SVD_Lin(), in which SVD coefficients evolve
according to a Lin() prior.n_comp argument of SVD
priors. Previously n_comp was half the number of components
of the ssvd argument (rounded up), giving a value of
3 for the SSVD objects implemented in bage. It
is now simply 3, with no link to the ssvd
argument.rows argument to augment() and
forecast() allowing users to request results for a subset
of data. (#106)s_seas argument in
RW_Seas() and RW2_Seas() priors, which was
previous 0. In practice, the default of 0 was
not intuivitely obvious, but neither was any other value (eg
1), so it is safest to force the user to specify
explicitly.mod_pois(),
mod_binom(), and mod_norm() no longer insist
that the outcome variable is included in data. Specifying
models where the outcome model is not included in data is
useful when carrying out simulations or when drawing from the prior
distribution but not the posterior distribution. The outcome variable in
data is now checked when function fit() is
called. (#109)mod_pois(), the expression
exposure = NULL is now equivalent to
exposure = 1, and weights = NULL, or the
omission of the weights arugment, is equivalent to
weights = 1.DRW(), DRW2() implementing
damped versions of random walks and second-order random walks. In the
DRW() case, the basic idea is
x[t] = phi * x[t-1] + error[t]. In the DRW2()
case it is
x[t] = x[t-1] + phi * (x[t-1] - x[t-2]) + error[t]. These
priors are useful for forecasting.SVD_DRW() and SVD_DRW2(). As
above, but the random walks are on the coefficients for SVDs.prt_deaths where deaths and population
datasets had not merged properly.usa_deaths, where the deaths column had
class "ts" rather than "numeric".CSA, HIMD_R,
HIMD_P1, HIMD_P5, WMD_C,
WMD_E, for use in modeling school attendance, internal
migration, and marriage. Also added help page for all scaled SVDs.ssvd(), for creating scaled SVDs is now
exported. (Previously it was an internal function.)v that can be used to specify a version.generate() method for scaled SVDs now use the term
sex rather thansexgender in results, for
consistency with components().generate() and components() methods for
scaled SVDs now default to showing results disaggregated by sex/gender.
Previously the default was to show results for the total
population.set_covariates() now use
sparse rather than dense matrices, which has significantly speeded up
computations, and reduced memory usage, for models with categorical
covariates.fit() silently aggregated rows
with duplicated values for the predictors (ie the variables to the right
side of the ~ in the model formula. From 0.9.9 this
aggregation only occurs with Poisson models where dispersion has been
set to 0, or with Normal models. Even in these cases, aggregation no
longer happens if the model includes a data model. In cases where
aggregation would previously have occurred, fit() warns
users about the behavior change. Thank you to Salah Merad for pointing
out that the aggregation was leading to unexpected results."inner-outer" fitting method can now be used with
models that include covariates.dispersion() has been added, which
returns prior or posterior values of the disp parameters.
(Previously the only ways to obtain values of disp was to
use function components().)unfit() was not previously removing
information on the "inner-outer" fitting method, making it
appear that a model had been fit using this method, when it had in fact
been re-fit using the standard method. This has been corrected.set_datamod_exposure(),
set_datamod_miscount(), set_datamod_noise(),
set_datamod_over(), and set_datamod_under().
In introduction to data models is given in vignette 10.n_draw() for querying the
n_draw value of a model object.replicate_data(), changed the default for
condition_on to "fitted" in cases where
dispersion is zero, since the "expected" option is not
permitted in these cases.Lin() prior in Mathematical
Details vignette, and on online help, to describe case where
s = 0. Also updated description of prior in online help to
match current implementation.exposure = 1, and aggregation of weights in normal models
where weights = 1.mod_pois(), mod_binom(),
mod_norm() to clarify prior for dispersion. Added default
value (of 1) to set_disp(). This does not affect behaviour,
but is a bit clearer for users. (#94)set_covariates(), added paragraph in online
documentation to note that calling set_covariates() on a
model that already has covariates deletes the existing covariates. Also
added a warning message to set_covariates() when covariates
being overwritten. (#95).mod_pois() emits message (not warning) if one or more
rates are suspiciously high (which is often a sympton of inaccurate
exposures.) (#96)rpois_rvec(), which sets
x[i] equal to lambda[i] for lambda[i] > 1e8. This avoids numeric
problems which can lead to valgrind errors. The guarded version is
called by augment() and by replicate_data().
The user is warned when the threshold of 1e8 is exceeded.set_n_draw(), which was failing to thin
covariate coefficient draws. (#97)mod_norm() and in the
Mathematical Details vignette (vignette 2), give a new parameterisation
of the mod_norm() model, expressed on the original scale,
not the transformed scale (and closer to the paramterisation used by
mod_pois() and mod_binom().)bage_mod_norm method for
replicate_data() so that it returns results on the original
scale.bage_mod_norm methods for the helper
functions for augment() so that they properly incorporate
weights.original_scale argument to
components(), to be used with normal models. Also added
message remining users that, with normal models, components were on a
log scale (#88)NaN and no
Inf permitted.fit_default() refactored in
0.9.1. Bug meant that when optimizer switched from nlminb()
to optim() on non-convergence, optim() was not
starting from old parameter values.RW2_Infant() to priors table. Hat-tip
to Luke Morris for noticing that entry was missing. (#87)set_covariates(). Models can now include
covariates. Covariates are predictors other than the cross-classifying
dimensions such as age, sex, or time – though covariates can be formed
from these dimensions.gdp_pc_2023 and dens_2020
to dataset kor_births.prt_deaths dataset.set_seeds() function, allowing users to reset
random seeds (though this would be uncommon in normal use.)NAs in offset and predictor
variables."bage_mod" objects"multi" option for optimizer
argument to fit(). With "multi", the
fit() function first tries nlminb() and if
that fails switches to optim() with method
"BFGS"."bage_mod" objects to show
the time spent by TMB::sdreport rather than the time spent
by drawing from the multivariate normal (which, since
bage started using sparseMVN, is very
short).gdp_pc_2023 and dens_2020 variables
to kor_birthsreport_sim() excludes comparisons of
"hyper" parameters (eg standard deviations) if the
simulation model and estimation model use different priors with
different classes for that term. For instance if the simulation model
uses a RW() prior for age and the estimation model uses a
RW2() prior for age, then report_sim() will
not report on the standard deviation parameter for age.report_sim() stating
that the interface is still under development.zero_sum argument to con (short
for “constraint”). con = "none" corresponds to
zero_sum = FALSE, and con = "by" corresponds
to zero_sum = TRUE. Additional options will be added in
future.sd argument to RW(),
RW2(), SVD_RW() and SVD_RW2().
The initial value of the random walks are drawn from a
N(0, sd^2) prior. By default sd equals
1, but it can be set to
AR() and Lin_AR()
priors so that the coefficients no longer need to be consistent with
stationarity. The Stan user guide recommends against building in
stationarity:
https://mc-stan.org/docs/stan-users-guide/time-series.html#autoregressive.section
Also, testing for stationarity often causes numerical problems.AR() and
Lin_AR() priors.AR() and
Lin_AR() priors, so that partical autocorrelation function
(PACF), rather than the AR coefficients themselves, are restricted to
(-1, 1). Restricting the PACF to (-1,1) ensures stationarity.optimizer argument to fit(), giving
choice between three ways of optimizingquiet argument to
fit() so that when it is TRUE, trace output
from the optimizer is shown.start_oldpar argument to fit(), to
allow calculations to be restarted on a model that has already been
fitted."bage_mod" object.computations part of models so
that it works with models fitted using the “inner-outer” method.
Extended the print() method for "bage_mod" so
that it shows extra output for models fitted using the
"inner-outer" method."bage_mod" objects. (Thank you to Andrew Taylor for
suggesting this.)s = 0 in Lin() priorszero_sum argument to priors with an
along dimension. When zero_sum is
TRUE, values for each combination of a by
variable and the along variable are constrained to sum to
zero. This can allow better identification of higher-level terms in
complicated models. It can also slow computations, and has virtually no
effect on estimates of the lowest-level rates, probabilities, and
means.RW2_Infant() prior for modelling age-patterns of
mortality rates.s_seas parameter in RW_Seas() and
RW2_Seas() now defaults to 0, rather than 1, so that
seasonal effects are by default fixed over time rather than varying.
Using varying seasonal effects can greatly increase computation
times.computations(), which can be used to
extract this information from fitted model objects.quiet argument to fit(). When
quiet is TRUE (the default), warnings
generated by nlminb() are suppressed. (These warnings are
virtually always about NAs early in the optimization process and are
nothing to worry about.)HFD, a scaled SVD object holding data from the
Human Fertiltiy Databasedeaths –> isl_deathsexpenditure –> nld_expendituredivorces –> nzl_divorcesinjuries –> nzl_injuriesus_acc_deaths –> usa_deathskor_births, births in South
Koreareport_sim() now works on fitted models. Thank you to
Ollie Pike for pointing out that it previously did not.age variable in
divorces.rr3().
Call poputils function rr3() instead.newdata argument to forecast().Lin() and
Lin_AR() priors.method and vars_inner to
fit(). When method is "standard"
(the default) fit() uses the existing calculation methods.
When method is "inner-outer",
fit() uses a new, somewhat experimental calculation method
that involves fitting an inner model using a subset of variables, and
then an outer model using the remaining variables. With big datasets,
"inner-outer" can be faster, and use less memory, but give
very similar results.fit() now internally aggregates input data before
fitting, so that cells with the same combinations of predictor variables
are combined. This increases speed and reduces memory usage.print.bage_modssvd() no longer exported. Will export once
package bssvd matures.augment() so it runs fasterdivorces datasetset_datamod_outcome_rr3(), which deals with the case where
the outcome variable has been randomly rounded to base 3.augment() now creates a new version of the outcome
variable if (i) the outcome variable has NAs, or (ii) a
data model is being applied to the outcome variable. The name of the new
variable is created by added a . to the start of the name
of the outcome variable.standardization
argument: "terms", "anova", and
"none". With "terms", all effects, plus
assoicated SVD coefficients, and trend, cyclical, and seasonal terms,
are centered independently. With "anova", the type of
standardization descibed in Section 15.6 of Gelman et al (2014) Bayesian
Data Analysis, is applied to the effects.SVDS(), SVDS_AR(),
SVDS_AR1(), SVDS_RW(), and
SVDS_RW2() priors. Added indep argument to
corresponding SVD priors. SVD priors now
choose between ‘total’, ‘independent’ and ‘joint’ models based on (1)
the value of indep argument, (2) the value of
var_sexgender and the name of the term.HMD now contains 5 components, rather than
10.Lin() and
LinAR() priorsSVD_AR(), SVDS_AR(),
SVD_AR1(), SVDS_AR1(), SVD_RW(),
SVDS_RW(), SVD_RW2(),
SVDS_RW2()draws_linpred, added draws_effectfree,
draws_spline, and draws_svd. Modified/added
downstream functions.compose_time()report_sim()components().augment() method for bage_mod objects now
calculated value for .fitted in cases where the outcome or
exposure/size is NA, rather than setting the value of
.fitted to NA.components() is
called. augment() uses the linear predictor (which does not
need standardization.)disp are stored, rather than the
full standardized components.ssvd_comp().forecast.bage_mod() Forecasting. Interface not yet
finalised.generate.bage_ssvd() Generate random age-sex profiles
from SVD.draw_vals_effect_mod() was
malfunctioning on models that contained SVD priors.
Need a high-speed mirror for your open-source project?
Contact our mirror admin team at info@clientvps.com.
This archive is provided as a free public service to the community.
Proudly supported by infrastructure from VPSPulse , RxServers , BuyNumber , UnitVPS , OffshoreName and secure payment technology by ArionPay.