NEWS

systemfit 1.1-30 (2023-03-22)

adjusted bread.systemfit() and estfun.systemfit() to recent changes of the 'sandwich' package

some adjustments to make the 'systemfit' package compatible with future versions (i.e., >= 1.4.2) of the 'Matrix' package

corrected a typo in the vignette (in the equation for calculating the variance-covariance matrix of the coefficients estimated by the "3SLS-Schmidt" estimator in the presence of restrictions)
fixed a few minor issues pointed out by 'R CMD check'

implemented 2SLS and 3SLS (instrumental variable) estimations for panel data
fixed a bug that occurred in panel data estimations when one variable name was a substring of another variable name
internal improvements to make this package compatible with R 4.0.0

users can now use pdata.frame() (in addition to plm.data()) to specify the panel structure of (panel) data sets (patch provided by Kevin Tappe -- thanks a lot!) [note: the use of plm.data() is deprecated]

added argument 'residCovDiag' to logLik.systemfit()
improved documentation of logLik.systemfit():
slightly improved documentation of systemfit():
imported functions from packages 'methods' and 'stats' (as indicated by 'R CMD check --as-cran' when using R-devel)

added (incomplete) estfun() method (from the "sandwhich" package) for objects of class "systemfit" (currently, the residual covariance matrix is always assumed to be the identity matrix, although this is only true for OLS estimations!); therefore, the sandwich package is now "imported"
changed dependency ("Imports") from "stats (>= 2.15.0)" to "stats (>= 2.14.0)"

slightly revised the package vignette so that it works with version 2.0 of the "sem" package

the systemfit package now works with version 2.0 of the "car" package (linear.hypothesis() has been renamed as linearHypothesis())

the logLik method now uses the average number of non-NA observations per equation (instead of the number of rows of the residual matrix, which might include NAs) to calculate the log-likelihood value: hence, the likelihood value is correct now even if there are NAs in the data; the calculation of the log-likelihood value for unbalanced systems uses the conditional density function for the unbalanced observations assuming that the errors of not-included observations are zero (the log-likelihood value was incorrect before if there were NAs in the data)
the logLik() method now takes the number of (linearly independent) coefficients from element "rank" to calculate the degrees of freedom, because this is correct even if there are NAs in the data (it was incorrect before if there were NAs in the data)
the logLik() method now returns (as attribute "nobs") the correct number of observations even if there are NAs in the data (it was incorrect before if there were NAs in the data)

model.matrix.systemfit.equation() now returns a matrix (and not a vector) even if the equation has only one regressor (including the constant); this fixes bugs in model.matrix.systemfit() and predict.systemfit() that occured if (at least) one equation has only one regressor

the subfolder "tests" and the test scripts in this subfolder are no longer included in the R package in order to reduce the time for checking this package on CRAN

added a NAMESPACE file
the returned fitted values, residuals, and the model frame include *all* observations now, where fitted values and residuals of observation that were not included in the estimation are NA
systemfit can estimate systems of equations with unequal numbers of observations now