NEWS.md
In ModalCens: Parametric Modal Regression with Right Censoring

ModalCens 0.3.1

Bug fix for family = "invgauss": the optimizer could stop at a spurious solution on the boundary lambda = 3 * max(mode) of the feasible region when the fixed starting values were far from the data. A second, data-driven starting point (least squares coefficients and a moment-type lambda inside the feasible region) is now also tried, and the solution with the larger likelihood is returned. Fits that were already correct are unchanged.
Bug fix for family = "beta": the inverse logit is now evaluated in an overflow-safe way. Extreme trial values of the linear predictor during the line search could produce NaN and stop the fit with an error, which was observed mainly under heavy censoring.
family = "beta" also tries a second starting point (least squares coefficients) and keeps the solution with the larger likelihood. Under heavy censoring the default start could lead to a degenerate solution with phi close to zero.

Added the Log-t ("logt") distribution family, in which the logarithm of the response follows a Student-t distribution. The mode is reparameterized in closed form, and the family supports right-censoring and randomized quantile residuals. It reduces to the Log-Normal family as the degrees of freedom grow.
New argument nu in modal_cens() for the degrees of freedom of the Log-t family, treated as known (default 4). It is ignored by the other families.
summary() reports the degrees of freedom when family = "logt", and the fitted object stores them in the new element nu.
A warning is issued when the Log-t dispersion estimate lies at the boundary of the region that guarantees an interior mode.

Added the Log-Logistic ("loglogistic") and Birnbaum-Saunders ("bisa") distribution families, providing highly flexible modeling options outside the exponential family framework. Both support mode reparameterization, right-censoring, and randomized quantile residuals.
Added a caveat in the documentation regarding the strong parametric restriction of the Inverse Gaussian distribution.
Added the Log-Normal ("lognormal") distribution family, with mode reparameterization, right-censoring support, and randomized quantile residuals.