match after changes in R devel.getOption("ddhazard_max_threads") defaults to one.all.equal.PF_get_score_n_hess are fixed. One is
that a off diagonal block in the observed information matrix was not
computed. The other is that parts of the score and observed information
matrix was only correct if parts of them were multiplied by the
duplication matrix.nlopt is no longer used in mode optimization. A
Newton–Raphson method is used instead. This seems a bit faster in some
cases and does not fail in some cases where nlopt did.fix_seed argument is added to
PF_control. fix_seed = FALSE combined with
averaging and a low number of particles seems to yield better
results.PF_EM when some periods do not have any
observations.PF_EM.PF_get_score_n_hess is added to compute the approximate
negative observation matrix and score vector.nu in PF_control scales the scale matrix
to get an identical covariance matrix.get_Q_0.predict.ddhazard has been re-written. The output with
type = "term" has changed. It now yields a list of lists.
Each list contains a list for each new_data row. The time
zero index is no longer included if tstart and
tstop is not matched in new_data. Parallel
computing is no longer supported. It likely did not yield any reduction
in computation time with the previous implementation. Calls with
type = "term" now uses the tstart and
tstop argument and supports predictions in the future. A
covariance matrix is added to the terms in the predictions.type == "VAR" in particle filters in the
smoothing proposal distribution. This has a major impact for most
calls.type == "VAR" in particle filters where
the transition from the time zero state to time one was not used in the
M-step estimation. This only has a larger impact for short series.method == "bootstrap_filter" where a
wrong covariance matrix was used for the proposal distribution.method == "AUX_normal_approx_w_particles" where a wrong
covariance matrix was used for the proposal distribution.logLik.PF_clouds. The log-likelihood
approximation was too high especially for the auxiliary particle
filters.PF_EM...._w_particles methods so
results have changed.random and fixed argument is added to
PF_EM as an alternative way to specify the random and fixed
effect parts.PF_EM and can be
estimated with model = "exponential".ddhazard objects are no longer
degenerate (e.g., in the case where a second order random walk is used).
Instead the dimension is equal to the dimension of the error term.PF_EM has been moved from the
control list. Further, there is a PF_control
which should preferably be used to construct the object for the
control argument of PF.static_glm.PF_EM uses \(Q_0\)
instead of \(Q\) for the artificial
prior and a bug have been fixed for sampling in the initial state in the
backward filter. This have changed the output.PF_EM with seed argument. The new way to
get reproducible is to call
f1 <- PF_EM(...); .GlobalEnv$.Random.seed" <- f1$seed; f2 <- eval(f1$call)
kinda as in simulate.lm.model argument to ddhazard
should be changed from "exp_bin",
"exp_clip_time" or "exp_clip_time_w_jump" to
"exponential".glm is used to find the first state
vector.ddhazard control argument is
changed.The following has been changed or added:
get_risk_obj when
is_for_discrete_model = TRUE in the call. The issue was
that individuals who were right censored in the middle of an interval
were included despite that we do not know that they survive the entire
interval. This will potentially affect the output for logit fits with
ddhazard.ddhazard_boot now provides the option of different
learning rates to be used rather than one if the first fit
succeeds.ddhazard with
control = list(criteria = "delta_likeli", ...). The
relative change in coefficient seems “preferable” as a default since it
tends to not converge when the fit is has large “odd” deviation due to a
few observations. The likelihood method though stops earlier for model
does not have such deviation.ddhazard.hatvalues method for ddhazard.
These described “ddhazard” vignette and examples of usage are shown the
vignette “Diagnostics”.residuals method and a
vignette “Diagnostics” with examples of usage of the
residuals function.rug call in the shiny app demo, fixed a bug
with the simulation function for the logit model and added the
computation time of the estimation to the output.ddhazard with
control = list(use_pinv = FALSE, ...).exp_combined method.The following have been added:
weights argument when calling ddhazard.ddhazard_boot. See the new
vignette ‘Bootstrap_illustration’ for details.print is added.