NEWS.md
mmrm now returns score per subject in empirical covariance. It can be accessed by component(obj, name = "score_per_subject").empirical_g_mat in the mmrm object, instead of the previous empirical_df_mat matrix. The model fit is now much faster and does not exhaust the memory anymore. If old model fit objects are used, the empirical_df_mat will still be used correctly, however a deprecation warning will be issued. Please consider re-fitting the model to get the new empirical_g_mat matrix.mmrm from source using a TMB version below 1.9.15, and installing a newer TMB of version 1.9.15 or above, would render the mmrm package unusable. This is fixed now, by checking in the dynamic library of mmrm whether the version of TMB has been sufficient.TMB was switched on, a warning would be given by fit_mmrm(), instructing users to turn off the tape optimizer. However, this is not necessary for reproducible results. Instead, it is now checked whether the deterministic hash for the TMB tape optimizer is used, and a warning is issued otherwise.fit_mmrm() was not visible to the user when calling mmrm() because it was caught internally, causing the first fit in each session to fail for the first tried optimizer and falling back to the other optimizers. The warning is now issued directly by mmrm(). This change ensures that the first model fit is consistent regarding the chosen optimizer (and thus numeric results) with subsequent model fits, avoiding discrepancies observed in version 0.3.13.TMB package versions below 1.9.15, MMRM fit results are not completely reproducible. While this may not be relevant for most applications, because the numerical differences are very small, we now issue a warning to the user if this is the case. We advise users to upgrade their TMB package versions to 1.9.15 or higher to ensure reproducibility.mmrm ignored contrasts defined for covariates in the input data set. This is fixed now.predict always required the response to be valid, even for unconditional predictions. This is fixed now and unconditional prediction does not require the response to be valid or present any longer.model.frame has been updated to ensure that the na.action works correctly.emmeans::emmeans returned NA for spatial covariance structures. This is fixed now.car::Anova gave incorrect results if an interaction term is included and the covariate of interest was not the first categorical variable. This is fixed now.car::Anova failed if the model did not contain an intercept. This is fixed now.TMB is turned on. If so, a warning is issued to the user once per session.mmrm now checks on the positive definiteness of the covariance matrix theta_vcov. If it is not positive definite, non-convergence is messaged appropriately.model.matrix has been updated to ensure that the NA values are dropped. Additionally, an argument use_response is added to decide whether records with NA values in the response should be discarded.predict has been updated to allow duplicated subject IDs for unconditional prediction.Anova is implemented for mmrm models and available upon loading the car package. It supports type II and III hypothesis testing.start for mmrm_control() is updated to allow better choices of initial values.confint on mmrm models will give t-based confidence intervals now, instead of the normal approximation.mmrm_control(), the allowed vcov definition is corrected to “Empirical-Jackknife” (CR3), and “Empirical-Bias-Reduced” (CR2).df_md, it will return statistics with NA values.method of mmrm() now only specifies the method used for the degrees of freedom adjustment.vcov argument of mmrm().model.matrix() and terms() methods to assist in post-processing.predict() method to obtain conditional mean estimates and prediction intervals.simulate() method to simulate observations from the predictive distribution.residuals() method to obtain raw, Pearson or normalized residuals.tidy(), glance() and augment() methods to tidy the fit results into summary tables.tidymodels framework support via a parsnip interface.covariance to mmrm() to allow for easier programmatic access to specifying the model’s covariance structure and to expose covariance customization through the tidymodels interface.mmrm() follows the global option na.action and if it is set other than "na.omit" an assertion would fail. This is now fixed and hence NA values are always removed prior to model fitting, independent of the global na.action option.model.frame() call on an mmrm object with transformed terms, or new data, e.g. model.frame(mmrm(Y ~ log(X) + ar1(VISIT|ID), data = new_data), would fail. This is now fixed.mmrm() always required a data argument. Now fitting mmrm can also use environment variables instead of requiring data argument. (Note that fit_mmrm is not affected.)emmeans() failed when using transformed terms or not including the visit variable in the model formula. This is now fixed.mmrm() might provide non-finite values in the Jacobian calculations, leading to errors in the Satterthwaite degrees of freedom calculations. This will raise an error now and thus alert the user that the model fit was not successful.options(mmrm.max_visits = ) to specify the maximum number of visits allowed in non-interactive mode.free_cores() in favor of parallelly::availableCores(omit = 1).model.frame() method has been updated: The full argument is deprecated and the include argument can be used instead; by default all relevant variables are returned. Furthermore, it returns a data.frame the size of the number of observations utilized in the model for all combinations of the include argument when na.action= "na.omit".component(., "optimizer") instead of previously attr(., "optimizer").mmrm function call with argument method. Options are “Kenward-Roger”, “Kenward-Roger-Linear” and “Satterthwaite” (which is still the default). Subsequent methods calls will respect this initial choice, e.g. vcov(fit) will return the adjusted coefficients covariance matrix if a Kenward-Roger method has been used.mmrm arguments to allow users more fine-grained control, e.g. mmrm(..., start = start, optimizer = c("BFGS", "nlminb")) to set the starting values for the variance estimates and to choose the available optimizers. These arguments will be passed to the new function mmrm_control.drop_visit_levels to allow users to keep all levels in visits, even when they are not observed in the data. Dropping unobserved levels was done silently previously, and now a message will be given. See ?mmrm_control for more details.mmrm calls, the weights object in the environment where the formula is defined was replaced by the weights used internally. Now this behavior is removed and your variable weights e.g. in the global environment will no longer be replaced.emmeans package for computing estimated marginal means (also called least-square means) for the coefficients.summary, logLik, etc.