 TN Total

28 ACC I train my model using training data. I test my model using, unsurprisingly, testing data. I calculate how many of the cases in the test data were correctly identified. My model scores 99%. So, is ACC a good metric? Mark Whitehorn

29 ACC Suppose 1% are Male. And suppose that my model assumes all are Female. My model scores an impressive 99%. Female Not Female Total Female P 99 TP 99 FN 0 99 Male N 1 FP 1 TN 0 1 Total Mark Whitehorn

30 ACC Hmmmm. We need another metric. Happily the good this about metrics is that they are like standards, and the wonderful thing about standards is that there are so many from which to choose. * *Attributed to Grace Hopper but also to others. Mark Whitehorn

31 There are many such metrics: Actual Positive P Actual Negative N Classified Positive CP True Positive TP False Positive FP Classified Negative CN False Negative FN True Negative TN Test Accuracy (ACC) True Negative Rate - Specificity (SPC) True Positive Rate (TPR) - Sensitivity or Recall False Positive Rate (FPR) - Fall-out False Negative Rate (FNR) Miss rate Positive Predictive Value (PPV) - Precision Negative Predictive Value (NPV) Formula (TP + TN)/Total TN/N = TN/(FP + TN) TP/P = TP/(TP + FN) FP/N = FP/(FP + TN) FN/P = FN/(TP+FN) = 1 - TPR TP/(TP + FP) TN/(TN + FN)

32 And we can start building more complex ones: Matthews Correlation Coefficient (MCC) = TP * TN FP *FN Sqrt((TP + FP)(TP + FN)(TN + FP)(TN+FN))

33 Good news everybody!! ROC curves only use two. Actual Positive P Actual Negative N Classified Positive CP True Positive TP False Positive FP Classified Negative CN False Negative FN True Negative TN Test Formula True Positive Rate (TPR) - Sensitivity or Recall TP/P = TP/(TP + FN) The fraction of positive examples that are correctly classified. How many of the positives do we get right? False Positive Rate (FPR) - Fall-out FP/N = FP/(FP + TN) The fraction of negative examples that are incorrectly classified. How many of the negatives do we get wrong? * Note that it is possible to get Division by zero errors if P or N is zero

34 Base Line We plot the FPR against TPR and produce ROC space. It has a base line as shown. All classification should be will be above this line. Why?

35 Base Line Well, what happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 50 TPR = 50

36 Base Line What happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 50 TPR = 50

37 Base Line What happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 50 TPR = 50

38 Base Line What happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 50 TPR = 50

39 Base Line What happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 50 TPR = 50

40 Base Line What happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 50 TPR = 50

41 Base Line What happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 50 TPR = 50

42 Base Line What happens if we simply guess? Random Random Positive Negative Prediction Prediction Total Actual Positive Actual Negative Total FPR = 80 TPR = 80

43 Base Line So anything above the line is doing better than average.

44 Base Line Anything below the line?

45 Base Line We simply reverse the classification.

46 Base Line Earlier we had a problem using ACC. Reality was producing: Positive 99% Negative 1% We had an algorithm that predicted all Positive which scores 99% on ACC. How does ROC space cope?

47 Base Line Well, here it is. Positive Negative Total Prediction Prediction Actual Positive Actual Negative Total FPR = TPR = 100.0

48 Base Line But change it by the merest smidgen: Positive Negative Total Prediction Prediction Actual Positive Actual Negative Total FPR = 0.0 TPR = 100.0

49 Base Line Suppose you try five different models, each of which produces a single point. Positive Negative Total Prediction Prediction Actual Positive Actual Negative Total FPR = 4.6 TPR = 96.5

50 Base Line You can then assess them. Nominally the best one is closest to the top left hand corner. This is not yet a classic ROC curve. For that we need a model that has a moveable threshold.

51 Zombie Apocalypse Zombies are rife in the city. After infection, people don t show symptoms for about 48 hours. We have people in quarantine. We discover that proto-zombies have higher level of a given antibody than the uninfected, so we measure the level of antibodies in the blood samples of these people.

52 Efficiency We measure the Antibody level (Ab) for 4,000 people for whom we know the diagnosis. (If the numbers look very convenient, that is because they are.) Ab Healthy Infected n 2,000 2,000 mean SD 10 10

53 Very approximately: 70% of all values fall within 1 SD of the mean 95% within 2 SD 99% within 3 SD 35% 35% 0.5% 2% 12.5% 12.5% 2% 0.5% Values SD

54 Efficiency So the distributions overlap, the means differ by ten SD and the SD is ten. (Sometimes the world is just too kind.) Healthy Infected Ab level

55 Efficiency As a side issue, do bear in mind that, if we look at the population of untested patients we don t see two overlapping normal distributions because they are additive. Healthy Infected Ab level

56 Efficiency Mixed patients - equal numbers of healthy and infected So what we see is a normal distribution with a greater standard deviation. Healthy Infected Ab level

57 Efficiency But if we know who is healthy or infected (as we do in our training data) then we do see two overlapping normal distributions. Healthy Infected Ab level

58 Efficiency So, if we were trying to classify unknown patients as Healthy (positive) we choose a threshold Ab value. Below that threshold we say the person is healthy, above it we say they are not healthy (a zombie). Healthy Here be living Infected Ab level Here be Living dead

59 Efficiency Here be living Here be Not Living We can set the threshold wherever we want, but every time we move it we change the numbers of TP, FP, TN and FN. Healthy Infected Ab level

60 Efficiency Here be living Here be Not Living Assume we are actively trying to identify the healthy. Healthy is therefore supposed to be classified as positive. Infected is supposed to be classified as negative. Healthy TP FN Ab level

61 Efficiency Here be living Here be Not Living These people are all healthy (Positive). The majority are correctly classified (True Positive). Some are falsely classified as negative (False Negative). Healthy TP FN Ab level

62 Ab level Efficiency Here be living Here be Not Living These people are infected (Negative). Most are correctly classified (True Negative). But some are incorrectly classified as healthy (False Positive). FP TN Infected

63 Efficiency if we decide to set the threshold at 70, clearly almost everyone is classified as not healthy (infected). So, massive numbers of True and False Negatives. But what of FPR and TPR? Healthy Infected Ab level

64 Efficiency The TPR is the fraction of positive examples that are correctly classified. 0.05% of the healthy people are correctly classified, so TPR = 0.05% The FTR is the fraction of negative examples that are incorrectly classified. 0% of the infected people are incorrectly classified so FPR=0%

65 Efficiency 2.5% healthy correctly classified, so TPR = 2.5% 0.5% infected are incorrectly classified, so FTP = 0.05% Healthy Infected Ab level

66 Efficiency TPR =15% FPR = 2.5% Healthy Infected Ab level

67 Efficiency TPR =50% FPR = 15% Healthy Infected Ab level

68 Efficiency Carrying on with collecting TPR and FPR TPR =85% FPR = 50% Healthy Infected Ab level

69 Efficiency TPR =97.5% FPR = 85% Healthy Infected Ab level

70 Efficiency TPR =99.5% FPR = 97.5% Healthy Infected Ab level

71 Efficiency TPR =100% FPR = 99.5% Healthy Infected Ab level

72 Efficiency TPR =100% FPR = 100% Healthy Infected Ab level

73 Efficiency Our model Random Given these figures, we can plot a ROC curve which shows the efficiency of the model.

74 ROC Curves These are very important estimators of the efficiency of a machine learning algorithm. However it is worth noting what ROC curves do and do not measure. They are great, but they have limitations.

75 ROC Curves Note that the different thresholds do not appear on the ROC curve and, from the curve alone, you cannot work out where they were.

76 Efficiency If the curves are further part, what happens to the ROC curve? Healthy Infected Ab level

77

78

79 AUC (Area Under the Curve) Clearly the area under the curve is another useful metric.

80 Does the ROC curve tell us where to put the threshold? It helps a great deal (in other words, No.) Why not? Well, the efficiency of classification is not the only consideration. For example, what about the cost (to the business, to humans) of an FP and an FN?

81 Does it tell us where to put the threshold? In our example, assume that all people judged negative (infected) are caged together and those judged positive (healthy) are released. (Both of these are very bad options but we are in the middle of an apocalypse.) The former decision condemns the FN in the cage to living death, the latter condemns the entire population to the same.

82 Does it tell us where to put the threshold? Clearly I do not envisage a real zombie apocalypse (I wouldn t joke about it if I did) but even in very simple classifications, the impact of an FP can be very different from that of a FN. Identifying a fraudulent transaction (positive). A FN may cost us the value of the transaction, a FP may cost us a valued customer.

83 Does it tell us where to put the threshold? Diagnosing a disease (positive). A false positive may cost a 5 retest. A false negative may kill the patient. So where do we set the threshold?

84 This has been merely an introduction. My advice would be not to forget those other metrics. Oh, but what about the origins? Test Accuracy (ACC) True Negative Rate - Specificity (SPC) True Positive Rate (TPR) - Sensitivity or Recall False Positive Rate (FPR) - Fall-out False Negative Rate (FNR) Miss rate Positive Predictive Value (PPV) - Precision Negative Predictive Value (NPV) Formula (TP + TN)/Total TN/N = TN/(FP + TN) TP/P = TP/(TP + FN) FP/N = FP/(FP + TN) FN/P = FN/(TP+FN) = 1 - TPR TP/(TP + FP) TN/(TN + FN)

85 Origins of ROC Curves Today the BBC has an article by Tim Harford on the origins of radar. ROC curves have their origins in Signal Detection Theory which is essentially about trying to distinguish between noise and not-noise (in other words, signal).

86 Rather confusingly the ROC was also a mythical bird. The ROC was VERY big but this illustration somewhat fails to show the scale of the problem. René Bull ( )

87 The ROC was VERY big

88 Now, suppose you wanted some warning that a ROC was coming. You d probably invent radar.

89 12 Count of Pixels Light Intensity

90 12 Not Blip Blip Count of Pixels Light Intensity To detect a signal, you simply look for anything to the right of the threshold.

91 But in the early days of radar it wasn t like that. The image was much less clear.

92 But in the early days of radar it wasn t like that. The image was much less clear.

93 12 Count of Pixels Light Intensity Does the question Where are we going to put the threshold? sound familiar?

94 ROC curves Thank you for ROCing up, any questions?

INTRODUCTION TO MACHINE LEARNING. Decision tree learning

INTRODUCTION TO MACHINE LEARNING. Decision tree learning INTRODUCTION TO MACHINE LEARNING Decision tree learning Task of classification Automatically assign class to observations with features Observation: vector of features, with a class Automatically assign

More information

Machine learning II. Juhan Ernits ITI8600

Machine learning II. Juhan Ernits ITI8600 Machine learning II Juhan Ernits ITI8600 Hand written digit recognition 64 Example 2: Face recogition Classification, regression or unsupervised? How many classes? Example 2: Face recognition Classification,

More information

METHODS FOR DETECTING CERVICAL CANCER

METHODS FOR DETECTING CERVICAL CANCER Chapter III METHODS FOR DETECTING CERVICAL CANCER 3.1 INTRODUCTION The successful detection of cervical cancer in a variety of tissues has been reported by many researchers and baseline figures for the

More information

Performance Evaluation of Machine Learning Algorithms in the Classification of Parkinson Disease Using Voice Attributes

Performance Evaluation of Machine Learning Algorithms in the Classification of Parkinson Disease Using Voice Attributes Performance Evaluation of Machine Learning Algorithms in the Classification of Parkinson Disease Using Voice Attributes J. Sujatha Research Scholar, Vels University, Assistant Professor, Post Graduate

More information

Receiver operating characteristic

Receiver operating characteristic Receiver operating characteristic From Wikipedia, the free encyclopedia In signal detection theory, a receiver operating characteristic (ROC), or simply ROC curve, is a graphical plot of the sensitivity,

More information

Behavioral Data Mining. Lecture 4 Measurement

Behavioral Data Mining. Lecture 4 Measurement Behavioral Data Mining Lecture 4 Measurement Outline Hypothesis testing Parametric statistical tests Non-parametric tests Precision-Recall plots ROC plots Hardware update Icluster machines are ready for

More information

Various performance measures in Binary classification An Overview of ROC study

Various performance measures in Binary classification An Overview of ROC study Various performance measures in Binary classification An Overview of ROC study Suresh Babu. Nellore Department of Statistics, S.V. University, Tirupati, India E-mail: sureshbabu.nellore@gmail.com Abstract

More information

When Overlapping Unexpectedly Alters the Class Imbalance Effects

When Overlapping Unexpectedly Alters the Class Imbalance Effects When Overlapping Unexpectedly Alters the Class Imbalance Effects V. García 1,2, R.A. Mollineda 2,J.S.Sánchez 2,R.Alejo 1,2, and J.M. Sotoca 2 1 Lab. Reconocimiento de Patrones, Instituto Tecnológico de

More information

Sensitivity, Specificity, and Relatives

Sensitivity, Specificity, and Relatives Sensitivity, Specificity, and Relatives Brani Vidakovic ISyE 6421/ BMED 6700 Vidakovic, B. Se Sp and Relatives January 17, 2017 1 / 26 Overview Today: Vidakovic, B. Se Sp and Relatives January 17, 2017

More information

Predictive Models for Healthcare Analytics

Predictive Models for Healthcare Analytics Predictive Models for Healthcare Analytics A Case on Retrospective Clinical Study Mengling Mornin Feng mfeng@mit.edu mornin@gmail.com 1 Learning Objectives After the lecture, students should be able to:

More information

Reliability, validity, and all that jazz

Reliability, validity, and all that jazz Reliability, validity, and all that jazz Dylan Wiliam King s College London Introduction No measuring instrument is perfect. The most obvious problems relate to reliability. If we use a thermometer to

More information

Reliability, validity, and all that jazz

Reliability, validity, and all that jazz Reliability, validity, and all that jazz Dylan Wiliam King s College London Published in Education 3-13, 29 (3) pp. 17-21 (2001) Introduction No measuring instrument is perfect. If we use a thermometer

More information

Example - Birdkeeping and Lung Cancer - Interpretation. Lecture 20 - Sensitivity, Specificity, and Decisions. What do the numbers not mean...

Example - Birdkeeping and Lung Cancer - Interpretation. Lecture 20 - Sensitivity, Specificity, and Decisions. What do the numbers not mean... Odds Ratios Example - Birdkeeping and Lung Cancer - Interpretation Lecture 20 - Sensitivity, Specificity, and Decisions Sta102 / BME102 Colin Rundel April 16, 2014 Estimate Std. Error z value Pr(> z )

More information

An Improved Patient-Specific Mortality Risk Prediction in ICU in a Random Forest Classification Framework

An Improved Patient-Specific Mortality Risk Prediction in ICU in a Random Forest Classification Framework An Improved Patient-Specific Mortality Risk Prediction in ICU in a Random Forest Classification Framework Soumya GHOSE, Jhimli MITRA 1, Sankalp KHANNA 1 and Jason DOWLING 1 1. The Australian e-health and

More information

VU Biostatistics and Experimental Design PLA.216

VU Biostatistics and Experimental Design PLA.216 VU Biostatistics and Experimental Design PLA.216 Julia Feichtinger Postdoctoral Researcher Institute of Computational Biotechnology Graz University of Technology Outline for Today About this course Background

More information

Predictive performance and discrimination in unbalanced classification

Predictive performance and discrimination in unbalanced classification MASTER Predictive performance and discrimination in unbalanced classification van der Zon, S.B. Award date: 2016 Link to publication Disclaimer This document contains a student thesis (bachelor's or master's),

More information

4 Diagnostic Tests and Measures of Agreement

4 Diagnostic Tests and Measures of Agreement 4 Diagnostic Tests and Measures of Agreement Diagnostic tests may be used for diagnosis of disease or for screening purposes. Some tests are more effective than others, so we need to be able to measure

More information

Essential Skills for Evidence-based Practice: Statistics for Therapy Questions

Essential Skills for Evidence-based Practice: Statistics for Therapy Questions Essential Skills for Evidence-based Practice: Statistics for Therapy Questions Jeanne Grace Corresponding author: J. Grace E-mail: Jeanne_Grace@urmc.rochester.edu Jeanne Grace RN PhD Emeritus Clinical

More information

Review. Imagine the following table being obtained as a random. Decision Test Diseased Not Diseased Positive TP FP Negative FN TN

Review. Imagine the following table being obtained as a random. Decision Test Diseased Not Diseased Positive TP FP Negative FN TN Outline 1. Review sensitivity and specificity 2. Define an ROC curve 3. Define AUC 4. Non-parametric tests for whether or not the test is informative 5. Introduce the binormal ROC model 6. Discuss non-parametric

More information

Knowledge Discovery and Data Mining. Testing. Performance Measures. Notes. Lecture 15 - ROC, AUC & Lift. Tom Kelsey. Notes

Knowledge Discovery and Data Mining. Testing. Performance Measures. Notes. Lecture 15 - ROC, AUC & Lift. Tom Kelsey. Notes Knowledge Discovery and Data Mining Lecture 15 - ROC, AUC & Lift Tom Kelsey School of Computer Science University of St Andrews http://tom.home.cs.st-andrews.ac.uk twk@st-andrews.ac.uk Tom Kelsey ID5059-17-AUC

More information

Week 2 Video 3. Diagnostic Metrics

Week 2 Video 3. Diagnostic Metrics Week 2 Video 3 Diagnostic Metrics Different Methods, Different Measures Today we ll continue our focus on classifiers Later this week we ll discuss regressors And other methods will get worked in later

More information

Data that can be classified as belonging to a distinct number of categories >>result in categorical responses. And this includes:

Data that can be classified as belonging to a distinct number of categories >>result in categorical responses. And this includes: This sheets starts from slide #83 to the end ofslide #4. If u read this sheet you don`t have to return back to the slides at all, they are included here. Categorical Data (Qualitative data): Data that

More information

4. Model evaluation & selection

4. Model evaluation & selection Foundations of Machine Learning CentraleSupélec Fall 2017 4. Model evaluation & selection Chloé-Agathe Azencot Centre for Computational Biology, Mines ParisTech chloe-agathe.azencott@mines-paristech.fr

More information

COMP90049 Knowledge Technologies

COMP90049 Knowledge Technologies COMP90049 Knowledge Technologies Introduction Classification (Lecture Set4) 2017 Rao Kotagiri School of Computing and Information Systems The Melbourne School of Engineering Some of slides are derived

More information

The Logic of Data Analysis Using Statistical Techniques M. E. Swisher, 2016

The Logic of Data Analysis Using Statistical Techniques M. E. Swisher, 2016 The Logic of Data Analysis Using Statistical Techniques M. E. Swisher, 2016 This course does not cover how to perform statistical tests on SPSS or any other computer program. There are several courses

More information

Statistics, Probability and Diagnostic Medicine

Statistics, Probability and Diagnostic Medicine Statistics, Probability and Diagnostic Medicine Jennifer Le-Rademacher, PhD Sponsored by the Clinical and Translational Science Institute (CTSI) and the Department of Population Health / Division of Biostatistics

More information

SISCR Module 7 Part I: Introduction Basic Concepts for Binary Biomarkers (Classifiers) and Continuous Biomarkers

SISCR Module 7 Part I: Introduction Basic Concepts for Binary Biomarkers (Classifiers) and Continuous Biomarkers SISCR Module 7 Part I: Introduction Basic Concepts for Binary Biomarkers (Classifiers) and Continuous Biomarkers Kathleen Kerr, Ph.D. Associate Professor Department of Biostatistics University of Washington

More information

3. Model evaluation & selection

3. Model evaluation & selection Foundations of Machine Learning CentraleSupélec Fall 2016 3. Model evaluation & selection Chloé-Agathe Azencot Centre for Computational Biology, Mines ParisTech chloe-agathe.azencott@mines-paristech.fr

More information

STATISTICAL METHODS FOR DIAGNOSTIC TESTING: AN ILLUSTRATION USING A NEW METHOD FOR CANCER DETECTION XIN SUN. PhD, Kansas State University, 2012

STATISTICAL METHODS FOR DIAGNOSTIC TESTING: AN ILLUSTRATION USING A NEW METHOD FOR CANCER DETECTION XIN SUN. PhD, Kansas State University, 2012 STATISTICAL METHODS FOR DIAGNOSTIC TESTING: AN ILLUSTRATION USING A NEW METHOD FOR CANCER DETECTION by XIN SUN PhD, Kansas State University, 2012 A THESIS Submitted in partial fulfillment of the requirements

More information

Predicting Breast Cancer Survivability Rates

Predicting Breast Cancer Survivability Rates Predicting Breast Cancer Survivability Rates For data collected from Saudi Arabia Registries Ghofran Othoum 1 and Wadee Al-Halabi 2 1 Computer Science, Effat University, Jeddah, Saudi Arabia 2 Computer

More information

Module Overview. What is a Marker? Part 1 Overview

Module Overview. What is a Marker? Part 1 Overview SISCR Module 7 Part I: Introduction Basic Concepts for Binary Classification Tools and Continuous Biomarkers Kathleen Kerr, Ph.D. Associate Professor Department of Biostatistics University of Washington

More information

Screening (Diagnostic Tests) Shaker Salarilak

Screening (Diagnostic Tests) Shaker Salarilak Screening (Diagnostic Tests) Shaker Salarilak Outline Screening basics Evaluation of screening programs Where we are? Definition of screening? Whether it is always beneficial? Types of bias in screening?

More information

A Learning Method of Directly Optimizing Classifier Performance at Local Operating Range

A Learning Method of Directly Optimizing Classifier Performance at Local Operating Range A Learning Method of Directly Optimizing Classifier Performance at Local Operating Range Lae-Jeong Park and Jung-Ho Moon Department of Electrical Engineering, Kangnung National University Kangnung, Gangwon-Do,

More information

Evaluation of diagnostic tests

Evaluation of diagnostic tests Evaluation of diagnostic tests Biostatistics and informatics Miklós Kellermayer Overlapping distributions Assumption: A classifier value (e.g., diagnostic parameter, a measurable quantity, e.g., serum

More information

Pooling Subjective Confidence Intervals

Pooling Subjective Confidence Intervals Spring, 1999 1 Administrative Things Pooling Subjective Confidence Intervals Assignment 7 due Friday You should consider only two indices, the S&P and the Nikkei. Sorry for causing the confusion. Reading

More information

Running head: How large denominators are leading to large errors 1

Running head: How large denominators are leading to large errors 1 Running head: How large denominators are leading to large errors 1 How large denominators are leading to large errors Nathan Thomas Kent State University How large denominators are leading to large errors

More information

Hypothesis-Driven Research

Hypothesis-Driven Research Hypothesis-Driven Research Research types Descriptive science: observe, describe and categorize the facts Discovery science: measure variables to decide general patterns based on inductive reasoning Hypothesis-driven

More information

How to assess the strength of relationships

How to assess the strength of relationships Publishing Date: April 1994. 1994. All rights reserved. Copyright rests with the author. No part of this article may be reproduced without written permission from the author. Meta Analysis 3 How to assess

More information

Chapter 7: Descriptive Statistics

Chapter 7: Descriptive Statistics Chapter Overview Chapter 7 provides an introduction to basic strategies for describing groups statistically. Statistical concepts around normal distributions are discussed. The statistical procedures of

More information

1 The conceptual underpinnings of statistical power

1 The conceptual underpinnings of statistical power 1 The conceptual underpinnings of statistical power The importance of statistical power As currently practiced in the social and health sciences, inferential statistics rest solidly upon two pillars: statistical

More information

Unraveling Recent Cervical Cancer Screening Updates and the Impact on Your Practice

Unraveling Recent Cervical Cancer Screening Updates and the Impact on Your Practice Transcript Details This is a transcript of a continuing medical education (CME) activity accessible on the ReachMD network. Additional media formats for the activity and full activity details (including

More information

Objectives. Quantifying the quality of hypothesis tests. Type I and II errors. Power of a test. Cautions about significance tests

Objectives. Quantifying the quality of hypothesis tests. Type I and II errors. Power of a test. Cautions about significance tests Objectives Quantifying the quality of hypothesis tests Type I and II errors Power of a test Cautions about significance tests Designing Experiments based on power Evaluating a testing procedure The testing

More information

15.301/310, Managerial Psychology Prof. Dan Ariely Recitation 8: T test and ANOVA

15.301/310, Managerial Psychology Prof. Dan Ariely Recitation 8: T test and ANOVA 15.301/310, Managerial Psychology Prof. Dan Ariely Recitation 8: T test and ANOVA Statistics does all kinds of stuff to describe data Talk about baseball, other useful stuff We can calculate the probability.

More information

Treating Lung Cancer: Past, Present, & Future Dr. Ramiswamy Govindan Washington University November, 2009

Treating Lung Cancer: Past, Present, & Future Dr. Ramiswamy Govindan Washington University November, 2009 Treating Lung Cancer: Past, Present, & Future Dr. Ramiswamy Govindan Washington University November, 2009 GRACE, the Global Resource for Advancing Cancer Education, is pleased to provide the following

More information

An Improved Algorithm To Predict Recurrence Of Breast Cancer

An Improved Algorithm To Predict Recurrence Of Breast Cancer An Improved Algorithm To Predict Recurrence Of Breast Cancer Umang Agrawal 1, Ass. Prof. Ishan K Rajani 2 1 M.E Computer Engineer, Silver Oak College of Engineering & Technology, Gujarat, India. 2 Assistant

More information

Patrick Breheny. January 28

Patrick Breheny. January 28 Confidence intervals Patrick Breheny January 28 Patrick Breheny Introduction to Biostatistics (171:161) 1/19 Recap Introduction In our last lecture, we discussed at some length the Public Health Service

More information

Tolerance of Effectiveness Measures to Relevance Judging Errors

Tolerance of Effectiveness Measures to Relevance Judging Errors Tolerance of Effectiveness Measures to Relevance Judging Errors Le Li 1 and Mark D. Smucker 2 1 David R. Cheriton School of Computer Science, Canada 2 Department of Management Sciences, Canada University

More information

Why we get hungry: Module 1, Part 1: Full report

Why we get hungry: Module 1, Part 1: Full report Why we get hungry: Module 1, Part 1: Full report Print PDF Does Anyone Understand Hunger? Hunger is not simply a signal that your stomach is out of food. It s not simply a time when your body can switch

More information

Research methods in sensation and perception. (and the princess and the pea)

Research methods in sensation and perception. (and the princess and the pea) Research methods in sensation and perception (and the princess and the pea) Sensory Thresholds We can measure stuff in the world like pressure, sound, light, etc. We can't easily measure your psychological

More information

Chapter 12. The One- Sample

Chapter 12. The One- Sample Chapter 12 The One- Sample z-test Objective We are going to learn to make decisions about a population parameter based on sample information. Lesson 12.1. Testing a Two- Tailed Hypothesis Example 1: Let's

More information

EXPERT INTERVIEW Diabetes Distress:

EXPERT INTERVIEW Diabetes Distress: EXPERT INTERVIEW Diabetes Distress: A real and normal part of diabetes Elizabeth Snouffer with Lawrence Fisher Living successfully with type 1 or type 2 diabetes requires the very large task of managing

More information

Step 2 Challenging negative thoughts "Weeding"

Step 2 Challenging negative thoughts Weeding Managing Automatic Negative Thoughts (ANTs) Step 1 Identifying negative thoughts "ANTs" Step 2 Challenging negative thoughts "Weeding" Step 3 Planting positive thoughts 'Potting" Step1 Identifying Your

More information

Probability Revision. MED INF 406 Assignment 5. Golkonda, Jyothi 11/4/2012

Probability Revision. MED INF 406 Assignment 5. Golkonda, Jyothi 11/4/2012 Probability Revision MED INF 406 Assignment 5 Golkonda, Jyothi 11/4/2012 Problem Statement Assume that the incidence for Lyme disease in the state of Connecticut is 78 cases per 100,000. A diagnostic test

More information

Christina Martin Kazi Russell MED INF 406 INFERENCING Session 8 Group Project November 15, 2014

Christina Martin Kazi Russell MED INF 406 INFERENCING Session 8 Group Project November 15, 2014 INFERENCING (HW 8) 1 Christina Martin Kazi Russell MED INF 406 INFERENCING Session 8 Group Project November 15, 2014 Page 2 The Clinical Decision Support System designed to utilize the Training Set data

More information

Introduction. We can make a prediction about Y i based on X i by setting a threshold value T, and predicting Y i = 1 when X i > T.

Introduction. We can make a prediction about Y i based on X i by setting a threshold value T, and predicting Y i = 1 when X i > T. Diagnostic Tests 1 Introduction Suppose we have a quantitative measurement X i on experimental or observed units i = 1,..., n, and a characteristic Y i = 0 or Y i = 1 (e.g. case/control status). The measurement

More information

Brain Tumor segmentation and classification using Fcm and support vector machine

Brain Tumor segmentation and classification using Fcm and support vector machine Brain Tumor segmentation and classification using Fcm and support vector machine Gaurav Gupta 1, Vinay singh 2 1 PG student,m.tech Electronics and Communication,Department of Electronics, Galgotia College

More information

Diagnostic tests, Laboratory tests

Diagnostic tests, Laboratory tests Diagnostic tests, Laboratory tests I. Introduction II. III. IV. Informational values of a test Consequences of the prevalence rate Sequential use of 2 tests V. Selection of a threshold: the ROC curve VI.

More information

Classification with microarray data

Classification with microarray data Classification with microarray data Aron Charles Eklund eklund@cbs.dtu.dk DNA Microarray Analysis - #27612 January 8, 2010 The rest of today Now: What is classification, and why do we do it? How to develop

More information

OCW Epidemiology and Biostatistics, 2010 David Tybor, MS, MPH and Kenneth Chui, PhD Tufts University School of Medicine October 27, 2010

OCW Epidemiology and Biostatistics, 2010 David Tybor, MS, MPH and Kenneth Chui, PhD Tufts University School of Medicine October 27, 2010 OCW Epidemiology and Biostatistics, 2010 David Tybor, MS, MPH and Kenneth Chui, PhD Tufts University School of Medicine October 27, 2010 SAMPLING AND CONFIDENCE INTERVALS Learning objectives for this session:

More information

Reframing Perspectives

Reframing Perspectives Page 1 Reframing Perspectives Reframing is an essential part of the coaching process as it helps others to see things differently and, as a result, come to different, more empowering conclusions or feelings

More information

Risk Aversion in Games of Chance

Risk Aversion in Games of Chance Risk Aversion in Games of Chance Imagine the following scenario: Someone asks you to play a game and you are given $5,000 to begin. A ball is drawn from a bin containing 39 balls each numbered 1-39 and

More information

August 29, Introduction and Overview

August 29, Introduction and Overview August 29, 2018 Introduction and Overview Why are we here? Haavelmo(1944): to become master of the happenings of real life. Theoretical models are necessary tools in our attempts to understand and explain

More information

EXERCISE: HOW TO DO POWER CALCULATIONS IN OPTIMAL DESIGN SOFTWARE

EXERCISE: HOW TO DO POWER CALCULATIONS IN OPTIMAL DESIGN SOFTWARE ...... EXERCISE: HOW TO DO POWER CALCULATIONS IN OPTIMAL DESIGN SOFTWARE TABLE OF CONTENTS 73TKey Vocabulary37T... 1 73TIntroduction37T... 73TUsing the Optimal Design Software37T... 73TEstimating Sample

More information

Economics 2010a. Fall Lecture 11. Edward L. Glaeser

Economics 2010a. Fall Lecture 11. Edward L. Glaeser Economics 2010a Fall 2003 Lecture 11 Edward L. Glaeser Final notes how to write a theory paper: (1) A highbrow theory paper go talk to Jerry or Drew don t listen to me. (2) A lowbrow or applied theory

More information

Something to think about. What happens, however, when we have a sample with less than 30 items?

Something to think about. What happens, however, when we have a sample with less than 30 items? One-Sample t-test Remember In the last chapter, we learned to use a statistic from a large sample of data to test a hypothesis about a population parameter. In our case, using a z-test, we tested a hypothesis

More information

Color naming and color matching: A reply to Kuehni and Hardin

Color naming and color matching: A reply to Kuehni and Hardin 1 Color naming and color matching: A reply to Kuehni and Hardin Pendaran Roberts & Kelly Schmidtke Forthcoming in Review of Philosophy and Psychology. The final publication is available at Springer via

More information

Analysis of Environmental Data Conceptual Foundations: En viro n m e n tal Data

Analysis of Environmental Data Conceptual Foundations: En viro n m e n tal Data Analysis of Environmental Data Conceptual Foundations: En viro n m e n tal Data 1. Purpose of data collection...................................................... 2 2. Samples and populations.......................................................

More information

Critical reading of diagnostic imaging studies. Lecture Goals. Constantine Gatsonis, PhD. Brown University

Critical reading of diagnostic imaging studies. Lecture Goals. Constantine Gatsonis, PhD. Brown University Critical reading of diagnostic imaging studies Constantine Gatsonis Center for Statistical Sciences Brown University Annual Meeting Lecture Goals 1. Review diagnostic imaging evaluation goals and endpoints.

More information

Feature selection methods for early predictive biomarker discovery using untargeted metabolomic data

Feature selection methods for early predictive biomarker discovery using untargeted metabolomic data Feature selection methods for early predictive biomarker discovery using untargeted metabolomic data Dhouha Grissa, Mélanie Pétéra, Marion Brandolini, Amedeo Napoli, Blandine Comte and Estelle Pujos-Guillot

More information

Introduction to Machine Learning. Katherine Heller Deep Learning Summer School 2018

Introduction to Machine Learning. Katherine Heller Deep Learning Summer School 2018 Introduction to Machine Learning Katherine Heller Deep Learning Summer School 2018 Outline Kinds of machine learning Linear regression Regularization Bayesian methods Logistic Regression Why we do this

More information

Teaching Statistics with Coins and Playing Cards Going Beyond Probabilities

Teaching Statistics with Coins and Playing Cards Going Beyond Probabilities Teaching Statistics with Coins and Playing Cards Going Beyond Probabilities Authors Chris Malone (cmalone@winona.edu), Dept. of Mathematics and Statistics, Winona State University Tisha Hooks (thooks@winona.edu),

More information

THE IMPORTANCE OF ETHICS IN EVERYDAY DECISIONS

THE IMPORTANCE OF ETHICS IN EVERYDAY DECISIONS THE IMPORTANCE OF ETHICS IN EVERYDAY DECISIONS Presented to Government Finance Officers Association Of Texas Fall Conference Presented by Dr. Raymond J. Clay, CPA Emeritus Professor of Accounting University

More information

PERFORMANCE MEASURES

PERFORMANCE MEASURES PERFORMANCE MEASURES Of predictive systems DATA TYPES Binary Data point Value A FALSE B TRUE C TRUE D FALSE E FALSE F TRUE G FALSE Real Value Data Point Value a 32.3 b.2 b 2. d. e 33 f.65 g 72.8 ACCURACY

More information

FMEA AND RPN NUMBERS. Failure Mode Severity Occurrence Detection RPN A B

FMEA AND RPN NUMBERS. Failure Mode Severity Occurrence Detection RPN A B FMEA AND RPN NUMBERS An important part of risk is to remember that risk is a vector: one aspect of risk is the severity of the effect of the event and the other aspect is the probability or frequency of

More information

Sheila Barron Statistics Outreach Center 2/8/2011

Sheila Barron Statistics Outreach Center 2/8/2011 Sheila Barron Statistics Outreach Center 2/8/2011 What is Power? When conducting a research study using a statistical hypothesis test, power is the probability of getting statistical significance when

More information

Modifying ROC Curves to Incorporate Predicted Probabilities

Modifying ROC Curves to Incorporate Predicted Probabilities Modifying ROC Curves to Incorporate Predicted Probabilities C. Ferri, P. Flach 2, J. Hernández-Orallo, A. Senad Departament de Sistemes Informàtics i Computació Universitat Politècnica de València Spain

More information

Analysis of Diabetic Dataset and Developing Prediction Model by using Hive and R

Analysis of Diabetic Dataset and Developing Prediction Model by using Hive and R Indian Journal of Science and Technology, Vol 9(47), DOI: 10.17485/ijst/2016/v9i47/106496, December 2016 ISSN (Print) : 0974-6846 ISSN (Online) : 0974-5645 Analysis of Diabetic Dataset and Developing Prediction

More information

Applying Data Mining for Epileptic Seizure Detection

Applying Data Mining for Epileptic Seizure Detection Applying Data Mining for Epileptic Seizure Detection Ying-Fang Lai 1 and Hsiu-Sen Chiang 2* 1 Department of Industrial Education, National Taiwan Normal University 162, Heping East Road Sec 1, Taipei,

More information

Glossary From Running Randomized Evaluations: A Practical Guide, by Rachel Glennerster and Kudzai Takavarasha

Glossary From Running Randomized Evaluations: A Practical Guide, by Rachel Glennerster and Kudzai Takavarasha Glossary From Running Randomized Evaluations: A Practical Guide, by Rachel Glennerster and Kudzai Takavarasha attrition: When data are missing because we are unable to measure the outcomes of some of the

More information

BMI 541/699 Lecture 16

BMI 541/699 Lecture 16 BMI 541/699 Lecture 16 Where we are: 1. Introduction and Experimental Design 2. Exploratory Data Analysis 3. Probability 4. T-based methods for continous variables 5. Proportions & contingency tables -

More information

Healthy Communities Conference Ana Diez Roux 1. Okay, good afternoon. It s a pleasure to be here. I guess by, I

Healthy Communities Conference Ana Diez Roux 1. Okay, good afternoon. It s a pleasure to be here. I guess by, I 1 Okay, good afternoon. It s a pleasure to be here. I guess by, I don t know, things happen in life. I ve become an academic somehow, although I started out as a pediatrician a long time ago. And it s

More information

SEED HAEMATOLOGY. Medical statistics your support when interpreting results SYSMEX EDUCATIONAL ENHANCEMENT AND DEVELOPMENT APRIL 2015

SEED HAEMATOLOGY. Medical statistics your support when interpreting results SYSMEX EDUCATIONAL ENHANCEMENT AND DEVELOPMENT APRIL 2015 SYSMEX EDUCATIONAL ENHANCEMENT AND DEVELOPMENT APRIL 2015 SEED HAEMATOLOGY Medical statistics your support when interpreting results The importance of statistical investigations Modern medicine is often

More information

Hat Puzzles. Tanya Khovanova MIT s Research Science Institute.

Hat Puzzles. Tanya Khovanova MIT s Research Science Institute. Problems Hat Puzzles Tanya Khovanova MIT s Research Science Institute tanya@math.mit.edu Abstract: This paper serves as the announcement of my program a joke version of the Langlands Program. In connection

More information

The Fallacy of Taking Random Supplements

The Fallacy of Taking Random Supplements The Fallacy of Taking Random Supplements Healthview interview with Dr. Paul Eck Healthview: We can see from our conversations that you are totally against people taking random supplements even if people

More information

CHAPTER 2 MAMMOGRAMS AND COMPUTER AIDED DETECTION

CHAPTER 2 MAMMOGRAMS AND COMPUTER AIDED DETECTION 9 CHAPTER 2 MAMMOGRAMS AND COMPUTER AIDED DETECTION 2.1 INTRODUCTION This chapter provides an introduction to mammogram and a description of the computer aided detection methods of mammography. This discussion

More information

The Wellbeing Course. Resource: Mental Skills. The Wellbeing Course was written by Professor Nick Titov and Dr Blake Dear

The Wellbeing Course. Resource: Mental Skills. The Wellbeing Course was written by Professor Nick Titov and Dr Blake Dear The Wellbeing Course Resource: Mental Skills The Wellbeing Course was written by Professor Nick Titov and Dr Blake Dear About Mental Skills This resource introduces three mental skills which people find

More information

Section 3.2 Least-Squares Regression

Section 3.2 Least-Squares Regression Section 3.2 Least-Squares Regression Linear relationships between two quantitative variables are pretty common and easy to understand. Correlation measures the direction and strength of these relationships.

More information

CHAPTER 5 DECISION TREE APPROACH FOR BONE AGE ASSESSMENT

CHAPTER 5 DECISION TREE APPROACH FOR BONE AGE ASSESSMENT 53 CHAPTER 5 DECISION TREE APPROACH FOR BONE AGE ASSESSMENT The decision tree approach for BAA makes use of the radius and ulna wrist bones to estimate the bone age. From the radius and ulna bones, 11

More information

Introduction to diagnostic accuracy meta-analysis. Yemisi Takwoingi October 2015

Introduction to diagnostic accuracy meta-analysis. Yemisi Takwoingi October 2015 Introduction to diagnostic accuracy meta-analysis Yemisi Takwoingi October 2015 Learning objectives To appreciate the concept underlying DTA meta-analytic approaches To know the Moses-Littenberg SROC method

More information

S4. Summary of the GALNS assay validation. Intra-assay variation (within-run precision)

S4. Summary of the GALNS assay validation. Intra-assay variation (within-run precision) S4. Summary of the GALNS assay validation (i.) Intra-assay variation (within-run precision) Intra-assay variation was determined by measuring standard blood samples (low activity standard; medium activity

More information

CHAPTER 6 HUMAN BEHAVIOR UNDERSTANDING MODEL

CHAPTER 6 HUMAN BEHAVIOR UNDERSTANDING MODEL 127 CHAPTER 6 HUMAN BEHAVIOR UNDERSTANDING MODEL 6.1 INTRODUCTION Analyzing the human behavior in video sequences is an active field of research for the past few years. The vital applications of this field

More information

The 3 Things NOT To Do When You Quit Smoking

The 3 Things NOT To Do When You Quit Smoking The 3 Things NOT To Do When You Quit Smoking Here are the 3 common mistakes people make when they try to quit smoking: 1. They think quitting will be hard not true 2. They think they have no willpower

More information

Worksheet for Structured Review of Physical Exam or Diagnostic Test Study

Worksheet for Structured Review of Physical Exam or Diagnostic Test Study Worksheet for Structured Review of Physical Exam or Diagnostic Study Title of Manuscript: Authors of Manuscript: Journal and Citation: Identify and State the Hypothesis Primary Hypothesis: Secondary Hypothesis:

More information

A supported model of decisionmaking:

A supported model of decisionmaking: A supported model of decisionmaking: The starting point: The starting point is not a test of capacity, but the presumption that every human being is communicating all the time and that this communication

More information

Name Psychophysical Methods Laboratory

Name Psychophysical Methods Laboratory Name Psychophysical Methods Laboratory 1. Classical Methods of Psychophysics These exercises make use of a HyperCard stack developed by Hiroshi Ono. Open the PS 325 folder and then the Precision and Accuracy

More information

Ingredients of Difficult Conversations

Ingredients of Difficult Conversations Ingredients of Difficult Conversations Differing Perceptions In most difficult conversations, there are different perceptions of the same reality. I think I'm right and the person with whom I disagree

More information

Worried about your memory?

Worried about your memory? Worried about your memory? Find out more about memory problems and what you can do next alzheimers.org.uk Worried about memory problems? Everybody forgets things from time to time. But if you or other

More information

An Empirical and Formal Analysis of Decision Trees for Ranking

An Empirical and Formal Analysis of Decision Trees for Ranking An Empirical and Formal Analysis of Decision Trees for Ranking Eyke Hüllermeier Department of Mathematics and Computer Science Marburg University 35032 Marburg, Germany eyke@mathematik.uni-marburg.de Stijn

More information