What’s New in SAS/ETS 12.1

TIMEDATA Procedure

SASEXFSD Interface Engine

AUTOREG Procedure

COUNTREG Procedure

MODEL Procedure

PANEL Procedure

QLIM Procedure

SASECRSP Interface Engine

SASEXCCM Interface Engine

SEVERITY Procedure

SSM Procedure

TCOUNTREG Procedure

X12 Procedure

The heteroscedasticity- and autocorrelation-consistent (HAC) covariance matrix estimator is supported, which consistently estimate the covariance matrix even when the heteroscedasticity and autocorrelation structure might be unknown or misspecified. Five types of kernel functions—Bartlett, Parzen, quadratic spectral, truncated, and Tukey-Hanning kernels—are supported. The bandwidth parameter can be estimated using the Andrews (1991) method, the Newey and West (1994) method, or a flexible equation based on sample size. The prewhitening feature and adjustment of degrees of freedom are supported. The well-known Newey-West estimator is also supported.

Multiple structural change tests proposed by Bai and Perron (1998) are supported. Specifically, these are the test of no break versus a fixed number of breaks ( test); the equal and unequal weighted versions of double maximum tests of no break versus an unknown number of breaks given some upper bound ( test and test); and the test of versus breaks ( test). The tests can be applied to both pure and partial structural change models. The p-value of each test, based on the simulation of the limiting distribution, and the confidence intervals of parameter estimators, including the break dates, are also provided. The constraints on the distribution of the errors and regressors across segments can be imposed. For estimating the covariance matrix the HAC estimator is supported.

The Shin cointegration tests with p-values are supported.

The p-values for the ERS optimal point unit root test, ERS DF-GLS unit root test, and KPSS unit root test are provided.

The status of ERS and Ng-Perron unit root tests changed from experimental to production.

A new variable selection method is provided. The greedy search method can be used with either forward selection or backward elimination. In each step, the AIC or BIC criterion is evaluated, and the selection continues until the selection criterion is met.

Multiple MODEL statements are supported. This enables multiple count models to be fitted under one PROC COUNTREG call.

The OPTIMIZE option has been added to the SOLVE statement to permit the simulation of models that include constraints on the solve variables in the model program’s system of equations. Upper and lower bounds on the solve variables can be imposed by using the BOUNDS statement, and linear or nonlinear constraints on functions of the solve variables can be imposed by using the RESTRICT statement. The OPTIMIZE option limits the solution space for simulations to the feasible region defined by constraints. When no feasible solution exists for a problem, information about how the constraints were violated are included in the OUT= data set if the OUTOBJVALS or OUTVIOLATIONS option is specified. The OPTIMIZE solution method computes constrained solutions by casting the simulation problem into a nonlinear optimization problem then solving the optimization problem.

Diagnostic reports that summarize the occurrence of missing values in both estimation(FIT) and simulation(SOLVE) steps have been added to the MODEL procedure. The new REPORTMISSINGS option generates tables that describe which variables in the model and which observations in the DATA= data set contribute missing values within FIT, or SOLVE calculations. The REPORTMISSINGS option produces output that is easier to interpret when debugging model and data specification problems than the ObsUsed table, which often lacks sufficient detail, or the PUT statement, which can produce too much output. The amount of diagnostic information that the REPORTMISSINGS tables include can be limited by using the MAXERRORS= option. The tables that the REPORTMISSINGS option produces can also attribute missing quantities in the model program to missing values of independent variables in the DATA= data set.

The ANALYZEDEP= option has been added to the MODEL procedure to provide more information on the nature of misspecification errors in simulations. When the system of equations specified in a SOLVE step does not consistently determine the solve variables, the system is partitioned into those equations that overdetermine, underdetermine, and consistently determine the solve variables. The partitioning of equations and solve variables is performed by using a Dulmage-Mendelsohn (Dulmage and Mendelsohn, 1958) decomposition of the system, which is invariant to the order in which equations and variables are specified. You can display the partitioning of the system graphically by using the BLOCK plot option in the ANALYZEDEP= option.

The BLOCK and DETAILS options for visualizing the dependency structure among equations and variables within a model program have been improved. General form equations can now be analyzed and incorporated in the dependency analysis. Also, you can produce a graphical representation of the dependence of equations on solve variables by using the DETAILS option in the ANALYZEDEP= option. The new dependency plot can display the relationship among many more equations and variables than was previously possible by using the DepStructure table. You can also customize the dependency plot to depict a subset of the equations and variables in the model by using the new EQGROUP and VARGROUP statements.

Three new copula options have been added to the MODEL procedure. Monte Carlo simulations can now use the CLAYTON, GUMBEL, and FRANK Archimedean copulas to specify the correlation structure among model equations in multivariate simulations.

The panel unit root tests have been added to test the hypothesis of a unit root. Several different specifications including six groups of deterministic variables, lag specifications, and kernel and bandwith specifications can be calculated for each test. The tests include the following:

Poolability tests for panel data models including F test and LR tests

The heteroscedasticity- and autocorrelation-consistent (HAC) covariance matrix estimator is supported, which consistently estimates the covariance matrix even when the heteroscedasticity and autocorrelation structure might be unknown or misspecified. Five types of kernel functions—Bartlett, Parzen, quadratic spectral, truncated, and Tukey-Hanning kernels—are supported. The bandwidth parameter can be estimated with the Andrews method, Newey-West method, and sample size based method, or a fixed value for the bandwidth can be provided. The prewhitening feature is also available with the HAC option. The well-known Newey-West estimator is also supported.

Bayesian Estimation Features. Most of the univariate models available in the QLIM procedure can be estimated in a Bayesian framework with the BAYES statement. The main features are as follows:

Heckman Selection Model – Two-Step Estimator. The QLIM procedure now supports Heckman’s two-step estimation method, as an alternative to maximum likelihood estimation of selection models. The standard errors of the second-step OLS estimates are corrected for consistency by default. However, if the uncorrected ones are requested for testing purposes, they are available with the UNCORRECTED option.

A new variable selection method. The greedy search method can be used with either forward selection or backward elimination. In each step, the AIC or BIC criterion is evaluated, and the selection continues until the selection criterion is met.

ODS Graphics plots for Bayesian and frequentist estimation methods. For the frequentist framework, the QLIM procedure can produce a graphical representation of the output that is produced with the OUTPUT statement. For the Bayesian approach, the QLIM procedure can produce the plots of the prior and the posterior predictive analysis.

Estimation algorithms have been modified to use multiple threads of execution in parallel, which enables PROC SEVERITY to fully utilize all the CPU cores of the machine where it is being run to complete the estimation tasks significantly faster.

A new plot, the Q-Q plot, has been added. You can request this plot by specifying the PLOTS=QQPLOT or PLOTS=ALL option in the PROC SEVERITY statement. For a distribution named dist, the quantile for a given value of the cumulative distribution function (CDF) is computed either by evaluating the dist _QUANTILE function, if it is defined for the distribution, or by inverting the dist _CDF function of the distribution.

Standard errors and confidence intervals are now available for the empirical distribution function (EDF) estimates. They are written to the OUTCDF= data set. If you specify the PLOTS=CDFPERDIST option, then the lower and upper confidence limits of EDF estimates are plotted in the CDFDistPlot plots. You can specify the confidence level for the confidence interval by specifying the new EDFALPHA= option in the PROC SEVERITY statement. For standard EDF estimators (no censoring or truncation), the standard errors are computed using the normal approximation. For Kaplan-Meier and modified Kaplan-Meier estimators (truncation with one type of censoring), Greenwood’s formula is used. For Turnbull’s estimator (both types of censoring with or without truncation), standard errors are computed from the estimate of the covariance matrix that is computed by inverting the Hessian matrix of Turnbull’s nonparametric log-likelihood. If the Hessian matrix is singular or results in missing values for the standard errors of any of the intervals, then the normal approximation method is used.

If you specify the SCALEMODEL statement, then the scale of the distribution depends on the values of regressors. For a given distribution family, each observation implies a different scaled version of the distribution. PROC SEVERITY needs to construct a single representative distribution from all such distributions in order to compute estimates of CDF and the probability density function (PDF) that are comparable across different distribution families. Prior to this release, the representative distribution was constructed as the weighted mixture of distributions implied by all observations. For that method, estimation of CDF or PDF for one observation requires computations, where denotes the total number of observations. So estimation of CDF or PDF for all observations requires computations, which can dominate the runtime of PROC SEVERITY even for moderately large values of . Starting with this release, you can specify the new DFMIXTURE= option in the SCALEMODEL statement to choose one of four methods to construct the representative mixture distribution. The prior method is used when you specify DFMIXTURE=FULL option. The default method is DFMIXTURE=MEAN, which uses a distribution with scale equal to the mean of scale values. It is significantly faster than the FULL method. The other two methods construct a mixture of distributions each with one of scale values, which are either the -quantiles from the sample of scale values (DFMIXTURE=QUANTILE) or the scale values implied by randomly chosen observations (DFMIXTURE=RANDOM). For , the QUANTILE and RANDOM methods can be significantly faster than the FULL method.

The DIST statement now supports two more keywords in addition to the _PREDEFINED_ keyword. If you specify the _USER_ keyword, then PROC SEVERITY includes all the custom distributions that you have defined in the libraries specified in the CMPLIB= system option. The _ALL_ keyword includes all the predefined distributions and your custom distributions. It also includes the Tweedie and scaled-Tweedie distributions that are not included by the _PREDEFINED_ keyword. The DIST statement also has two new options, LISTONLY and VALIDATEONLY. The LISTONLY option lists the names of the distributions that you have specified in the DIST statement and the distributions implied by any keywords that you specify. This option is especially useful in conjunction with the keywords. The VALIDATEONLY option validates all the specified distributions and writes the distribution’s information to the OUTMODELIFO= data set and a new ODS table, DistributionInfo. This option is especially useful in conjunction with your custom distributions, because it enables you to check whether the definitions of the functions and subroutines that make up your distribution satisfy PROC SEVERITY’s requirements.

A trend component that satisfies a two-factor (nonseasonal and seasonal) ARIMA(p,d,q)(P,D,Q)s model can be specified.

A state subsection that satisfies a first-order, vector ARMA model—VARMA(p,q) with and —can be specified.

Diagnostic plots are available for residual analysis and structural break analysis.

New printing options enable printing of series and component forecasts and smoothed estimates. In addition, you can print estimated system matrices.

A table that identifies extreme additive outliers is printed. Additionally, structural breaks that are associated with state shocks can also be printed.

A new option, MATCHPARM, in the TREND statement simplifies parameter specification when the CROSS= option is specified.

New options enable finer control over the nonlinear optimization of the likelihood in the parameter estimation phase.

ODS Graphics plots are provided. The TCOUNTREG procedure can produce plots of various important predictive functions as well as model diagnostics.

A new variable selection method is provided. The greedy search method can be used with either forward selection or backward elimination. In each step, the AIC or BIC criterion is evaluated, and the selection continues until the selection criterion is met.

The PICKMDL statement. The PICKMDL statement causes the X12 procedure to automatically select a regARIMA model from a list of candidate models defined by the user in the MDLINFOIN= data set. The METHOD= option in the PICKMDL statement controls how the model selection is performed. The selected regARIMA model then extends the time series prior to performing the X-12-ARIMA seasonal adjustment. The PICKMDL statement is experimental for this release.

The SEATSDECOMP statement. The SEATSDECOMP statement first computes the B1 series by using the X-12-ARIMA method and then performs a seasonal adjustment of the B1 series by using the SEATS decomposition method. SEATS is a polynomial-based seasonal decomposition method developed by Gomez and Maravall (1997a, 1997b). You can write the resulting components to a data set by specifying the OUT= option in the SEATSDECOMP statement. The SEATSDECOMP statement is experimental in this release.

The NOAPPLY option has been added as a general option to the REGRESSION statement. The NOAPPLY option specifies whether specific regression effects are to be included in the B1 series that is seasonally adjusted.

The AICTEST option has been added as a general option to the REGRESSION statement. The AICTEST option enables you to specify a regression effect, but the effect is not included in the regARIMA model unless the results of an AIC test determine that the effect should be included in the model. Thus, the AICTEST option can be used to automatically select regressors for the regARIMA model.

Dulmage, A. L. and Mendelsohn, N. F. (1958), “Coverings of Bipartite Graphs,” Canadian Journal of Mathematics, 10, 517–534.

Gomez, V. and Maravall, A. (1997a), Guide for Using the Programs TRAMO and SEATS, Beta Version, Banco de España.

Gomez, V. and Maravall, A. (1997b), Program TRAMO and SEATS: Instructions for the User, Beta Version, Banco de España.