PcGive 10.4 and GiveWin 2.3: enhancements and solved problems

GiveWin 2.3

• Uses log scale for axis with large negative numbers would send GiveWin (or Ox) into infite loop, using more and more memory.
• Entering a zero for the `scale' in regression/scale in graphics, and on `apply', hangs GiveWin
• Selecting a graphics area, and then shutting down leads to a crash.
• Support for tsp .chm help file.
• reading .dat file with embedded names/samples goes wrong if the sample periods are not all identical.

PcGive 10.4

• Help: first looks in bin\..\PcGive\doc\_xyz.html, where xyz is the package name. If that isn't found, it will look for bin\xyz.html.
• Simultaneous equations model with identities. If lagged identities are used in the model, and these are entered in the system *before* any lagged endogenous variables, then dynamic forecasts and dynamic (long-run) analysis would go wrong.
• GARCH forecasts: plot of actual variable in forecast period (when withholding data for forecasting) was wrong (it was minus the forecast error). There was no problem with the forecasts themselves.
• Omitted variables test in the presence of Seasonals would get the timing of the selected variables wrong, and therefore produce meaningless results.
• Hetero test (full output) for IVE: labelling of coefficients is wrong
• Panel data: Further Output/Save Covariance matrix: doesn't work for xls file.
• Further Output/Covariance matrix: if HCSE/HACSE in Options, also print this matrix.
• Initialization error when more than 1 regressor in variance of GARCH.
• Cointegration (general restrictions): automatic removal of scaling was not always applied correctly, resulting in restrictions not exhibited in beta. This problem would be directly observable in that the restricted beta is not as specified when using non-unit values for standardization. This problem did not apply when standardizing on unity.
• Store forecasts in database with derived variables didn't work
• Batch code to map CVAR to model...: missing some commas

PcGive 10.3: enhancements and solved problems

• NLS/ML: can now plot marginal likelihood grids (keeping other parameters fixed).
• RLS: coefficients corresponding to singularity are now set to .NaN instead of 0 (this improves the graphs).
• Chow p-value for recursive estimation now remembered between runs.
• Forecasting and not computing any standard errors would lead to runtime error.
• Portmanteau statistic in Descriptive statistics did not subtract sample mean (but printed ACF is in deviation from sample mean). Now always reporting Portmanteau statistic in deviation from sample mean in all packages (this has a small impact on the reported statistic for GARCH and ARFIMA).
• Simultaneous equations model estimated by FIML or CFIML: some information was not properly reset after the Vector AR test (EGE-AR). This affected some output, which was wrong subsequently: test for restrictions; listed coefficients under Further Output/HCSE; information criteria (for CFIML).
• Further Output/Large residuals led to run-time error.
• GIVE model after FIML: identities from SEM would be appended to names of Ys, so that variable labels in output could be wrong.
• Simultaneous equations model with identities and a longer lag length on the identity endogenous variables than the normal endogenous variables: Dynamic simulation was wrongly ignoring the extra identity lags
• Orthogonalized impulse responses for model with identities would give runtime error.
• Impulse responses of simultaneous equations model ignored the instantaneous impact of endogenous variables that are on the right-hand side.
• RRF coefficient standard errors did not print SEs of identity part; Now also printing correlation of RRF residuals (as in version 9)
• Eigenvalues of long-run matrix: now sorted in decreasing modulus; always include imag in printing (even if all zero) Same for companion matrix.
• Long-run covariance: was printing negative covariances as zero.
• Added check for separation to discrete choice modelling.
• Panel data batch couldn't handle multiple group variables (for example, a variable that is both X and I: regressor and instrument)
• Panel data model without regressors was not allowed.

GiveWin 2.2 and PcGive 10.2: enhancements and solved problems

GiveWin 2.2

• Context sensitive help didn't work for for Mozilla or Netscape 7
• It appears that exit; only works from the batch editor, but not when running an fl file directly (nor when using Ctrl+B).
• Airline data on log scale: no top large label Improved labelling if the data range small
• Change pixel coords of area, then half scale (which adjust pixel coords), then full scale will leave area at bottom.
• Create Ox file, Save and cancel: is marked as clean, but because it doesn't exist as disk file, it cannot be run: add check, because this can be very confusing.
• $\theta_{12}L^{12}$ in graphics moves superscript of L left
• Printing from command line didn't work.

PcGive 10.2

• Volume II Section 11.9.1.1 Portmanteau refers to running mean version, which is not used since version 9.
• Linear restrictions or general restrictions without specifying any restrictions leads to crash.
• Analysis of lag structure in multiple eqn models: Seasonals were treated as lags. Also: outcome could be wrong because using incorrect variable ordering.
• Cointegration: restrictions on alpha of type A*theta was not allowed without beta restrictions.
• Batch commands for ATGARCH didn't work.
• Principal Components with column of zeros gave run-time error
• Added Save section to Further Output in Panel data, to allow saving of covariance matrix or parameters~stderr to a file.
• NLS involving Ya only, but Yb/Yc is missing (first half Yb, 2nd half Yc): gives no observations in this selection message, then failed to load ... message. Should only check actual and fitted (and loglik) for this.
  Also, if an irrelevant variable is deleted from the database, NLS would still complain about its absence
• store("forecasts"); undocumented; no forecast command in batch.
• Hetero/Hetero-X test was sometimes unecessarily unavailable with message:
  HeteroTest(): problems with det|V'V|
  Renamed this to
  Hetero test: non-positive auxiliary residual variance
  Also renamed singular W'W message to
  regressors are linearly dependent (numerically or exact - some coefficients set to zero)
• Graph simulation residuals over sub sample: Runtime error: 'matrix[1][200] - matrix[1][100]' bad operand
• Added Further Output to Panel data package.
• Extended principal component algebra code. Analysis with missing values used the wrong scale.
• Latex output: aa_bb must be changed to aa\_bb
• Forecasting with derived functions: now allowing H-step; found some examples where first std.error was zero (because instantaneous derivative was wrong).
• Restricted cointegration with lagged variable and trend restricted to the cointegration space: general restrictions dialog could show restricted variables in the wrong order.
• For monthly data, the first seasonal variable did not correspond to January.
• SEM: static long-run after test summary gives a run-time error.
• estsystem("NLS", 1980, 1, 1990,2) ignores the estimation sample and uses all available observations.
• PCFIML does now include HCSE in the output of FIML, however, the headers were missing, and they are not also used for testing restrictions. Problem with Options in selecting HCSE but not HACSE.
• Store in database of Garch: doesn't have fitted values.
• general restrictions for DPD/GARCH/Arfima/Logit.
• Unit Root test: If Trend (or Seasonals) selected, but Constant not selected, the Constant would be in the model, but not used in the critical values.
• PcGive Volume III reports the finite sample corrected standard errors as default for two-step GMM (when robust standard errors are selected). Message in output changed from using robust SE to using finite sample corrected SE.
• if I load PcGive 10.1 directly through its shortcut but exit via GiveWin (with the PcGive module open), I get a Runtime error "abnormal program termination" for PcGive10.exe.
• "help" does not work when using the descriptive statistics package. This was caused by some missing descriptive.html files.
• LaTeX output of SEM with one variable is wrong
• Write results for Impulse responses still wrong with Identities. Labels for custom impulses improved; now including identity eqns (with option not to).
• When running a GARCH(1,1) with OPG and robust std errors selected: robust se equals standard se: message was incomplete: `Robust-SE based on and analytical OPG matrix'
• Write results instead of graphing for Impulse responses leads to run-time error

GiveWin 2.1 and PcGive 10.1: enhancements and solved problems

GiveWin 2.1

• Text windows
  • "Find Next" (F3) doesn't work.
  • Text font can now be set to bold or italic.
• PostScript setup was not sticky between runs
• Automatic dialogs: now adjust to font size.
• Workspace
  • Maximized windows in GiveWin: keep maximized if switching through work space. Always process click on module, even if already selected.
  • Clicking on PcGive in Modules list does not bring it up when it is minimized.
  • Right-click on OxGauss file in workspace: Says Run ... where ... is rubbish.
• GiveWin's "Module Help Index" and "Search Module Help Index" previously only worked for Ox and TSP files. Now it will start the active module help if possible.
• Graphics
  • Minor improvement to axis in logs when not enough space for all labels.
  • Change graph scale with multiple areas goes wrong
  • Created separate All page for graphics properties The `Line Attributes' page of the Graphics Property dialog has been renamed `Lines'.
  • Reloading graph: would add box if that is default.
  • Exponentially-weighted moving average and correlation EWMA is an option in scatter plots.
  • Cubic spline control: has spline selected, but doesn't show bandwidth options.
• Algebra/Batch
  • Algebra/Batch save: didn't have warning when failed to save because file was read only.
  • delete Algebra function didn't work.
  • New Algebra functions: ewma and ewmc0.
• File input/output
  • Readinf version 2.1 .xls files: integers beween 32000 and 64000 are read as negative numbers (not a problem in version 3 onwards types; also note that GiveWinnever saves a .xls file with integers).
  • Saving as in7 file: didn't check for spaces (could be problem, e.g., when loading from .xls, and then saving).
  • Saving new file: offering default extension.
  • Added support for .csv files.

PcGive 10.1

• Logitjd: switched to binary when using weights.
• Descriptive Statistics
  • Moved to the Package menu.
  • Added principal components analysis.
  • Added ACF.
• Recall accross packages was problematic.
• HCSE
  • If HCSE selected in options: now including them in the test for restrictions.
  • HCSE for FIML: request in options was ignored.
  • HCSE for OLS: request in options now also in standard equation output.
• Further Output: option to print residuals exceeding a certain level.
• Treating infinity as missing in cross-section regression.
• Cointegrated VAR
  • general restrictions: now checking that restriction does not access higher restriction (lower is allowed). Fixed a potential problem with code in which a restriction that accesses a lower restriction.
  • Reduced form beta now includes -I block for clarity; printing standard errors if no additional restrictions imposed.
  • with restrictions, but zero degrees of freedom: would standardize beta.
  • Beta entirely restricted gives run-time error in recursive estimation.
  • Error when restricting last cointegrating vector entirely.
  • Variable names are back above the parameters in the cointegration restrictions.
  • Store recursive eigenvalues gave error message.
  • Runtime error: 'fCointData' wrong number of arguments in cointegration graphics with lagged DY and U removed.
  • Recursive restricted cointegrated VAR: including SE of beta when identified
• Progress: now ordering correctly, and showing intermediate reductions.
• RALS
  • When doing RALS, and when static long-run is checked under further output, you get error messages.
  • RALS sample automatically adjusted if necessary.
• Options: added Correlation matrix/Cointegration test/ Cointegration test: with Max test
• Added critical values to ADF summary table header; sample selection is now as for model (if necessary, adjustment for lags will be made automatically).
• Dynamic analysis
  • lag analysis on reverse order; printing intermediate lag length F-tests for single eqn.
  • Added lag structure analysis for VAR/multivariate models
• Heteroscedasticity test
  • now using internal scaling for enhanced accuracy of the auxiliary regression. This can give different results with data that is very unevenly scaling.
  • Reintroduced output of Heteroscedasticity coefficients from Test/Test
• Batch
  • added "testres { ... }" command.
  • for SEM: didn't allow non-token Y variables (such as "log(Y)").
  • Batch commands to do cointegration tests after estimating a VAR: test("I1"); test("I2");
  • This did not work: rank(0); estimate("COINT", 1964, 1, 1998, 1);
• Forecasting with derived variables could lead to a crash.
• CFIML: fixed problem with standard errors in the presence of unrestricted variables.
• 1SLS
  • skip order/rank condition check
  • with identities: SEs in wrong location
• Testing Linear&General restrictions for FIML
  • would use coefficient values of zero.
  • names in help text now include eqns.
• Linear restrictions: now remembers previous set.
• FIML estimation with unrestricted variables or more variables in URF than RRF would give a harmless warning: MaxBFGS(): initial Hessian dimension error, using I
• Arfima
  • show bands on ACF.
  • Added summary statistics for Arfima/GARCH forecasts.
  • naive forecasts as default.
• Store in database
  • dates are out by one year when seasonals are used. Store in database of recursive equation standard errors could also be wrong.
  • Renamed `Store in Database' to `Store Residuals etc. in Database'
• GARCH
  • changed 1e-6 default starting value for high ARCH lags to 1e-3, which avoids that they can get stuck at 1e-6.
  • dynamic forecasts with lagged Y: error bars wrong (were constant).
  • Random search in GARCH had run-time error.

GiveWin 2.01 and PcGive 10.0b: enhancements and solved problems

GiveWin 2.01

• Ctrl+M didn't work because it was doubly defined, now using Ctrl+I for Run as Command.
• 3D graphs from table did not handle entries with missing z coordinates correctly. Now turning patch into triangle if only one corner missing. Also added rotate commands from keyboard (Alt+ <>-+)
• On some versions of Windows 98 GiveWin 2 experienced intermittent crashes.
• Support for v96 dat files in GiveWin.
• Small fix for reading certain excel files.

PcGive 10.0b

• decision whether Hetero-X could be computed (based on degrees of freedom) was too stringent
• SEM equation with one variable would not write coefficient
• SEM dialogs now have identities not selected
• Print static forecasts: labelling of one-step forecasts is out by one (general header is correct).
• IVE static forecasts had argument error.
• RALS max grid (which corresponds to the graphs in section 6.3 of Volume I) did not appear in the graph.
• DPD didn't allow lags; DPD didn't handle Gmm/GmmLevel in batch.
• non-linear model initialization when not starting counting parameters at zero: index access out of range.
• non-linear model: graphic analysis has undefined variable same for store in database (neither is problem if normal model estimated first).
• Panel data help would not be launched from F1.
• Restricted cointegration could generate all zeros as starting values if only alpha's set to zero; now can restrict all alpha's or all beta's.
• If the static long-run has one variable, the label becomes the first character of that string (as an integer) rather than the string (single-equation modelling).
• Estimate SEM, then single equation: exiting PcGive could lead to crash.
• SEM: slight miscount in degrees of freedom of heteroscedasticity (and hetero-X) test
• CFIML: vector AR-test was not reported.

PcGive 10 documentation, Volume I

• Log-likelihood now includes the likelihood constant. Results in Volume 1 do not yet include this, so report a different log-likelihood. This also affects the Progress report, see e.g. p.87/89.
• Rals grid dialog has changed (p.72). To replicate the graphs, select Max grid.
• Encompassing test (p.90) printed output has M1 and M2 results switched in numerical output.

PcGive 10 documentation, Volume II

• On page 26: Hetero-X now printed.
• page 41: Select Model/Dynamic analysis should be Test/Dynamic analysis.
• Appendix A2, unlisted changes from PcFiml 9.3:
  • SEM equation-by-equation ARCH test, 2nd degrees of freedom now comparable to single equation estimation: T-kbar-2s, where kbar is average number of parameters per equation (rounded down), and s lag length.
  • SEM equation-by-equation Hetero and Hetero-X test, 2nd degrees of freedom now uses kbar (instead of URF k).
  • Multiple equation modelling: equation-by-equation portmanteau is different from PcFiml 9.3 because it does not use the running mean version of the ACF. Same for residual correlogram.