Introduction to ROC Curves

About ROC curves

A ROC curve provides a graphical representation of the relationship between the true-positive and false-positive prediction rate of a model. The y-axis corresponds to the sensitivity of the model, i.e. how well the model is able to predict true positives (real cleavages) from sites that are not cleaved, and the y-coordinates are calculated as:

The x-axis corresponds to the specificity (expressed on the curve as 1-specificity), i.e. the ability of the model to identify true negatives. An increase in specificity (i.e. a decrease along the X-axis) results in an increase in sensitivity. The x-coordinates are calculated as:

The greater the sensitivity at high specificity values (i.e. high y-axis values at low X-axis values) the better the model. A numerical measure of the accuracy of the model can be obtained from the area under the curve, where an area of 1.0 signifies near perfect accuracy, while an area of less than 0.5 indicates that the model is worse than just random. The quantitative-qualitative relationship between area and accuracy follows a fairly linear pattern, such that the following could be used as a guide:

• 0.9-1: Excellent
• 0.8-0.9: Very good
• 0.7-0.8: Good
• 0.6-0.7: Average
• 0.5-0.6: Poor

Introduction to ROC Curves

The sensitivity and specificity of a diagnostic test depends on more than just the "quality" of the test--they also depend on the definition of what constitutes an abnormal test. Look at the the idealized graph at right showing the number of patients with and without a disease arranged according to the value of a diagnostic test. This distributions overlap--the test (like most) does not distinguish normal from disease with 100% accuracy. The area of overlap indicates where the test cannot distinguish normal from disease. In practice, we choose a cutpoint (indicated by the vertical black line) above which we consider the test to be abnormal and below which we consider the test to be normal. The position of the cutpoint will determine the number of true positive, true negatives, false positives and false negatives. We may wish to use different cutpoints for different clinical situations if we wish to minimize one of the erroneous types of test results.

We can use the hypothyroidism data from the likelihood ratio section to illustrate how sensitivity and specificity change depending on the choice of T4 level that defines hypothyroidism. Recall the data on patients with suspected hypothyroidism reported by Goldstein and Mushlin (J Gen Intern Med 1987;2:20-24.). The data on T4 values in hypothyroid and euthyroid patients are shown graphically (below left) and in a simplified tabular form (below right).

T4 value	Hypothyroid	Euthyroid
5 or less	18	1
5.1 - 7	7	17
7.1 - 9	4	36
9 or more	3	39
Totals:	32	93

Suppose that patients with T4 values of 5 or less are considered to be hypothyroid. The data display then reduces to:

T4 value	Hypothyroid	Euthyroid
5 or less	18	1
> 5	14	92
Totals:	32	93

You should be able to verify that the sensivity is 0.56 and the specificity is 0.99.

Now, suppose we decide to make the definition of hypothyroidism less stringent and now consider patients with T4 values of 7 or less to be hypothyroid. The data display will now look like this:

T4 value	Hypothyroid	Euthyroid
7 or less	25	18
> 7	7	75
Totals:	32	93

You should be able to verify that the sensivity is 0.78 and the specificity is 0.81.

Lets move the cut point for hypothyroidism one more time:

T4 value	Hypothyroid	Euthyroid
<>	29	54
9 or more	3	39
Totals:	32	93

You should be able to verify that the sensivity is 0.91 and the specificity is 0.42.

Now, take the sensitivity and specificity values above and put them into a table:

Cutpoint	Sensitivity	Specificity
5	0.56	0.99
7	0.78	0.81
9	0.91	0.42

Notice that you can improve the sensitivity by moving to cutpoint to a higher T4 value--that is, you can make the criterion for a positive test less strict. You can improve the specificity by moving the cutpoint to a lower T4 value--that is, you can make the criterion for a positive test more strict. Thus, there is a tradeoff between sensitivity and specificity. You can change the definition of a positive test to improve one but the other will decline.

The next section covers how to use the numbers we just calculated to draw and interpret an ROC curve.
.

|

Plotting and Intrepretating an ROC Curve

This section continues the hypothyroidism example started in the the previous section. We showed that the table at left can be summarized by the operating characteristics at right:

T4 value	Hypothyroid	Euthyroid
5 or less	18	1
5.1 - 7	7	17
7.1 - 9	4	36
9 or more	3	39
Totals:	32	93

Cutpoint	Sensitivity	Specificity
5	0.56	0.99
7	0.78	0.81
9	0.91	0.42

The operating characteristics (above right) can be reformulated slightly and then presented graphically as shown below to the right:

Cutpoint	True Positives	False Positives
5	0.56	0.01
7	0.78	0.19
9	0.91	0.58

This type of graph is called a Receiver Operating Characteristic curve (or ROC curve.) It is a plot of the true positive rate against the false positive rate for the different possible cutpoints of a diagnostic test.

An ROC curve demonstrates several things:

1. It shows the tradeoff between sensitivity and specificity (any increase in sensitivity will be accompanied by a decrease in specificity).
2. The closer the curve follows the left-hand border and then the top border of the ROC space, the more accurate the test.
3. The closer the curve comes to the 45-degree diagonal of the ROC space, the less accurate the test.
4. The slope of the tangent line at a cutpoint gives the likelihood ratio (LR) for that value of the test. You can check this out on the graph above. Recall that the LR for T4 <> 9 is 0.2. This corresponds to the far right, nearly horizontal portion of the curve.
5. The area under the curve is a measure of text accuracy. This is discussed further in the next section.

.

The Area Under an ROC Curve

| Previous Section | Main Menu | Next Section |

The graph at right shows three ROC curves representing excellent, good, and worthless tests plotted on the same graph. The accuracy of the test depends on how well the test separates the group being tested into those with and without the disease in question. Accuracy is measured by the area under the ROC curve. An area of 1 represents a perfect test; an area of .5 represents a worthless test. A rough guide for classifying the accuracy of a diagnostic test is the traditional academic point system:

• .90-1 = excellent (A)
• .80-.90 = good (B)
• .70-.80 = fair (C)
• .60-.70 = poor (D)
• .50-.60 = fail (F)

Recall the T4 data from the previous section. The area under the T4 ROC curve is .86. The T4 would be considered to be "good" at separating hypothyroid from euthyroid patients.

ROC curves can also be constructed from clinical prediction rules. The graphs at right come from a study of how clinical findings predict strep throat (Wigton RS, Connor JL, Centor RM. Transportability of a decision rule for the diagnosis of streptococcal pharyngitis. Arch Intern Med. 1986;146:81-83.) In that study, the presence of tonsillar exudate, fever, adenopathy and the absence of cough all predicted strep. The curves were constructed by computing the sensitivity and specificity of increasing numbers of clinical findings (from 0 to 4) in predicting strep. The study compared patients in Virginia and Nebraska and found that the rule performed more accurately in Virginia (area under the curve = .78) compared to Nebraska (area under the curve = .73). These differences turn out not to be statistically different, however.

At this point, you may be wondering what this area number really means and how it is computed. The area measures discrimination, that is, the ability of the test to correctly classify those with and without the disease. Consider the situation in which patients are already correctly classified into two groups. You randomly pick on from the disease group and one from the no-disease group and do the test on both. The patient with the more abnormal test result should be the one from the disease group. The area under the curve is the percentage of randomly drawn pairs for which this is true (that is, the test correctly classifies the two patients in the random pair).

Computing the area is more difficult to explain and beyond the scope of this introductory material. Two methods are commonly used: a non-parametric method based on constructing trapeziods under the curve as an approximation of area and a parametric method using a maximum likelihood estimator to fit a smooth curve to the data points. Both methods are available as computer programs and give an estimate of area and standard error that can be used to compare different tests or the same test in different patient populations. For more on quantitative ROC analysis, see Metz CE. Basic principles of ROC analysis. Sem Nuc Med. 1978;8:283-298.

A final note of historical interest
You may be wondering where the name "Reciever Operating Characteristic" came from. ROC analysis is part of a field called "Signal Dectection Therory" developed during World War II for the analysis of radar images. Radar operators had to decide whether a blip on the screen represented an enemy target, a friendly ship, or just noise. Signal detection theory measures the ability of radar receiver operators to make these important distinctions. Their ability to do so was called the Receiver Operating Characteristics. It was not until the 1970's that signal detection theory was recognized as useful for interpreting medical test results.

Logistic Regression

Logistic Regression

What is the logistic curve? What is the base of the natural logarithm? Why do statisticians prefer logistic regression to ordinary linear regression when the DV is binary? How are probabilities, odds and logits related? What is an odds ratio? How can logistic regression be considered a linear regression? What is a loss function? What is a maximum likelihood estimate? How is the b weight in logistic regression for a categorical variable related to the odds ratio of its constituent categories?

This chapter is difficult because there are many new concepts in it. Studying this may bring back feelings that you had in the first third of the course, when there were many new concepts each week.

For this chapter only, we are going to deal with a dependent variable that is binary (a categorical variable that has two values such as "yes" and "no") rather than continuous.

[Technical note: Logistic regression can also be applied to ordered categories (ordinal data), that is, variables with more than two ordered categories, such as what you find in many surveys. However, we won't be dealing with that in this course and you probably will never be taught it. If our dependent variable has several unordered categories (e.g., suppose our DV was state of origin in the U.S.), then we can use something called discriminant analysis, which will be taught to you in a course on multivariate statistics.]

It is customary to code a binary DV either 0 or 1. For example, we might code a successfully kicked field goal as 1 and a missed field goal as 0 or we might code yes as 1 and no as 0 or admitted as 1 and rejected as 0 or Cherry Garcia flavor ice cream as 1 and all other flavors as zero. If we code like this, then the mean of the distribution is equal to the proportion of 1s in the distribution. For example if there are 100 people in the distribution and 30 of them are coded 1, then the mean of the distribution is .30, which is the proportion of 1s. The mean of the distribution is also the probability of drawing a person labeled as 1 at random from the distribution. That is, if we grab a person at random from our sample of 100 that I just described, the probability that the person will be a 1 is .30. Therefore, proportion and probability of 1 are the same in such cases. The mean of a binary distribution so coded is denoted as P, the proportion of 1s. The proportion of zeros is (1-P), which is sometimes denoted as Q. The variance of such a distribution is PQ, and the standard deviation is Sqrt(PQ). {Why can't all of stats be this easy?}

Suppose we want to predict whether someone is male or female (DV, M=1, F=0) using height in inches (IV). We could plot the relations between the two variables as we customarily do in regression. The plot might look something like this:

Points to notice about the graph (data are fictional):

1. The regression line is a rolling average, just as in linear regression. The Y-axis is P, which indicates the proportion of 1s at any given value of height. (review graph)
2. The regression line is nonlinear. (review graph)
3. None of the observations --the raw data points-- actually fall on the regression line. They all fall on zero or one. (review graph)

Why use logistic regression rather than ordinary linear regression?

When I was in graduate school, people didn't use logistic regression with a binary DV. They just used ordinary linear regression instead. Statisticians won the day, however, and now most psychologists use logistic regression with a binary DV for the following reasons:

1. If you use linear regression, the predicted values will become greater than one and less than zero if you move far enough on the X-axis. Such values are theoretically inadmissible.
2. One of the assumptions of regression is that the variance of Y is constant across values of X (homoscedasticity). This cannot be the case with a binary variable, because the variance is PQ. When 50 percent of the people are 1s, then the variance is .25, its maximum value. As we move to more extreme values, the variance decreases. When P=.10, the variance is .1*.9 = .09, so as P approaches 1 or zero, the variance approaches zero.
3. The significance testing of the b weights rest upon the assumption that errors of prediction (Y-Y') are normally distributed. Because Y only takes the values 0 and 1, this assumption is pretty hard to justify, even approximately. Therefore, the tests of the regression weights are suspect if you use linear regression with a binary DV.

The Logistic Curve

The logistic curve relates the independent variable, X, to the rolling mean of the DV, P (). The formula to do so may be written either

Or

where P is the probability of a 1 (the proportion of 1s, the mean of Y), e is the base of the natural logarithm (about 2.718) and a and b are the parameters of the model. The value of a yields P when X is zero, and b adjusts how quickly the probability changes with changing X a single unit (we can have standardized and unstandardized b weights in logistic regression, just as in ordinary linear regression). Because the relation between X and P is nonlinear, b does not have a straightforward interpretation in this model as it does in ordinary linear regression.

Loss Function

A loss function is a measure of fit between a mathematical model of data and the actual data. We choose the parameters of our model to minimize the badness-of-fit or to maximize the goodness-of-fit of the model to the data. With least squares (the only loss function we have used thus far), we minimize SS_res, the sum of squares residual. This also happens to maximize SS_reg, the sum of squares due to regression. With linear or curvilinear models, there is a mathematical solution to the problem that will minimize the sum of squares, that is,

b = (X'X)^-1X'y

Or

b = R^-1r

With some models, like the logistic curve, there is no mathematical solution that will produce least squares estimates of the parameters. For many of these models, the loss function chosen is called maximum likelihood. A likelihood is a conditional probability (e.g., P(Y|X), the probability of Y given X). We can pick the parameters of the model (a and b of the logistic curve) at random or by trial-and-error and then compute the likelihood of the data given those parameters (actually, we do better than trail-and-error, but not perfectly). We will choose as our parameters, those that result in the greatest likelihood computed. The estimates are called maximum likelihood because the parameters are chosen to maximize the likelihood (conditional probability of the data given parameter estimates) of the sample data. The techniques actually employed to find the maximum likelihood estimates fall under the general label numerical analysis. There are several methods of numerical analysis, but they all follow a similar series of steps. First, the computer picks some initial estimates of the parameters. Then it will compute the likelihood of the data given these parameter estimates. Then it will improve the parameter estimates slightly and recalculate the likelihood of the data. It will do this forever until we tell it to stop, which we usually do when the parameter estimates do not change much (usually a change .01 or .001 is small enough to tell the computer to stop). [Sometimes we tell the computer to stop after a certain number of tries or iterations, e.g., 20 or 250. This usually indicates a problem in estimation.]

Where on Earth Did This Stuff Come From?

Suppose we only know a person's height and we want to predict whether that person is male or female. We can talk about the probability of being male or female, or we can talk about the odds of being male or female. Let's say that the probability of being male at a given height is .90. Then the odds of being male would be

.

(Odds can also be found by counting the number of people in each group and dividing one number by the other. Clearly, the probability is not the same as the odds.) In our example, the odds would be .90/.10 or 9 to one. Now the odds of being female would be .10/.90 or 1/9 or .11. This asymmetry is unappealing, because the odds of being a male should be the opposite of the odds of being a female. We can take care of this asymmetry though the natural logarithm, ln. The natural log of 9 is 2.217 (ln(.9/.1)=2.217). The natural log of 1/9 is -2.217 (ln(.1/.9)=-2.217), so the log odds of being male is exactly opposite to the log odds of being female. The natural log function looks like this:

Note that the natural log is zero when X is 1. When X is larger than one, the log curves up slowly. When X is less than one, the natural log is less than zero, and decreases rapidly as X approaches zero. When P = .50, the odds are .50/.50 or 1, and ln(1) =0. If P is greater than .50, ln(P/(1-P) is positive; if P is less than .50, ln(odds) is negative. [A number taken to a negative power is one divided by that number, e.g. e^-10 = 1/e^10. A logarithm is an exponent from a given base, for example ln(e¹⁰) = 10.]

Back to logistic regression.

In logistic regression, the dependent variable is a logit, which is the natural log of the odds, that is,

So a logit is a log of odds and odds are a function of P, the probability of a 1. In logistic regression, we find

logit(P) = a + bX,

Which is assumed to be linear, that is, the log odds (logit) is assumed to be linearly related to X, our IV. So there's an ordinary regression hidden in there. We could in theory do ordinary regression with logits as our DV, but of course, we don't have logits in there, we have 1s and 0s. Then, too, people have a hard time understanding logits. We could talk about odds instead. Of course, people like to talk about probabilities more than odds. To get there (from logits to probabilities), we first have to take the log out of both sides of the equation. Then we have to convert odds to a simple probability:

The simple probability is this ugly equation that you saw earlier. If log odds are linearly related to X, then the relation between X and P is nonlinear, and has the form of the S-shaped curve you saw in the graph and the function form (equation) shown immediately above.

An Example

Suppose that we are working with some doctors on heart attack patients. The dependent variable is whether the patient has had a second heart attack within 1 year (yes = 1). We have two independent variables, one is whether the patient completed a treatment consistent of anger control practices (yes=1). The other IV is a score on a trait anxiety scale (a higher score means more anxious).

Our data:

Person	2nd Heart Attack	Treatment of Anger	Trait Anxiety
1	1	1	70
2	1	1	80
3	1	1	50
4	1	0	60
5	1	0	40
6	1	0	65
7	1	0	75
8	1	0	80
9	1	0	70
10	1	0	60
11	0	1	65
12	0	1	50
13	0	1	45
14	0	1	35
15	0	1	40
16	0	1	50
17	0	0	55
18	0	0	45
19	0	0	50
20	0	0	60

Our correlation matrix:

	Heart	Treat	Anx
Heart	1
Treat	-.30	1
Anx	.59**	-.23	1
Mean	.50	.45	57.25
SD	.51	.51	13.42

Note that half of our patients have had a second heart attack. Knowing nothing else about a patient, and following the best in current medical practice, we would flip a coin to predict whether they will have a second attack within 1 year. According to our correlation coefficients, those in the anger treatment group are less likely to have another attack, but the result is not significant. Greater anxiety is associated with a higher probability of another attack, and the result is significant (according to r).

Now let's look at the logistic regression, for the moment examining the treatment of anger by itself, ignoring the anxiety test scores. SAS prints this:

Response Variable: HEART

Response Levels: 2

Number of Observations: 20

Link Function: Logit

Response Profile

Ordered

Value HEART Count

1 0 10

2 1 10

SAS tells us what it understands us to model, including the name of the DV, and its distribution.

Then we calculate probabilities with and without including the treatment variable.

Model Fitting Information and Testing Global Null Hypothesis BETA=0

Criterion Intercept Intercept Chi-sq

Only and

Covariates

-2 LOG L 27.726 25.878 1.848

1df (p=.17)

The computer calculates the likelihood of the data. Because there are equal numbers of people in the two groups, the probability of group membership initially (without considering anger treatment) is .50 for each person. Because the people are independent, the probability of the entire set of people is .50²⁰, a very small number. Because the number is so small, it is customary to first take the natural log of the probability and then multiply the result by -2. The latter step makes the result positive. The statistic -2LogL (minus 2 times the log of the likelihood) is a badness-of-fit indicator, that is, large numbers mean poor fit of the model to the data. SAS prints the result as -2 LOG L. For the initial model (intercept only), our result is the value 27.726. This is a baseline number indicating model fit. This number has no direct analog in linear regression. It is roughly analogous to generating some random numbers and finding R² for these numbers as a baseline measure of fit in ordinary linear regression. By including a term for treatment, the loss function reduces to 25.878, a difference of 1.848, shown in the chi-square column. The difference between the two values of -2LogL is known as the likelihood ratio test.

When taken from large samples, the difference between two values of -2LogL is distributed as chi-square:

Recall that multiplying numbers is equivalent to adding exponents (same for subtraction and division of logs).

This says that the (-2Log L) for a restricted (smaller) model - (-2LogL) for a full (larger) model is the same as the log of the ratio of two likelihoods, which is distributed as chi-square. The full or larger model has all the parameters of interest in it. The restricted is said to be nested in the larger model. The restricted model has one or more of parameters in the full model restricted to some value (usually zero). The parameters in the nested model must be a proper subset of the parameters in the full model. For example, suppose we have two IVs, one categorical and once continuous, and we are looking at an ATI design. A full model could have included terms for the continuous variable, the categorical variable and their interaction (3 terms). Restricted models could delete the interaction or one or more main effects (e.g., we could have a model with only the categorical variable). A nested model cannot have as a single IV, some other categorical or continuous variable not contained in the full model. If it does, then it is no longer nested, and we cannot compare the two values of -2LogL to get a chi-square value. The chi-square is used to statistically test whether including a variable reduces badness-of-fit measure. This is analogous to producing an increment in R-square in hierarchical regression. If chi-square is significant, the variable is considered to be a significant predictor in the equation, analogous to the significance of the b weight in simultaneous regression.

For our example with anger treatment only, SAS produces the following:

Analysis of Maximum Likelihood Estimates
Variable	DF	Par Est	Std Err	Wald Chisq	Pr > Chi- sq	Stand. Est	Odds Ratio
Intercept	1	-.5596	.6268	.7972	.3719	.	.
Treatment	1	1.2528	.9449	17566	.1849	.3525	3.50

The intercept is the value of a, in this case -.5596. As usual, we are not terribly interested in whether a is equal to zero. The value of b given for Anger Treatment is 1.2528. the chi-square associated with this b is not significant, just as the chi-square for covariates was not significant. Therefore we cannot reject the hypothesis that b is zero in the population. Our equation can be written either:

Logit(P) = -.5596+1.2528X

Or

The main interpretation of logistic regression results is to find the significant predictors of Y. However, other things can sometimes be done with the results.

The Odds Ratio

Recall that the odds for a group is :

Now the odds for another group would also be P/(1-P) for that group. Suppose we arrange our data in the following way:

	Anger Treatment
Heart Attack	Yes (1)	No (0)	Total
Yes (1)	3 (a)	7 (b)	10 (a+b)
No (0)	6 (c)	4 (d)	10 (c+d)
Total	9 (a+c)	11 (b+d)	20 (a+b+c+d)

Now we can compute the odds of having a heart attack for the treatment group and the no treatment group. For the treatment group, the odds are 3/6 = 1/2. The probability of a heart attack is 3/(3+6) = 3/9 = .33. The odds from this probability are .33/(1-.33) = .33/.66 = 1/2. The odds for the no treatment group are 7/4 or 1.75. The odds ratio is calculated to compare the odds across groups.

If the odds are the same across groups, the odds ratio (OR) will be 1.0. If not, the OR will be larger or smaller than one. People like to see the ratio be phrased in the larger direction. In our case, this would be 1.75/.5 or 1.75*2 = 3.50.

Now if we go back up to the last column of the printout where is says odds ratio in the treatment column, you will see that the odds ratio is 3.50, which is what we got by finding the odds ratio for the odds from the two treatment conditions. It also happens that e^1.2528 = 3.50. Note that the exponent is our value of b for the logistic curve.

K-fold Cross Validation

Cross validation is a method for estimating the true error of a model. When a model is built from training data, the error on the training data is a rather optimistic estimate of the error rates the model will achieve on unseen data. The aim of building a model is usually to apply the model to new, unseen data--we expect the model to generalise to data other than the training data on which it was built. Thus, we would like to have some method for better approximating the error that might occur in general. Cross validation provides such a method.

Cross validation is also used to evaluate a model in deciding which algorithm to deploy for learning, when choosing from amongst a number of learning algorithms. It can also provide a guide as to the effect of parameter tuning in building a model from a specific algorithm.

Test sample cross-validation is often a preferred method when there is plenty of data available. A model is built from a training set and its predictive accuracy is measured by applying the model a test set. A good rule of thumb is that a dataset is partitioned into a training set (66%) and a test set (33%).

To measure error rates you might build multiple models with the one algorithm, using variations of the same training data for each model. The average performance is then the measure of how well this algorithm works in building models from the data.

The basic idea is to use, say, 90% of the dataset to build a model. The data that was removed (the 10%) is then used to test the performance of the model on ``new'' data (usually by calculating the mean squared error). This simplest of cross validation approaches is referred to as the holdout method.

For the holdout method the two datasets are referred to as the training set and the test set. With just a single evaluation though there can be a high variance since the evaluation is dependent on the data points which happen to end up in the training set and the test set. Different partitions might lead to different results.

A solution to this problem is to have multiple subsets, and each time build the model based on all but one of these subsets. This is repeated for all possible combinations and the result is reported as the average error over all models.

In k-­fold cross ­validation, sometimes called rotation estimation, the dataset D is randomly split into k mutually exclusive subsets (the folds) D1 ; D2 ; ...; Dk of approximately equal size. The inducer is trained and tested k times; each time t from {1; 2;... ; k}, it is trained on D - Dt and tested on Dt . The cross­validation estimate of accuracy is the overall number of correct classifica­
tions, divided by the number of instances in the dataset. Formally, let D (i) be the test set that includes instance x i = (vi ; yi), then the cross­validation estimate of accuracy

The cross­validation estimate is a random number that depends on the division into folds. Complete cross ­validation is the average of all

possibilities for choosing m-k instances out of m, but it is usually too expensive. Except for leave­one­one (n­fold cross­validation), which is always complete, k­fold cross­ validation is estimating complete k­fold cross­validation using a single split of the data into the folds. Repeat­ing cross­validation multiple times using different splits into folds provides a better Monte­Carlo estimate to the complete cross­validation at an added cost. In strati­fied cross­validation, the folds are stratified so that they contain approximately the same proportions of la­bels as the original dataset.