The LOGISTIC Procedure

displays a label on diagnostic plots to aid in identifying the outlying observations. This option enhances the plots produced by the DFBETAS, DPC, INFLUENCE, LEVERAGE, and PHAT options. If an ID statement is specified, then the plots are labeled with the ID variables. Otherwise, the observation number is displayed.

suppresses the plots produced by the DFBETAS, DPC, INFLUENCE, LEVERAGE, and PHAT options if there are more than number observations. Also, observations are not displayed on the EFFECT plots when the cutoff is exceeded. The default is MAXPOINTS=5000. The cutoff is ignored if you specify MAXPOINTS=NONE.

displays only specifically requested plot-requests.

suppresses paneling. By default, multiple plots can appear in some output panels. Specify UNPACKPANEL to display each plot separately.

produces all appropriate plots. You can specify other options with ALL. For example, to display all plots and unpack the DFBETAS plots, you can specify plots=(all dfbetas(unpack)).

displays calibration plots for the fitted model. For binomial response data, a loess curve is fit to the observed events/trials ratios versus the predicted probabilities. For binary data, an indicator variable is set to 1 if the response is an event and set to 0 otherwise, and a loess curve is fit to this indicator versus the predicted probabilities. For polytomous response data, a panel of plots is produced that displays one plot for each response level as follows: an indicator variable is set to 1 if the observed response equals that level and set to 0 otherwise, and a loess curve is fit to this indicator versus the predicted probability of that level. See Output 79.19.4 for an example of this plot.

You can specify the following calibration-options:

displays plots of DFBETAS versus the case (observation) number. This option is available only for binary and binomial response models. This displays the statistics that are generated by the DFBETAS=_ALL_ option in the OUTPUT statement. The UNPACK option displays the plots separately. See Output 79.6.5 for an example of this plot.

displays plots of DIFCHISQ and DIFDEV versus the predicted event probability, and displays the markers according to the value of the confidence interval displacement C. This option is available only for binary and binomial response models. See Output 79.6.8 for an example of this plot. You can specify the following dpc-options:

displays and enhances the effect plots for the model. For more information about effect plots and the available effect-options, see the section PLOTS=EFFECT Plots.

Note: The EFFECTPLOT statement provides much of the same functionality and more options for creating effect plots. See Outputs Model-Predicted Probabilities by Sex, Model-Predicted Probabilities, Model-Predicted Probabilities, Predicted Probability and 95% Prediction Limits, and Model-Predicted Probabilities for examples of effect plots.

displays index plots of RESCHI, RESDEV, leverage, confidence interval displacements C and CBar, DIFCHISQ, and DIFDEV. This option is available only for binary and binomial response models. These plots are produced by default when any plot-request is specified and the MAXPOINTS= cutoff is not exceeded. The UNPACK option displays the plots separately. The STDRES option also displays index plots of STDRESCHI, STDRESDEV, and RESLIK. See Outputs Residuals, Hat Matrix, and CI Displacement C and CI Displacement CBar, Change in Deviance and Pearson Chi-Square for examples of these plots.

displays plots of DIFCHISQ, DIFDEV, confidence interval displacement C, and the predicted probability versus the leverage. This option is available only for binary and binomial response models. The UNPACK option displays the plots separately. See Output 79.6.7 for an example of this plot.

suppresses all plots.

displays and enhances the odds ratio plots for the model. For more information about odds ratio plots and the available oddsratio-options, see the section Odds Ratio Plots. See Outputs Plot of Odds Ratios of Heat at Several Values of Soak,Plot of the ODDSRATIO Statement Results, Plot of Odds Ratios for Additive, and Plot of Odds Ratios for Style for examples of this plot.

displays plots of DIFCHISQ, DIFDEV, confidence interval displacement C, and leverage versus the predicted event probability. This option is available only for binary and binomial response models. The UNPACK option displays the plots separately. See Output 79.6.6 for an example of this plot.

displays the precision-recall (PR) curve for binary and binomial response models. If you specify a method in the SELECTION= option, then an overlaid plot of all the PR curves for each step of the selection process is displayed. If you specify ROC statements, then an overlaid plot of the model (or the selected model if you specify the SELECTION= option) and the ROC statement models is displayed. If you specify the OUTROC= option in a SCORE statement, then the PR curve for the scored data set is also displayed.

Because PR curves are remappings of the ROC curve (Davis and Goadrich 2006), the CROSSVALIDATE, EPS, ID, WEIGHTED, METHOD=, THINX=, THINY=, and OPTIMAL suboptions in the ROCOPTIONS option are also applied to PR curves. The PRIOREVENT= and PRIOR= options in the SCORE statement apply to the curves produced by that statement.

The STEPLEN= option specifies the length of a step along the ROC curve, which is used to determine extra points at which the PR curve is computed. For more information, see the section Precision-Recall Curves.

displays the ROC curve. This option is available only for binary and binomial response models. If you also specify a SELECTION= method, then an overlaid plot of all the ROC curves for each step of the selection process is displayed. If you specify ROC statements, then an overlaid plot of the model (or the selected model if a SELECTION= method is specified) and the ROC statement models is displayed. If you specify the OUTROC= option in a SCORE statement, then the ROC curve for the scored data set is displayed.

The ID= option labels certain points on the ROC curve. Typically, the labeled points are closest to the upper left corner of the plot, and points directly below or to the right of a labeled point are suppressed. This option is identical to, and has the same keywords as, the ID= suboption of the ROCOPTIONS option.

You can define the following macro variables to modify the labels and titles on the graphic:

To revert to the default labels and titles, you can specify the macro variables in a %SYMDEL statement. For example:

   %let _ROC_ENTRYTITLE=New Title;
   Submit PROC LOGISTIC statement
   %symdel _ROC_ENTRYTITLE;

See Output 79.8.3 and Example 79.8 for examples of these ROC plots.

sets the significance level for creating confidence limits of the areas and the pairwise differences. The ALPHA= value that is specified in the PROC LOGISTIC statement is the default. If neither ALPHA= value is specified, then ALPHA=0.05 by default.

displays a table of the area under the empirical ROC curve, the area under the convex hull of the empirical ROC curve when METHOD=CONVEXHULL, the area under the binormal ROC curve when METHOD=BINORMAL, the area under the lower ROC curve when METHOD=LOWER, and the area under the PR curve when PLOTS=PR.

uses cross validated predicted probabilities instead of the model-predicted probabilities for all ROC and area under the ROC curve (AUC) computations and for all PR curve and area computations; for more information, see the section Classification Table. The cross validated probabilities are also used in computations for the "Association of Predicted Probabilities and Observed Responses" table. If you use a SCORE statement, then the OUTROC= data set and the AUC statistic from the FITSTAT option use the cross validated probabilities only when you score the original data set; otherwise, the model-predicted probabilities are used.

is an alias for the ROCEPS= option in the MODEL statement; however, if you also specify that ROCEPS= value, then this option is ignored. This value is used to determine which predicted probabilities are equal. The default value is the square root of the machine epsilon, which is about 1E–8.

displays labels on certain points on the individual ROC and PR curves and also on the SCORE statement’s ROC and PR curves. This option overrides the ID= suboption of the PLOTS=ROC option. If several observations lie at the same place on the curve, the value for the last observation is displayed. If you specify the ID option with no keyword, any variables that are listed in the ID statement are used. If no ID statement is specified, the observation number is displayed. The following keywords are available:

The SENSIT, 1MSPEC, POSPRED, and NEGPRED statistics are defined in the section Receiver Operating Characteristic Curves. The misclassification rate is the number of events that are predicted as nonevents and the number of nonevents that are predicted as events as calculated by using the given cutpoint (predicted probability) divided by the number of observations. If the PEVENT= option is also specified, then POSPRED and NEGPRED are computed using the first PEVENT= value and Bayes’ theorem, as discussed in the section Positive Predictive Values, Negative Predictive Values, and Correct Classification Rates Using Bayes’ Theorem.

specifies the method that is used to estimate the ROC and PR curves. By default, the empirical curves are produced as described in the section Receiver Operating Characteristic Curves. This is the area estimation method used for ROC comparisons.

The LOWER method modifies the empirical curve by always using a step function created by stepping up at the observed thresholds. This provides a more conservative estimate of the true ROC curve than the empirical estimate.

The CONVEXHULL method modifies the empirical curve by skipping thresholds that could be improved by replacing the observations with a weighted coin flip; it is sometimes called the attainable curve. The convex hull always has a larger AUC than the empirical curve.

The BINORMAL method assumes that the predicted values for your events and for your nonevents both have normal distributions (Metz 1978). If your events are distributed as and your nonevents are distributed as , and , then the binormal curve is given by

where

The area under the binormal ROC curve is

Although the binormal curve is smoother than the empirical curves, it can easily extend below the diagonal and might not be a good fit to your data. The METHOD= option is not available when you specify ROC statements or the ROCCI option, because the required standard errors are based on the empirical ROC curves.

suppresses the display of the model fitting information for the models specified in the ROC statements.

marks observed thresholds on the ROC and PR curves that optimize the given statistics, and adds these statistics to the OUTROC= data set. Note that there can be more than one threshold that optimizes the statistic. You can specify more than one optimal statistic, and if the same threshold is optimal for several statistics, the marks are jittered. In the following list, let P be the prevalence of the events in the population; by default, the prevalence is the proportion of events in the training data set , where is the frequency of events and n is the total frequency. However, you can change this value by specifying the PEVENT= option in the MODEL statement or the PRIOR= or PRIOREVENT= option in the SCORE statement. For each threshold on the ROC curve, denote the true positive fraction (sensitivity, recall) for that threshold as , the false positive fraction (1 – specificity) as , the true negative fraction (specificity) as , the precision (positive predictive value) as , and the negative predictive value as . The following statistics are available; the default is OPTIMAL=YOUDEN:

is an alias for the OUTROC= option in the MODEL statement.

suppresses labels on the ROC curve for any threshold that is too close to the preceding threshold’s label. Let X = sensitivity and Y = 1 – specificity; suppose the point is labeled; and let the next point be . If both and , then is not labeled. Similarly, this option suppresses labels on the PR curve for any threshold that is too close to the preceding threshold’s label, where Y = precision and X = recall. Optimal threshold labels are never thinned. By default, THINX=0.05.

suppresses labels on the ROC and PR curves for any threshold that is too close to the preceding threshold’s label. See the THINX= option for details. Optimal threshold labels are never thinned. By default, THINY=0.05.

uses frequency weight in the ROC and PR computations (Izrael et al. 2002) instead of just frequency. Typically, weights are considered in the fit of the model only, and hence they are accounted for in the parameter estimates. The "Association of Predicted Probabilities and Observed Responses" table uses frequency (unless the BINWIDTH=0 option is also specified in the MODEL statement) and is suppressed when ROC comparisons are performed. This option also affects computations of the ROC and PR curves performed by the SCORE statement, as well as computations of all areas under the curves.