PcGive 13: enhancements and solved problems

Fixed in PcGive 13.3

Fixed in Logit: fixed run-time error in OutputFreqTable
Fixed in Negbin estimation: was not printing probabilities; added check that truncation is not set too low.

New and improved in PcGive 13.2

System with one Y: now writing name of dependent variable.
Count data estimation with y-value in excess of 51 would give runtime error in OutputFreqTable.
Panel estimation batch code should write setdummy("none"); instead of setdummy(); when there are no dummies.

New and improved in PcGive 13.1

X12arima store and option batch commands didn't work.
When selecting a subsample in discrete choice modelling, automatic model selection does not work. The initial GUM is the subsample model, but the remainder the full sample, so that the tests go haywire.
Further Output: write nonlinear code or batch code to map to I(0): need to write all names as string ("var") in case any names are not identifiers.
Algebra error: NLS would not work with a dated database if the date variable was called Date: there would be two variables with the same name which NLS would not allow. Fixed by both allowing duplicate names in NLS and emoving the duplication.
Generated Ox Batch code for regime-switching models would not run with message switchpack.dll not found.
batch/volume3_garch.ox was not updated to use the new class name PcGiveGarch.
PcNaive using Autometrics and recursive estimation could select the wrong sample if the selected model has shorter lag length than the GUM. PcNaive could fail in Autometrics initialization on an uninitialized variable.
Forecasting with derived functions didn't work with dated database, because the dependent variable(s) where expected to be at the front. Changed default naming, using DF: F1: F2: etc prefixes. Improved output from store batch command

New and improved in PcGive 13.0

Selection of robust standard errors is now done in estimation menu (is different covariance estimator). Now it is remembered when it is used, and part of the generated Ox code. The entry under options now solely relates to the additional table which is also available under Further Output. The Options following maximization settings now purely relate to output options.
Ox code for GARCH models: class now called PcGiveGarch (instead of Garch, which was the same as G@rch).
Pointwise asymptotic Std.Error bands on residual QQ-plots
Regime switching models: Markov Switching.
Diagnostic testing
- Added RESET23 and Vector RESET/RESET23 tests.
- Added p-values for Portmanteau, Portmanteau not part of system test summary if open lag structure.
- Hetero-test now removes variables that are identical when squared (were already removed from output, now removed at an earlier stage). (useful when many dummies are present)
- Now doing Hetero and ARCH degrees of freedom as in PcGets (ARCH still undersized)
- Added Hetero, Index and RESET23 to PcNaive
- Added Index and Vector Index test
- Index test removes variables that are identical when cubed
- Hetero test: removing observations with zero residuals, removing dummy variables from regressors. For 4 or more equations: using rotated form (n equations instead of n(n+1)/2). Now including unrestricted/fixed variables in test. Single eqn AR-test/Hetero-test diagnostic for system: only using variables with non-zero coefficients in the RF. Single eqn diagnostic for system: ordered by equation.
- Test Summary and Autometrics: now reports RESET23 (using squares and cubes) instead of RESET (just squares) whenever the RESET test is included.
Improved date printing for large residuals
Added sigma to Autometrics single eqn output (Not AdjR^2, but note that highest AdjR^2 corresponds to smallest sigma).
Autometrics added to cross-section
Autometrics for binary logit/probit and for count data.
Autometrics: can impose long-run sign restrictions on search space.
PcNaive can now run with Autometrics and impulse saturation, but dummies are not reported in the output.
Small change to Autometrics output: stages more clearly identified; now including coefficients of terminal models as well as p-values.
Moved DW to Test menu, added Adj.R^2 to output, replaced var(Y) by se(Y) (using T-1, so that AdjR^2 = 1 / sqr(sigma/se(Y))).
Slight change in handling `lagged variables' (i.e. variables ending in underscore followed by number, e.g. XYZ_2):
- cross-section regression now shows all, and treats them as normal variables
- time-series and other modules that allow formulation of lags:
  - if XYZ is in the database, then all XYZ_# are removed from the list in the model dialog
  - if XYZ is NOT in the database, then all XYZ_# are kept, but not treated as lagged variables.

Fixed in PcGive 13

Run batch that estimates SEM (e.g. MulTut1.in7) then switch to VAR (e.g. MulTutVAR.fl) produces error messages, because the SEM is set for the new VAR.
Autometrics with backtesting against current GUM: run-time error when printing results.
Testing for omitted variables is wrong when there are variables that are lags in the database preceeding the tested variables (i.e. testing for X when the database has W_1 and X).
Reset test would fail if no regressors in the model
PcNaive mislabelled recursive test rejection frequencies when saving the results in the _r output file (graphs are correct). The labels where ordered by nominal frequency instead of by test.
Sample selection in models for discrete data: did not skip over missing values.
Count data models: could not select k=1 radiobox in negative binomial (but could specify k=1 directly).
1SLS: fixed out-of-memory error on very large system.
Robust F test of restrictions: was not divided by the degrees of freedom

This file last changed .