H O M E / I N F O / P R O D U C T S / O R D E R / S U P P O R T

TSP / What is TSP? / Features / Requirements

What’s new in TSP? Versions 5.1 5.0 4.5 4.4 and earlier versions

VERSION 5.1

Oxmetrics compatibility

New linear programming command LP

ANALYZ for functions of series, improved output and options

Panel robust standard errors for PANEL, LSQ, ML, Probit

Generalized Impulse Response in VAR, plus improved plotting

Enhanced iteration in LAD and LMS, looking for multiple solutions

LogL added to LIML

FORM creates unnormalized equations

Enhancements to SORT

Speeded up FIML is some cases

New nonlinear option for stepsize, improving iteration behavior.

Circle graphs (where each data point’s importance is represented by circle size)

Greatly improved Excel spreadsheet reading – more versions, and multiple sheets

Reading of Stata dta files up to Version 10

Improved SHOW SERIES output.

Reading up to 500 characters per line interactively (used to be 80)

Allowing various end-of-line characters in batch files (for unix/MAC compatibility)

Labelling printed matrices

Decimal style dates for quarterly and monthly frequencies, allowing computation

More efficient long programs with loops

VERSION 5.0

Most of the new features in TSP 5.0 involve panel data estimation, but there are a wide variety of improvements, including many to the graphics routines, censored quantile estimation, simple kernel estimation and interval regression.

New commands:

• INTERVAL - Interval regression. This is like ordered probit, but the category bounds are known in advance.
• KERNEL - univariate kernel density estimation and simple kernel regression.
• LAD – Censored (and uncensored) quantile regression with bootstrapped standard errors
• TOBIT – general two-limit Tobit model
• LP – linear programming, using the Simplex method.

Improved panel data commands:

• PANEL (REI,REIT) - ML estimation of one-way and two-way random effects linear regression models. The effects in the two-way model are normally orthogonal (Individual and Time), but they can be partially or fully nested (such as State and Region). As usual in TSP, the data can be unbalanced, there is no limit on the number of individuals or time periods, and the standard errors are computed from analytic second derivatives. The new FEPRINT option prints the fixed effect intercepts and their standard errors for the WITHIN model. A summary table of different panel models estimated with ML (TOTAL, WITHIN, REI, REIT) is printed, showing the model which minimizes SBIC.
• PROBIT (REI,FEI) - ML estimation of one-way random effects or fixed effects Probit models. For random effects, the number of points in the Hermite quadrature can be specified (default 20).
• AR1 (REI,FEI) - ML estimation of one-way random effects or fixed effects AR(1) regression models. The fixed effects model can be estimated with exact ML (retain first observations) or conditional ML (drop first observations); for conditional ML a common factor test is automatically computed to compare with an unrestricted fixed effects regression with lagged dependent and independent variables. AR1 and AR1(FEI) also work on irregularly spaced data with conditional ML.
• 2SLS (FEI), LIML(FEI), linear SUR/3SLS/GMM/FIML (FEI) - fixed effects in 2SLS, LIML, and systems estimation. Normally the estimates of the actual fixed effects (intercepts) are not shown, but their values and SEs can be printed for 2SLS. Fixed effects are stored in the series @AI for further computations.
• Panel-grouped robust standard errors, via the HCOV=Q (QMLE) and HCOMEGA=BLOCK/DIAGONAL options. Only available in a few commands so far, such as OLSQ, 2SLS, PANEL, and PROBIT. But these SEs will be added to more commands in the future.

Improved graphics:

• GRAPH(LINE) and PLOT handle panel data, by inserting gaps between individuals.
• GRAPH (NOSORT,LINE) creates lines that are joined in observation order, rather than being sorted by X values.
• TSP/unix versions have graphics using gnuplot interface and .GIF output.
• HIST has graphical output (Windows/Givewin/Oxmetrics version only).
• BJIDENT, BJEST, BJFRCST, PLOTS (PREVIEW) (Windows/GiveWin/Oxmetrics version only) replaces the old text plots (differenced series, residuals, forecasts, CUSUM, CUSUMSQ) with graphics plots.
• GRAPH(SURFACE) x y z; (Windows/GiveWin2/Oxmetrics version only) gives a 3D surface plot.

Other improved commands:

• ANALYZ extensions: (1) allow expressions that vary over observations (such as elasticities for several time periods); (2) asymmetric confidence intervals, computed using Monte Carlo simulation. New options control how the resulting values are to be displayed when they vary across observations.
• BJEST(HCOV=U), ML PROC (HCOV=U) yield parameter standard errors from numeric second derivatives (new default). Usually much better than the BFGS or DFP rank 1 update methods. BJEST combined with ML PROC will estimate a linear or nonlinear regression with ARMA errors.
• BJIDENT(ESACF) - Extended Sample ACF, which is useful for identifying stationary or nonstationary ARMA models.
• BJEST(HCOV=U,EXACTML) - numeric second derivatives for reliable standard error computation. Can be used with ML PROC to estimate regression models with ARMA residuals by ML.
• CDF(F,DF1=DF2=) - now handles non-integer degrees of freedom, so it can compute the CDF for the beta distribution.
• CONVERT(MAP=) - a new way of creating series which are aggregated or averaged over an old series.
• FIML (HCOV=C,HITER=C) gives parameter standard errors from Discrete Hessian (a very close approximation to analytic Hessian, based on numerical differences of the analytic first derivatives). Can also be used for iteration, providing quadratic convergence speed and generally smaller SEs than the default HCOV=B.
• FORM(SUM) - makes an equation as a sum of terms. For example FORM(SUM) EQ S X1-X3; creates FRML EQ S = X1+X2+X3;
• GMM(MAP=) - sparse map matrices are handled much more efficiently, since zero rows and columns are no longer retained in the COVOC matrix.
• LIML stores the log likelihood for the estimates in @LOGL. The value is the same as that computed by FIML when equations for all the right hand side endogenous variables as a function of the instruments are included.
• LIML (BEKKER) gives Bekker standard errors, which are the new default and are more robust to problems with many/weak instruments. The estiimated concentration parameter and Cragg-Donald F-statistic are printed, for detecting models that may have this problem.
• LIST(SUFFIX=) - similar to PREFIX=, but adds the text to the end of each list item.
• RANDOM (MEAN=series, STDEV=series, VAR=series, REPLACE) - the series options allow different means and different variances for each observation. RANDOM(NOREPL) draws without replacement (like drawing cards, or could be used for permutations).
• READ now handles all current Excel files (2/3/4/5/7/98/2000/2002). This is especially helpful for datasets with more than 32768 rows, although there is still a limitation of 65535 rows. READ can also now handle Stata v7, v8, and v9 files.
• REGOPT CHOWHET - computes a Chow test which is robust to heteroskedasticity.
• SORT(RANK,AVETIES/MINTIES) - controls how ties are handled in the rank calculation (vs. assigning random ranks to ties). The SORT algorithm is improved to use QuikSort, which is much faster when there are a lot of observations.

Other improvements:

• The TSP User's Guide and TSP Reference Manual have been converted to the smaller 6" x 9" format. They are also available as PDF files (and HTML help systems), so you can read the texts on your PC when your printed manuals are in some other location.
• TSP now reads up to 500 characters per line in interactive mode (the old limit was 80). Input files may now have linefeed (LF) or carriage return characters (CR) at the end of the line, to aid compatibility across Windows, MAC, and unix platforms.
• Dashed lists handle leading zeros properly. E.g. x09-x11 expands to x09 x10 x11 instead of x9 x10 x11.
• The list @ALL is always available, and expands to all currently defined series. So MSD @ALL; may be convenient for looking at the data after several series have been read from a spreadsheet.
• Warning message for missing values have been streamlined and limited in number. So it should no longer be necessary to use commands like OPTIONS LIMWARN=0; to turn off all warnings. (this change was made early in TSP 4.5, so it has been in effect for quite awhile).

VERSION 4.5

· GiveWin compatibility (now superseded by Oxmetrics)

· We have converted our acclaimed TSP Reference Manual to Microsoft Help format, to facilitate easy lookup of command syntax while developing your TSP programs. But never fear! Unlike many software companies, we will still ship hard copies of our manuals as well.

New procedures

· POISSON estimates Poisson models that have a linear regression function (E(y)=Xb) using maximum likelihood.

· NEGBIN estimates Negative Binomial type 1 and type 2 models with linear mean function for the mean using maximum likelihood.

· ORDPROB estimates Ordered Probit models (models for categorical data generated from a latent variable with mean Xb that is normally distributed) using maximum likelihood.

· LMS (Least Median of Squares) estimates linear models where the objective function is the median squared residual. It is a robust estimator which is useful for detecting outliers.

Enhanced procedures

· AR1 combines grid and iterative methods, to ensure that a global maximum of the objective function has been achieved. Analytic second derivatives for faster iteration.

· ARCH analytic second derivatives, for fast Newton iterations and robust standard errors. Several initialization options for presample values of h(t) and e(t)**2.

· BJEST (ARIMA estimation) exact ML with stationarity and invertability imposed.

· BIDENT Burg algorithm for partial auto correlations.

· FORCST now can do forecasts after VAR, ARCH and TOBIT estimation.

· CNORM2(z1,z2,rho) new bivariate normal CDF function, for ML estimation of bivariate probit and trinomial probit models (differentiable).

· READ is able to read Stata datasets directly (v2-6 datasets with the .dta extension).

· Extended NOPRINT option can be used to suppress printing of INPUT files.

· SAMPSEL combines grid and iterative methods, to ensure that a global maximum of the objective function has been achieved. Analytic second derivatives for faster iteration.

If you have any questions or comments about TSP please send an email to info@tspintl.com.

Comments or questions about this website should also be sent to the info@tspintl.com. Lost? Please consult the site map.