NEWS

---

glmSTARMA 1.1.0 (2026-07-19)

New Features

• Added the argument const to the vgamma() family function. It is used in the transformation of past observations when the log-link is used: past observations are now transformed as log(y + const) instead of log(y).
• Added the argument dispersion_constant to the dglmstarma.control() and glmstarma_sim.control() functions. Its value is passed to the const argument of vgamma() when the log-link is used for the dispersion model in dglmstarma().

Minor Improvements

• The log-likelihood of the QuasiBinomial family is now more robust against numerical issues.
• Removed redundant checks for NA and infinite values in the ts argument of the glmstarma() and dglmstarma() functions.
• Removed unused C++ methods that were left over from a preliminary version of the package.
• Added withr to the Suggests field in the DESCRIPTION file; it is now used in the test-data.R test.

Bug Fixes

• Fixed a bug that prevented a dispersion model with an identity link from being fitted.
• Fixed a bug in the variance() and dev.resids() functions of the vbinomial() and vquasibinomial() family functions.
• Fixed a bug in the dev.resids() function of the vbinomial() and vquasibinomial() family functions that caused problems when observed values were on the boundary of the support.
• Fixed a bug in the variance() function of the vnormal() family function, which now returns the correct dimensions when ignore_dispersion = TRUE and mu is not a matrix.
• The default dispersion parameter for vquasibinomial() is now correctly set to 1 instead of 0 (not allowed) if the user does not specify a value and it must be estimated.
• Fixed a bug in the vnegative.binomial() family that did not allow zero values in the dispersion argument. For zero-values in the dispersion, the Negative Binomial distribution reduces to a Poisson distribution.
• Fixed a bug in the variance_fun of the InversGauss family in C++.
• Changed the observation transformation for vpoisson("sqrt"), vquasipoisson("sqrt"), and vnegative.binomial("sqrt") to sqrt(x), as sqrt(x + 3/8) * 2 was causing problems.
• Fixed a bug in the link_trafo() and derivative_link_trafo() functions of the SoftClippingBinomial and SoftClippingQuasiBinomial family functions in C++, which caused problems when different values of n were used across locations.
• Fixed a bug in dglmstarma.control() where parameter_init_dispersion was not checked.
• Fixed a bug in predict.dglmstarma() that ignored the specified copula for prediction type "sample".
• Fixed bugs in glmstarma_predict() and dglmstarma_predict() that always caused warnings when time-varying covariates were included in the model.
• Fixed a bug in predict.glmstarma() that prevented predictions for models with a homogeneous intercept.
• Fixed a bug in fitted.dglmstarma() that prevented initial time points from being removed when drop_init = TRUE.
• Fixed a bug that prevented residuals of a model from being scaled by the variance.
• print.summary.glmstarma() and print.summary.dglmstarma() can now handle empty models.
• delete_glmSTARMA_data() now gives correct messages when datasets do not exist in the cache.
• The documentation of the SST dataset now lists the correct number of locations/grid points.
• Removed an incorrect reference to a vgarch family from the stfamily help page.