What’s New in SAS/STAT 12.1

The ADAPTIVEREG procedure fits multivariate adaptive regression splines. The method is a nonparametric regression technique that combines both regression splines and model selection methods. It constructs spline basis functions in an adaptive way by automatically selecting appropriate knot values for different variables. The procedure performs model reduction by applying model selection techniques. Thus, the ADAPTIVEREG procedure is both a nonparametric regression procedure and a predictive modeling procedure.

The ADAPTIVEREG procedure supports models with classification variables, and it provides options for improving modeling speed. PROC ADAPTIVEREG extends the method to data with response variables that are distributed in the exponential family, including the binomial, Poisson, negative binomial, gamma, and inverse Gaussian distributions. PROC ADAPTIVEREG is multithreaded, enabling it to take advantage of multiple processors.

The QUANTLIFE procedure implements two quantile regression approaches that have been developed to account for right-censoring and provide valid estimates. Portnoy (2003) proposed a method to estimate conditional quantile functions from survival data by generalizing the idea of the Kaplan-Meier estimator of the survival function, and Peng and Huang (2008) developed a quantile regression approach motivated by the Nelson-Aalen estimator of the cumulative hazard function. Both methods are implemented with linear programming algorithms in the QUANTLIFE procedure. Like the standard quantile regression method for uncensored data, these two methods are distribution-free and are applicable to heteroscedastic data.

The QUANTLIFE procedure produces survival plots, conditional quantile plots, and quantile process plots. It performs semiparametric quantile regression when you specify spline effects. PROC QUANTLIFE takes advantage of parallel computing when multiple processors are available.

The QUANTSELECT procedure performs model selection for linear quantile regression. It provides capabilities similar to those offered by the GLMSELECT procedure (provides model selection for univariate linear models) including forward, backward, stepwise, LASSO, and adaptive LASSO selection methods. You can specify criteria such as AIC, SBC, AICC, adjusted R1, and significance levels to compare the fit of models, to determine when to stop the model selection process, and to choose the final selection model. The QUANTSELECT procedure supports constructed effects (such as smoothing splines and polynomials) and the SPLIT option for its CLASS statement. PROC QUANTREG produces parameter estimates progression plots and a variety of criterion progression plots.

PROC QUANTSELECT supports constructed effects such as regression spline, and it enables you to partition your data into training, validation, and testing roles. It is also multithreaded so that it can take advantage of multiple processors.

PROC QUANTSELECT is very efficient and can handle hundreds of variables and thousands of observations.

The STDRATE procedure computes direct and indirect standardized rates and risks for study populations. With direct standardization, you compute the weighted average of stratum-specific estimates in the study population, using weights such as population-time from a standard or reference population. With indirect standardization, you compute the weighted average of stratum-specific estimates in the reference population by using weights from the study population. The procedure provides summary statistics such as rate and risk estimates (and their confidence limits) for each stratum, as well as graphs.

Postprocessing macros are now available in the SAS autocall library. These macros duplicate the corresponding summary and diagnostic capabilities provided by the MCMC procedure and can be used with any SAS data set that contains posterior samples. These macros are documented with the MCMC procedure.

The %CIF macro implements nonparametric methods for estimating cumulative incidence functions with competing risks data. The macro can also be used to test the hypothesis that cumulative incidence functions are identical across groups. The %CIF macro is available from the SAS autocall library.

The CALIS procedure now provides the following features:

The CLUSTER option in the EFFECTPLOT statement modifies the INTERACTION plot-type by displaying the levels of the SLICEBY= effect in a side-by-side fashion. The CONNECT option in the EFFECTPLOT statement modifies the BOX and INTERACTION plot-types by connecting the predicted values with a line.

The FMM procedure is now production and adds the truncated exponential, truncated lognormal, truncated negative binomial, and truncated normal distributions.

The Agresti-Caffo and Miettinen-Nurminen confidence limits for the risk (proportion) difference are now available, and any confidence limit type can now be displayed in the risk difference plot. You can also request a continuity correction for the Wilson confidence limits for a binomial proportion.

The PLCORR option in the TEST statement produces Wald and likelihood ratio tests for the polychoric correlation coefficient.

The new DF= chi-square option specifies or adjusts the degrees of freedom for chi-square tests. The TESTF= chi-square option now permits you to provide null frequencies for a one-way chi-square test by using a secondary input data set. Similarly, the TESTP= chi-square option now permits you to provide null proportions by using a secondary input data set.

The LRCHISQ chi-square option in the TABLES statement produces a likelihood ratio chi-square test for one-way tables. This test can be based on a null hypothesis of equal proportions, specified proportions, or specified frequencies. The LRCHISQ option in the EXACT statement produces an exact likelihood ratio chi-square test for one-way tables.

The EXACT statement can now produce Barnard’s exact unconditional test for the risk difference as well as an exact likelihood ratio goodness-of-fit test.

The MAXLEVELS=n option in the TABLES statement specifies the maximum number of variable levels to display in one-way frequency tables; it also applies to one-way frequency plots.

The CROSSLIST(STDRES) option in the TABLES statement displays standardized residuals in the CROSSLIST table for two-way crosstabulation.

Mosaic plots are now produced with the PLOTS=MOSAICPLOT option. The GROUPBY= plot option specifies the primary grouping for two-way frequency plots. The new TWOWAY=CLUSTER plot option provides a cluster layout for two-way frequency plots that are displayed as bar charts.

The new DDFM=KENWARDROGER2 option applies the (prediction) standard error and degrees-of-freedom correction that are detailed by Kenward and Roger (2009). This correction further reduces the precision estimator bias for the fixed and random effects under nonlinear covariance structures.

The odds ratio, relative risk, and kappa plots now display the common (overall) statistic in addition to the statistics for each two-way table (stratum).

The LIFEREG procedure now supports numerous postprocessing statements including the ESTIMATE, EFFECTPLOT, LSMEANS, LSMESTIMATE, SLICE, STORE, and TEST statements.

The LIFETEST procedure now supports a WEIGHT statement. For survival plot enhancements, the MAXLEN= option of the ATRISK option specifies a maximum number of characters that can be used in displaying stratum labels, and the OUTSIDE option of the ATRISK option specifies that the at-risk table be drawn outside the plot area. If you assign a label to a strata variable, that label is used in all tables and graphs.

The LOESS procedure now supports an OUTPUT statement.

PROC LOGISTIC provides partial proportional odds logistic regression with the UNEQUALSLOPES option in the MODEL statement. The ESTIMATE, LSMEANS, LSMESTIMATE, SLICE, and STORE statements can now be used for a stratified analysis. The PCORR option in the MODEL statement computes the partial correlation statistic for each model parameter (excluding the intercept).

The ID statement specifies variables in the DATA= data set that are used for labeling ROC curves and influence diagnostic plots.

The NLOPTIONS statement controls the optimization process for conditional analyses (specified with a STRATA statement) and for constrained optimization (specified with the UNEQUALSLOPES option in the MODEL statement).

The EFFECTPLOT statement and the PLOTS=EFFECT option have two new options for displaying plots with a CLASS effect on the X axis. The CLUSTER option displays the levels of the SLICEBY= effect in a side-by-side fashion. The CONNECT option connects the predicted values with a line.

The MCMC procedure provides the following new capabilities:

The ID statement names one or more variables for identifying observations in the output and in the plots. In addition, the MANHATTAN plot option generates a Manhattan plot.

The PROFILE statement selects parameters for profiling for the assessment of their nonlinear characteristics. It can also gauge the influence of each observation on the selected parameters.

The following options are added to the PROC NPAR1WAY statement:

You can use the new ODS GRAPHICS statement option BYLINE=FOOTNOTE or BYLINE=TITLE to display BY-group lines in a footnote or title in graphs.

The DIST= option in the RANDOM statement enables you to specify either a gamma or lognormal distribution for the shared frailty. Bayesian frailty models are now supported. The DISPERSIONPRIOR= option in the BAYES statement specifies the prior distribution of the dispersion parameter.

Fleming-Harrington estimates can be requested with the METHOD=FM option in the BASELINE statement and in the OUTPUT statement. The DIRADJ option in the BASELINE statement specifies direct adjusted survival curves.

You can specify the multiple correlation between the tested predictor and the covariates with the CORR= option in the LOGISTIC statement.

The PROBIT procedure now supports numerous postprocessing statements including the ESTIMATE, EFFECTPLOT, LSMEANS, LSMESTIMATE, SLICE, STORE, and TEST statements.

The QUANTREG procedure now supports the ESTIMATE statement. The EFFECT statement now supports the effect-types COLLECTION, LAG, MULTIMEMBER, and POLYNOMIAL in addition to SPLINE.

The REG procedure creates heat maps for residual and fit plots when the MAXPOINTS= threshold is exceeded.

The STDI= option in the OUTPUT statement specifies a variable to contain the estimates of the standard errors of the individual predicted values.

The MODEL=INPUTNEVENTS option in the SAMPLESIZE statement specifies the number of events from a fixed-sample study of survival data. There are two new INPUTEVENTS options for the sample size computation: ACCRUAL= and LOSS=. The ACCRUAL= option specifies the method for individual accrual. The LOSS= option specifies the individual loss to follow-up in the sample size computation.

The STOP=BOTH option in the DESIGN statement specifies the condition of early stopping for the design. The new BETABOUNDARY=BINDING suboption computes the Type I error probability with the acceptance boundary, and the new BETABOUNDARY=NONBINDING suboption computes the Type I error probability without the acceptance boundary.

The BETABOUNDARY option in the PROC SEQTEST statement specifies whether the boundary is used in the computation of the Type I error level . The BETABOUNDARY=BINDING option computes the Type I error probability with the (acceptance) boundary, and the BETABOUNDARY=NONBINDING suboption computes the Type I error probability without the boundary.

The SURVEYFREQ procedure now produces mosaic plots for crosstabulation tables. The GROUPBY= plot option specifies the primary grouping for two-way weighted frequency plots.

You can now estimate geometric means for finite populations with the GEOMEAN keyword in the PROC SURVEYMEANS statement. The new POSTSTRATA statement provides poststratification analysis.

The SERATIO and VARRATIO options in the MODEL statement compute the ratio of two standard errors for the regression coefficients and the ratio of two variances for the regression coefficients, respectively.

The STB option in the MODEL statement produces standardized regression coefficients.

The SURVEYSELECT procedure now provides Bernoulli and Poisson sampling.

If you assign a label to a strata variable, the procedure now uses the label instead of the variable name in all tables and graphs.

The appearance of the default bar chart for two-way frequency plots has changed. The row level labels have been moved outside the plot so that the row grouping appears less dominant.

For two-way dot plots (TYPE=DOTPLOT) in nonstacked layouts, the default positions of the row and column variables are reversed to group graph cells by the column variable. You can specify GROUPBY=ROW to group graph cells by the row variable.

Random-effects parameter names are constructed using the formatted values rather than the unformatted values.

If the MISSING= option is not specified, PROC MCMC samples all missing values (including partial missing in some cases) by default. Observations for which the procedure failed to identify proper sampling algorithms are discarded prior to the simulation. If the MISSING= option is explicitly specified (AC or CC), the option is honored.

PROC MCMC avoids performing an optimization prior to the start of the simulation if the only sampling algorithms used in the program are conjugate or direct.

PROC MCMC now permits a model specification that has only RANDOM and MODEL statements; PRIOR and PARMS statements are no longer required in that case.

By default, the AFDR and PFDR are constrained to be greater than or equal to the raw p-value. The UNRESTRICT option of the PROC MULTTEST statement’s AFDR and PFDR options estimates the AFDR and PFDR as defined in Benjamini and Hochberg (2000), which allows the adjustment to reduce the raw p-value.

PROC SURVEYSELECT now uses the Mersenne-Twister random number generator by default. In previous releases, PROC SURVEYSELECT used the RANUNI random number generator. To reproduce samples that PROC SURVEYSELECT selected in releases prior to SAS/STAT 12.1, you can use the RANUNI option with the SEED= option (for the same input data set and selection parameters).

PCs running Windows 64-bit operating systems can now allocate more than 2Gb of memory to fit a model. This change affects the GLIMMIX, GLM, HPMIXED, and MIXED procedures.