TMODEL Procedure

PROC TMODEL includes changes to the underlying computational algorithms that are used in the majority of PROC MODEL analyses. The new algorithms improve the stability and convergence characteristics along with the computational efficiency for most problems; however, for some problems these improvements can cause PROC TMODEL to produce different results than PROC MODEL. In particular, both estimation and simulation tasks rely on matrix ordering and factorization algorithms that have been enhanced in PROC TMODEL to work more efficiently, especially for large problems. Also, PROC TMODEL processes input data in a different order than PROC MODEL to improve performance, and this can cause some estimation tasks to produce different results.

In addition to performance improvements, PROC TMODEL has the following new features:

The ability to specify cross-sectional variables in PROC TMODEL allows for the estimation of dynamic models by using multiple time series. By contrast, PROC MODEL can estimate dynamic models only for data that contain a single time series. Also, PROC TMODEL enhances the modeling capabilities of PROC MODEL by supporting models of the correlations among cross sections through the specification of random-effects parameters.

Models that depend on highly nonlinear parameters can cause the estimation process either to converge slowly or to fail to converge. PROC TMODEL includes two new features that address these problems. In PROC MODEL, a first-order approximation of the model problem’s Hessian matrix is used in the parameter search. PROC TMODEL has the option to use exact Hessian matrix values in the parameter search. PROC TMODEL also supports the use of an alternative nonlinear programming solver that improves convergence characteristics for many estimation problems.

PROC TMODEL breaks computationally intensive operations into multiple, concurrent threads to reduce the time it takes to complete many of the estimation and simulation tasks available in PROC MODEL. PROC MODEL performs all calculations sequentially. PROC TMODEL can break calculations up in the following ways:

For estimation tasks that do not involve dynamic models, PROC TMODEL breaks up the observed data into partitions and computes each partition’s contribution to the estimation of parameters concurrently. When PROC TMODEL analyzes a model for many BY groups, the BY groups can be analyzed concurrently. In Monte Carlo simulations, the random perturbations of the model variables can be evaluated concurrently in PROC TMODEL. In problems that involve a numerical optimization, it is sometimes necessary to perform many local optimizations by using separate initial estimates in order to determine a global solution. PROC TMODEL can perform these local optimizations concurrently to find the global solution more quickly than PROC MODEL, which solves local optimization problems sequentially.

PROC MODEL Features Not Available in PROC TMODEL

The following features in PROC MODEL are not currently available in PROC TMODEL:

• some features in the model file used by the OUTMODEL= and MODEL= options

• BY groups in the SDATA= and ESTDATA= data sets

• covariance matrices in output data sets

• the OUTSUSED= data set

• some diagnostic information in output tables

• Durbin-Watson autocorrelation statistics

• some features in the CMP system, such as the RUN_MACRO function

Model specifications in either PROC MODEL or PROC TMODEL can be saved to a file for use in subsequent PROC MODEL or PROC TMODEL operations; however, there are some limitations to the model files that can be shared between PROC MODEL and PROC TMODEL. The specification of instrumental variables that is saved to a model file in PROC MODEL or PROC TMODEL cannot be used in PROC TMODEL or PROC MODEL estimation tasks, respectively.

Monte Carlo simulations in PROC MODEL can apply a different error covariance matrix, different parameter values, and a different parameter covariance matrix for each BY group in the DATA= data set through the specification of corresponding BY groups in the SDATA= and ESTDATA= data sets. Also, in PROC MODEL estimations, different initial parameter estimates for each BY group can be specified in the ESTDATA= data set. In contrast, PROC TMODEL does not currently support the use of BY groups in either the SDATA= or ESTDATA= data set. All BY groups share the same covariance matrices and parameter value specifications. PROC TMODEL will support BY groups in the SDATA= and ESTDATA= data sets in a future release.

PROC MODEL adds perturbations to parameter values during Monte Carlo simulations when a parameter covariance matrix is specified in the ESTDATA= data set. PROC TMODEL does not currently support perturbation of parameters, but this functionality will be available in a future release.

PROC MODEL stores both parameter estimates and parameter covariance matrices in the OUTEST= data set when the OUTCOV option is specified. PROC TMODEL does not support the output of covariance matrices in the OUTEST= data set; however, the COV option, which prints parameter covariance matrices, is supported.

Many tables that PROC MODEL produces are not available in PROC TMODEL because they fall into one or more of the following categories:

Not available

report the intermediate states of PROC MODEL calculations that are not available or are computed differently in PROC TMODEL.

Replaced

are replaced by equivalent diagnostic output in PROC TMODEL in a different format.

Future

are not yet available in PROC TMODEL but are planned for a future release.

The reason that each table is not available in PROC TMODEL is summarized in Table 1.

Table 1: ODS Tables Not Available in PROC TMODEL

ODS Table Name	Description	Reason1
ODS Tables Created by the FIT Statement
AugGMMCovariance	Crossproducts matrix	F
ConfInterval	Profile likelihood confidence intervals	F
Crossproducts	Crossproducts matrix	F
DatasetOptions	Data sets used	F
DetResidCov	Determinant of the residuals	F
DWTest	Durbin-Watson test	F
EstSummaryMiss	Summary statistics for PAIRWISE option	F
GMMCovariance	Crossproducts matrix	F
GMMTestStats	GMM test statistics	F
Godfrey	Godfrey’s serial correlation test	F
HausmanTest	Hausman’s test table	F
InvXPXMat	XX inverse for system	N
IterInfo	Iteration printing	N R
ObsSummary	Identifies observations that contain errors	R
ObsUsed	Observations read, used, and missing	R
ParmChange	Parameter change vector	N
SizeInfo	Storage requirement for estimation	F
XPXMat	XX for system	N
YkVector	Marquardt iteration vector	N
ODS Tables Created by the SOLVE Statement
DatasetOptions	Data sets used	F
DescriptiveStatistics	Descriptive statistics	F
FitStatistics	Fit statistics for simulation	F
ObsSummary	Simulation trace output	R
ObsUsed	Observations read, used, and missing	R
SolutionVarList	Solution variable lists	F
TheilRelStats	Theil relative change error statistics	F
TheilStats	Theil forecast error statistics	F
ErrorVec	Iteration error vector	N
ResidualValues	Iteration residual values	N
PredictedValues	Iteration predicted values	N
SolutionValues	Iteration solved for variable values	N
ODS Tables Created by the FIT and SOLVE Statements
AdjacencyMatrix	Adjacency graph	R
CodeDependency	Variable cross reference	N
CodeList	Listing of compiled program code	N
CrossReference	Cross-reference listing for program	N
DepStructure	Dependency structure of the system	N
FirstDerivatives	First derivative table	N
IterIntg	Integration iteration output	N
MemUsage	Memory usage statistics	F
MissingDependencies	Missing values by dependency	F
MissingObservations	Missing values by observation	F
MissingSymbols	Missing values by symbol	F
ParmReadIn	Parameter estimates read in	F
SortAdjacencyMatrix	Sorted adjacency graph	R
TransitiveClosure	Transitive closure graph	R
1N - Not available, R - Replaced, F - Future

You can find information in PROC TMODEL for the PROC MODEL tables in category R as follows:

• IterInfo displays iteration information for the optimization process in the log for the ORMP optimizer.

• ObsSummary displays diagnostic information for observations that produce missing values in the log.

• ObsUsed displays observation counts in the EstSummaryStats table.

• AdjacencyMatrix, SortAdjacencyMatrix, TransitiveClosure have dependency information that you display by using the ANALYZEDEPS= option in the SOLVE statement.