LOGSELECT Procedure

specifies a global significance level for the construction of confidence intervals. The confidence level is 1–number. The value of number must be between 0 and 1. You can override the global specification by specifying the ALPHA= option in the OUTPUT statement. By default, ALPHA=0.05.

uses group and order information from the input data table. You can add this information to the input table by using the partition action in the table action set. For more information, see the section The APPLYROWORDER Option in Chapter 2, Shared Concepts. By default, the APPLYROWORDER option is not enabled. If you specify this option but the table does not contain group and order information, the procedure terminates with an error.

Note: You can use group and order information for a repeated measures analysis if the group information meets certain criteria. For more information, see the section Using a Preexisting Table Order.

displays measures of association between predicted probabilities and observed responses for binary, binomial, and ordinal response models. These measures assess the predictive ability of the model. The displayed statistics are the concordance index c (the area under the ROC curve, AUC), Somers’ D statistic (Gini’s coefficient), the Goodman-Kruskal gamma statistic, and Kendall’s tau-a statistic. These statistics are based on the number of pairs of observations that have different response values, the number of concordant pairs, and the number of discordant pairs. For more information, see the section Association Statistics.

specifies the precision to use for the association, classification, and LACKFIT option computations. The predicted probabilities or scores are rounded to the nearest multiple of number, which effectively bins and sorts the observations. You can specify a number between 0 and 1; if number does not evenly divide the unit interval, then it is reduced to a valid value. By default, BINEPS=1E–5. For more information, see the section Classification Table and ROC Curves.

displays Wald confidence limits for each of the parameter estimates. If you specify CLB=PL, then profile-likelihood confidence limits are also displayed. The confidence level is 0.95 by default; you can change it by specifying the ALPHA= option. This option is not available when you use either LASSO selection or elastic net selection.

creates the "Parameter Estimates Correlation Matrix" table. The correlation matrix is computed by normalizing the covariance matrix . That is, if is an element of , then the corresponding element of the correlation matrix is , where . The option is not available when you use either LASSO selection or elastic net selection.

creates the "Parameter Estimates Covariance Matrix" table. The covariance matrix is computed as the inverse of the negative of the matrix of second derivatives of the log-likelihood function with respect to the model parameters (the Hessian matrix). The option is not available when you use either LASSO selection or elastic net selection.

displays a table for binary or binomial response models that, for a set of probabilities or cutpoints, contains the frequencies of observations that are correctly and incorrectly classified as events and nonevents, and optionally the sensitivity, the 1–specificity, the positive and negative predictive values, the correct classification rate, the error rate, and the lift.

Because the number of cutpoints that are generated from your data can be very large, you can write the classification table to an output data table, or you can specify your own list of cutpoints.

If you specify the PRIOR= option, then the reported predicted probabilities are unchanged, but the statistics that are specified by the keywords PPV, NPV, PC, ACCURACY, MISCLASS, and LIFT are modified.

If you specify a PARTITION statement, then a table is created for each role. If you do not specify the PRIOR= option, then the training proportions are used as the prevalences for computing the statistics that are specified by the keywords PPV, NPV, PC, ACCURACY, MISCLASS, and LIFT for the validation and test data.

You can specify the following options. For more information, see the section Classification Table and ROC Curves.

names the input data table for PROC LOGSELECT to use. The default is the most recently created data table. libref.data-table is a two-level name, where

declares the DATA= table to be the same input data table that is used for building the model, when you also specify the RESTORE= option. The FITDATA option enables you to specify the PARTFIT option to produce more fit statistics, to compute all the statistics from the OUTPUT statement that are available for your model, and to group computations according to the partition roles.

names an input data table that contains starting values. libref.data-table is a two-level name, where libref refers to the library, and data-table specifies the name of the output data table. For more information about this two-level name, see the DATA= option and the section Using CAS Sessions and CAS Engine Librefs.

The input data table must contain an Estimate column and a ParmName column, and it can optionally contain your BY variables. You can create a version of this table by using a DISPLAYOUT statement to produce the "Parameter Estimates" table from a PROC LOGSELECT program. For example:

displayout parameterestimates=pe;

You can then modify the parameter estimates in this table and copy them back in by using the following statement:

proc logselect inparmest=mylib.pe;

generates the "Iteration History" table.

performs the Hosmer-Lemeshow goodness-of-fit test for binary and polytomous response models (Hosmer and Lemeshow 2000; Fagerland, Hosmer, and Bofin 2008; Fagerland and Hosmer 2013, 2016). The subjects are divided into at most G groups of approximately the same size. You can specify G as any integer greater than or equal to 5; by default, G = 10. The discrepancies between the observed and expected numbers of observations in these groups are summarized by the Pearson chi-square statistic, which is then compared to a chi-square distribution. The number of degrees of freedom that PROC LOGSELECT uses for this test depends on the specified model; you can override this number by specifying a nonnegative integer, df, and by modifying the reduction, r. A small p-value suggests that the fitted model is not an adequate model. For more information, see the section The Hosmer-Lemeshow Goodness-of-Fit Test.

The following options are available:

specifies the base regularization parameter for the LASSO model selection method. The regularization parameter for step i is r. By default, LASSORHO=0.8.

specifies the maximum number of steps for LASSO model selection. By default, LASSOSTEPS=20.

specifies the convergence tolerance for the optimization algorithm that solves for the LASSO parameter estimates at each step of LASSO model selection. By default, LASSOTOL=1E–6.

specifies the maximum number of response levels that are allowed when you have a polytomous response variable. By default, MAXRESPONSELEVELS=100.

specifies when to levelize the data—that is, when to generate the design rows for your model from the data. The default, MULTIPASS=FALSE, levelizes the data once and stores them in a temporary data table, then rereads the data from that table as needed; this method is fast but uses more memory. MULTIPASS=TRUE levelizes the data every time they are read; this method is slower but uses less memory. The MULTIPASS= option does not affect how the OUTPUT statement is processed. When MULTIPASS=TRUE, you can also use the PAGEOBS= option to improve performance.

disables the checking process that determines whether maximum likelihood estimates of the regression parameters exist. For more information, see the section Existence of Maximum Likelihood Estimates.

suppresses the display of the "Class Level Information" table if you do not specify number. If you specify number, the values of the classification variables are displayed for only those variables whose number of levels is less than number. Specifying number helps to reduce the size of the "Class Level Information" table if some classification variables have a large number of levels.

suppresses computation of the covariance matrix and the standard errors of the regression coefficients. When the model contains many variables (thousands), the inversion of the Hessian matrix to derive the covariance matrix and the standard errors of the regression coefficients can be time-consuming. The CORRB, COVB, and TYPE3 options, and the OUTPUT statement options for confidence limits, are not available when the NOSTDERR option is specified. This option also disables the quasi-complete separation check; for more information, see the section Existence of Maximum Likelihood Estimates.

suppresses Hessian and computations and invokes the NOSTDERR option. When the model contains many variables (thousands), computing the Hessian matrix during model fitting can be time-consuming. The CLB, CORRB, COVB, and TYPE3 options, the REPEATED statement, the OUTPUT statement options that rely on the covariance, and confidence limits for the parameters and for the ODDSRATIO, OUTPUT, RISK, and LSMEANS statements are not available when you specify the NOXPX option. The forward, backward, and stepwise selection methods are not available; however, you can specify the LASSO and elastic net methods. This option invokes the TECHNIQUE=LBFGS option by default, and the NEWRAP, NRRIDG, and TRUREG optimization techniques are not available. Because the method of identifying linearly dependent variables relies on the matrix, some parameters that should be assigned 0 degrees of freedom can be missed. This option also disables the quasi-complete separation check; for more information, see the section Existence of Maximum Likelihood Estimates.

specifies the maximum number of observations to be included in a batch. During the optimization, the LOGSELECT procedure reads at most number observations from the data table into memory, performs the appropriate log-likelihood, gradient, and Hessian computations on that batch of observations, then discards those observations and reads in the next batch of data for processing. Generally, a smaller number decreases memory usage but might lead to longer computation times, whereas a larger number might lead to shorter computation times but increases memory usage. The default PAGEOBS=AUTO option determines whether the entire data table can be conveniently held in memory; if it cannot, then number is set to 256. The PAGEOBS= option has the most benefit when you also specify the MULTIPASS=TRUE option; however, it can adversely affect performance when the default MULTIPASS=FALSE option is in effect. For more information, see the section Memory Usage.

displays fit statistics in the "Fit Statistics" table that are usually produced when your data are partitioned. This option is not required when you specify a PARTITION statement.

The additional statistics include the R-square, the adjusted R-square, McFadden’s R-square, the average square error (Brier score), the misclassification rate, the null-model log likelihood, and the difference in means.

controls the convergence criterion for computing profile-likelihood confidence intervals. The value is a positive number. By default, PLCONV=1E–4. This option has no effect if you do not request profile-likelihood confidence intervals.

specifies the maximum number of iterations for profile-likelihood confidence interval searches. By default, PLMAXITER=25. If you do not reach convergence in number iterations, the confidence limit is set to missing. This option has no effect if you do not request profile-likelihood confidence intervals.

specifies the tolerance for testing the singularity of the Hessian matrices that are computed during the profile-likelihood confidence interval searches. The test requires that a pivot for sweeping this matrix be at least the value times a norm of the matrix, where value is greater than 0. By default, value is the machine epsilon times 1E7, which is approximately 1E–9. This option has no effect if you do not request profile-likelihood confidence intervals.

specifies the name of the analytic store that contains the context and results of a model that is selected and stored from a previous statistical analysis. The analytic store is created by a STORE statement from a previous PROC LOGSELECT call or by a store parameter that is specified in a previous logistic action call.

The displayed output can include the following content:

See Example 18.7 for further discussion of these outputs.

You can specify the STB and CLB options in the PROC LOGSELECT statement to modify the "Parameter Estimates" table; you can specify the CORRB, COVB, and TYPE3 options to display those tables; you can specify the CODE statement to produce SAS code for scoring; you can specify LSMEANS statements to compute least squares means; and you can specify ODDSRATIO statements to compute odds ratios. If you specify a DATA= data table, then you can also specify the ASSOCIATION, CTABLE, LACKFIT, and PARTFIT options, and you can score that data table by specifying the OUTPUT statement.

If the analytic store is created using BY-group processing, then the results are displayed according to each BY group. If you specify a DATA= input data table, then you must also specify the same BY statement that was used to create the store.

If a PARTITION statement is used to create the stored model, then the Replay section displays only information that is computed from the original training data. For the Restore section, if you create the partition for the stored model by using a ROLE variable and also specify the FITDATA option, then the ROLE variable must exist in the DATA= table and the Restore section displays results according to the roles. However, if the partitions for the stored model are created using a FRACTION option, then the roles are ignored because the roles cannot be reliably re-created; this means that output statistics that rely on the input data are not computed.

Because the selected model from a previous analysis is stored, statements and options specific to model specification, selection, and optimization are not available. In particular, the CLASS, EFFECT, FREQ, MODEL, PARTITION, REPEATED, SELECTION, STORE, and WEIGHT statements are not available to be used with the RESTORE= option, and the MULTIPASS= option is always set to TRUE.

libref.data-table is a two-level name, where libref refers to the library, and data-table specifies the name of the input data table. For more information about this two-level name, see the DATA= option and the section Using CAS Sessions and CAS Engine Librefs.

specifies an integer to be used to start the pseudorandom number generator for the LSMEANS statement’s ADJUST=SIMULATE option and for the PARTITION statement’s FRACTION option. If you do not specify a seed, or if you specify a number less than or equal to 0, the seed is generated by reading the time of day from the computer’s clock.

displays the standardized estimates of the parameters in the "Parameter Estimates" table. The standardized estimate of is given by , where is the total sample standard deviation for the ith explanatory variable and

The sample standard deviations for parameters that are associated with CLASS variables are computed using their codings. The standardized estimates are not computed for the intercept parameters.

computes Wald statistics for Type 3 contrasts for each effect that you specify in the MODEL statement. This option is not available when you specify a REPEATED statement. The option is also not available when you use either LASSO selection or elastic net selection. For more information, see the section Joint Tests and Type 3 Tests.

continues to perform computations by using the last iteration of the optimization when the optimization fails. By default, computations cease when a failure occurs.