The model argument in bfactor() can now
be specified using the mirt.model() syntax to include more
cognitively friendly tracking of item names and respective locations
(requested by Afshin Khosravi)
Add reverse.score() function for reverse scoring
specific items within a matrix or
data.frame
Fixed issue related to missing data patterns that resulted in bias when estimating the hyper-parameters in single and multi-group models (reported by Paul Jewsbury)
mirt.model() syntax gains a negation operator for
omitting specific observed/latent groups from specifications. For
example, the following will omit “Group3” identifies from between groups
equality constraint definitions
CONSTRAINB[-Group3] = ...
RMSD_DIF() now supports datasets that follow
vertical scaling structures (i.e., when groups answer some items but not
others). Requested by Alexandre Jaloto
M2() functions now compute null model and SRMR fall
all models whenever possible, including the latent class variance
(reported by Hynek Cigler)
VCOV memory leak bugfix for mixture models (see Github issue #247)
Standardized residuals for point estimates now returned in
personfit() when passing return.resids=TRUE
(requested by Raymond Hernandez)
Fix for DIF() when sparse data included with mixed
item formats (reported by Heather Leigh Kayton)
When computing category-level information curves include the negative Hessian in computations (reported by Milica Kabic)
Allow missing data patterns in personfit(), as well
as a new option to return all raw item by person residuals (requested by
George Karabatsos)
Fix Zero-inflated model example in multipleGroup(),
which required the discontinuous trait location to be populated
explicitly with a customTheta syntax (reported by Brooke
Magnus)
Empirical reliability estimates in fscores() and
empirical_rxx() include option to use the true score
variance as an estimate of the observed score variance (suggested by
Hynek Cigler)
technical list gains a nconstrain
argument for specifying equality constraints with negative relationships
(e.g., a12 = -a21). Requested by Berend Terluin
Added unipolar log‑logistic model (Lucke, 2015) itemtype,
specified as itemtype = 'ULL'. Note that this automatically
changes a number of internal defaults, such as using a log-normal(0,1)
density for the latent traits, and where the theta_lim is
specified to be positive
Added complementary log‑log model (Shim, Bonifay, and Wiedermann,
2022) itemtype, specified as itemtype = 'CLL'
Added itemtype = '5PL' model for unidimensional
dichotomous data to included asymmetric response functions. Example in
help(mirt) also demonstrates asymmetric 2PL model as the
5PL itself is very unstable and requires strong priors
Methods using Quasi-Monte Carlo integration post-convergence were
not respecting correlated latent variable structures (reported by George
Kephart when using M2())
Bugfix for fscores() when supplying mixture models
that was introduced by changing previous classification default for
latent class models (reported by Karel Veldkamp)
residuals() gains a p.adjust argument
for FWE control
DRF() gains a DIF.cat argument to
compute statistics on a per-category basis when studying polytomous
items
expected.test() gains a probs.only
logical to return probability functions for each category (only used
when individual = TRUE)
Small bug fixes in C++ code that resulted in memory leaks
For models fit using mdirt() the
fscores() EAP and EAPsum methods now always returns
classification probabilities as the default (reported by Matthew
Madison)
SIBTEST() gains a DIF logical to
perform DIF tests across suspect_set
DIF() and SIBTEST() gain a
pairwise logical input to perform pairwise post-hoc
comparisons for multi-group applications
DRF() gains groups2test argument and
friends for multi-group models
infit and outfit statistics can now be computed in
itemfit() when missing data are present (requested by Hanif
on the mirt-package forum:
https://groups.google.com/g/mirt-package/c/_mA3YbMmbzM/m/CydOl-F4BQAJ?utm_medium=email&utm_source=footer)
coef(..., IRTpars=TRUE) is now applied to
multidimensional IRT models, provided that the item contains simple
structure (suggested by Sverre Ofstad)
Fixed match() bug in SIBTEST() when
total score is missing (reported by Ziying Li)
fscores(..., method ='EAPsum') now supports
returning the ACOV matrices, matching the behaviour of the other
estimators
Store previously defined customItems and
customGroup lists for use in secondary functions (e.g.,
DIF(), boot.mirt(), etc). Reported by Nataly
Beribisky
Combining priors with equality constraints no longer uses multiple prior definitions in the likelihood computations. Hence, constrained parameters are now treated as though they are a single parameter with only one prior distribution (reported by Matthias von Davier in the context of multiple-group models with between group item priors)
Added a groups2test argument to DIF()
to isolate individual grouping variable specification when using more
than 2 groups
Implicit argument ‘invariance’ stored in multiple-group objects
now automatically used in boot.mirt() (previously had to be
manually passed)
Bugfix when using items2test in DIF when input is a
character vector (reported by @jbuncher)
Bug fixes for multiple-group DIF testing with DIF()
when using more than two groups (reported by Ruben Neda and Davin Díaz
García)
boot.mirt() gains a boot.fun argument
to accept user-defined functions for extracting the associated
statistics to bootstrap
When verbose = TRUE in residuals() a
set of summary statistics is reported for easier flagging
itemfit() arguments changed to accommodate
outputting tables more consistently. Now a single
return.tables argument is used to specify which tables to
return
anova() removes support for the verbose
flag, and instead labels the rows of the resulting output to identify
the models
X2 and G2 classes of item-fit
statistics now better deal with large missing value vectors on a
per-item basis for better consistency
technical list gains a storeEMhistory
flag to store the EM history (requested by @netique)
DRF() gains best-fitting prior support (currently
limited to Gaussian distributions)
Correct index subset caused by tmp row removals in MG objects (fixes #227)
Progress bar added automatically (controlled via the
verbose argument) when using several of the package’s
secondary functions (e.g., fscores(), DIF(),
'DRF(), mdirt(), etc)
Added itemstats() function to give basic item
information statistics
Item-EFA models now automatically flips negative signs in rotate
solutions (e.g., via summary()) according to the sign of
the largest observed loading (allows easier interpretation of the
resulting correlation matrix)
response.pattern deals with completely missing
vectors now (issue #220)
residuals() gains a approx.z logical to
transform LD values into approximate z-ratios
mirt(), mixedmirt(), and
multipleGroup() now have model = 1 to fit a
unidimensional IRT model by default
Added covdata argument to fscores() to
allow latent regression covariate information as well. Example added to
fscores() documentation to demonstrate this
addition
Added RCI() function to compute reliable change
index via IRT modelling
Added delta method SE in coef(., IRTpars = TRUE) for
the nominal and nested-logit models
itemfit() gains a S_X2.plot argument to
visualize the expected-observed probability differences based on the
S-X2 conditional sum-score strategy
Added type = 'EAPsum' to plot() generic
to view an expected vs observed sum-scores plot
itemfit() gains a p.adjust argument to
allow for p-value adjustments in the output for all methods
anova() generic now supports a ...
input to compare many nested models, compared in sequence
Added type = 'threshold' to itemplot()
to plot cumulative probability information (requested by Azman
Sami)
Fixed Bug Error in if (any(SEtmp < 0)) that
appeared due new R 4.0+ behaviour (reported by Ziying Li and Caroline
Böhm)
Fix bug in itemfit() when plotting multiple-group
objects
Bugfix in fscores() report on which row failed to
converge when datsets contain response patterns that were completely
missing
Previous technical = list(removeEmptyRows = TRUE)
input now deprecated. Response patterns that are now completely missing
are supplied NA placeholders within estimation and post-estimation
supporting functions (e.g., fscores(),
personfit(), fixed(), etc)
Added converged element in DIF() output
to evaluate whether the nested model iteration converged
Added support for plausible-value draws in fscores()
when using response.pattern argument
Fix SE.type = 'Fisher' computation in multi-group
models (reported by Felix Zimmer)
Switch par and f inputs in
numerical_deriv()
Added gen.difficulty() to compute the generalized
difficulty statistics described by Ali, Chang, and Anderson (2015) for
polytomous response models (suggested by Alexander Freund)
Added RMSD_DIF() to compute marginal effect size
measure recently used in PISA anlayses when investigating
‘badness-of-fit’ DIF effects when using constrained multiple-group
models
extract.group() now explicitly requires the group
name to be passed rather than the group number (this is a far more
natural route)
plot(..., type =) now supports 'trace',
'infotrace', 'itemscore', and ''
for two-dimensional models to create faceted graphics
Added read.mirt() function back to package now that
plink is again available on CRAN
Syntax input from the car package’s
lht() function adopted within mirt’s
wald() function for easier specifications (see
examples)
Better cope with syntax definitions of models in
DIF(), particularly with the CONSTRAINB form
(reported by Hao Wu)
Corrected outer-product summation for
SE.type = 'Fisher' computation (reported by Felix
Zimmer)
Added fixedCalib() function to perform the five
fixed-calibration methods describe by Kim (2006)
Empirical histogram dentype convergence tolerance no
longer modified (default now the same as the Gaussian
dentype criteria)
Fix for GGUMs using model syntax input (was ignoring the slope loading specifications; reported by Ben Listyg)
fixed traditional2mirt() math for gpcm when 5 or
more category items are supplied (reported by Aiden Loe)
OpenMP support added to E-step portion of the package, where
number of threads can be specified via the mirtCluster()
function argument omp_threads. Special thanks to Matthias
von Davier for providing the omp reduction code in the
Estep.cpp file
Behaviour of mirt(..., large) has now been modified,
where large = TRUE now skips computing the unique response
patterns for datasets that likely contain little to no repeated response
patterns (suggested by Matthias von Davier). The previous two-step
behaviour is now achieved by passing large = 'return',
storing this list object, and passing it back to the large
input argument
Positive/negative sign remove from chi-square components in
residuals(type = 'LD') (requested by Cengiz Zopluoglu to
help avoid confusion). Sign is still however present in the standardized
correlation estimates
itemtype = 'rsm' reported the incorrect information
functions due to use of - instead of + from
traditional2mirt() (reported by Nasser Hasan)
column names of the fscores() results now correspond
to the model syntax definition names instead of the previous F#
convention
fix method = 'classify' option in
fscores() when more than two mixtures are fitted (reported
by Lisa Limeri)
fix bug in 'drop_sequential' scheme in
DIF() introduced in the previous version of mirt due to
some internal organization changes (reported by Balal Izanloo)
allow infit/outfit statistics to be computed for non-Rasch models (suggested by Alexander Freund for use with GGUMs)
added p.adjust argument to DRF()
(requested by Keri J. S. Brady)
support for computation of the ACOV matrix when the variance of
the specific factors are freely estimated in
bfactor()
fix for invariance = 'free_var' argument in
multipleGroup() for multidimensional models with correlated
traits, which previously fixed the correlation parameters inadvertently
(reported by Ruoyi Zhu)
use proper mins internal when using
extract.group() to keep the original minimum response
scoring pattern (reported by Adam Ťápal)
bugfix for single-group models for draw_parameters()
(reported by Keri Brady and @ddueber)
numeric model specification in bfactor() bug patched
when intervals were not 1 unit apart due to NA placeholders (reported by
Luis Manuel Lozano)
latent trait/class names now are forced to be different than the data column names (bug reported by Nathan Carter)
fixed X2*_df and PV_Q1* when missing
data pattern resulted in dropped categories (reported by Mac
Pank)
added likert2int() to convert Likert-type
character/factor responses to integer data
estfun() gains a centering argument to
center the scores (contributed by Rudolf Debelak)
impute argument in itemfit() and
M2() have been deprecated in favour of removing data
row-wise via na.rm=TRUE
Acceptance ratio when using MH samplers now returned prior to ‘Stage 2’ during estimation so that these ratios are better behaved. As well, an heuristic improved method for increasing/decreasing the acceptance ratios is now implemented
Added return_seq_model to DIF() to
return the final MG model on the last iteration of the sequential search
schemes
Bugfix in DIF() when sequential scheme was selected
but no items contained DIF on the first iteration (reported by Scott
Withrow)
SIBTEST() gains a plot argument to
create various plots depicting the (weighted) differences between the
focal subtest versus the matched subtest information
residuals() gains a 'JSI' type to
compute the JSI statistics proposed by Edwards et al. (2018)
residuals() gains an 'expfull' type to
compute an expected value table for all possible response patterns (not
just those observed in the data)
Fix for key variable for nested-logit models when
data are collapsed to have equal intervals (reported by Emil
Kirkegaard)
Added delta method for IRT parameter transformations when using multiple-group models (reported by Alex Miller)
empirical.poly.collapse argument added to
itemfit() to plot expected score functions for polytomous
items (suggested by Keri Brady)
SRMSR now reported in M2() for GGUMs (suggested by
Bo on the mirt-package forum)
weights argument added to
estfun.AllModelClass to allow for the inclusion of
survey.weights to calculate the scores
DIF() now simplifies the output by default rather
than returning lists from anova(). Wald tests are always
simplified
Where applicable, RMSEA statistics are reported in
itemfit() for tests that return suitable X2 and df
components
Fix negative TLI and CFI values when using the C2 statistic from
the M2() function (reported by Jake Kraska and Charlie
Iaconangelo)
Fix delta method SEs for 'gpcm' itemtype (reported
by Lennart Schneider)
When lower/upper bounded parameters are included the default optimizer is now ‘nlminb’ rather than ‘L-BFGS-B’. This is mainly due to the instability in the ‘L-BFGS-B’ algorithm which is prone to converging instantly for unknown reasons
mdirt() gains a item.Q list to specify
Q-matrices at the item-category level for each item
createItem() functions gain an optional argument to
the function definitions to allow for list-specified data from functions
such as mirt() via a silent
mirt(..., customItemsData) argument
lattice auto.key default now reports lines rather
than points. This is now more consistent when, for example, color theme
is changed to black and white in the trellis window
Added Differential Response Function (DRF) statistics from
upcoming publication (Chalmers, accepted) in a new function entitled
DRF(). These are related to compensatory and
non-compensatory measures of response bias for DIF, DBF, and DTF
available from the SIBTEST framework but for IRT model fitted within the
multiple-group estimation framework
structure argument added to mdirt()
function to allow log-linear models for simplifying the profile
probability model computations
export internally used traditional2mirt() function
to transform a small selection of classical IRT parameterizations into
the slope-intercept form
fix survey.weights input for multiple group models
(reported by Leigh Allison)
fix itemtype = "rsm" block restriction when items
contain unequal category lengths (reported by Aiden Loe)
SIBTEST() computation of beta coefficient changed to
match Shealy and Stout’s (1993) form of p_k * (Y_R - Y_F)
(was previously p_k * (Y_F - Y_R); reported by Craig
Wells). As well, Jmin default is increased to 5 to avoid
conservative Type I error behavior in longer tests
Fix negative chi-square differences in DIF()
function due to non-converged sub-models (reported by Daniel
McKelvey)
M2() function gains a type input to
distinguish between the univariate-bivariate collapsed M2* statistic and
the bivariate only collapsed C2 statistic (Cai and Monro, 2014). C2 can
be useful for polytomous items when there are too few degrees of freedom
to compute the fully collapsed M2*
multipleGroup() gains the dentype
argument to allow for mixture IRT models to be fitted (e.g.,
dentype = 'mixture-3' fits a three-class mixture model).
This also allow modifications such as the zero-inflated IRT model to be
fitted
technical gains a zeroExtreme logical
flag to assign survey weights of 0 to extreme response patterns (FALSE
by default). This may be required when Woods’
extrapolation-interpolation method is used with empirical histograms to
avoid ill defined extrapolated densities
fscores(), itemfit(),
M2(), and residuals() gain a
use_dentype_estimate argument to compute EAP-based scores
whenever the latent trait density was estimated (e.g., via empirical
histograms)
Empirical histograms can now be now scaled to [0,1] using Woods’
extrapolation-interpolation method via the input
dentype = 'empiricalhist_Woods'. Degrees of freedom updated
to reflect this change, and 121 quadrature points are used instead of
the previous 199 for better stability
Semi-parametric Davidian curve estimation of the shape of the
latent trait distribution in unidimensional IRT models was contributed
by Oguzhan Ogreden, as well the associated components used within this
framework (such as the interpolation-extrapolation method described by
Woods, 2006). This estimation method is available through the new
dentype input. mirt also now links to the
dcurver package to obtain the associated computation
functions in the EM algorithm
M2(), itemfit(),
SIBTEST(), and fscores() gain an
na.rm logical to remove rows of missing data
fscores() gains a
append_response.pattern logical to indicate whether
response patterns via the response.pattern input should be
appended to the factor score results
new dentype argument added to estimation-based
functions to specify the density structure of the latent traits (default
is 'Gaussian'). This update breaks the previous
empiricalhist logical option
anova() will accept a single fitted model object and
return information related to AIC, BIC, log-likelihood, etc
Hannan–Quinn (HQ) Criterion added to
anova()
Added multidimensional version of sequential response model
(e.g., Tutz, 1990). Includes itemtype = 'sequential' for
the multidimensional 2PL variant, and itemtype = 'Tutz' for
the Rasch variant
Printing IRT parameters via
coef(mod, IRTpars = TRUE) now computes the delta method for
the g and u terms as well. Interpreting these
is generally not recommended due to their bounded parameter nature (CIs
can be outside the range [0,1]), but are included for posterity
createItem() gains a bytecompile flag
to indicate whether the internal functions should be byte-compiled
before using (default is TRUE)
Special GROUP location holder in
mirt.model() to index the group-level hyper-parameter
terms
key2binary() gains a score_missing flag
to indicate whether missing values should be scored as 0 or left as
NA
createItem() gains support for
derivType = 'symbolic' and
derivType.hss = 'symbolic' to symbolically compute the
gradient/Hessian functions (template code-base contributed by Chen-Wei
Liu)
createItem() gains a derivType.hss
argument to distinguish gradient from Hessian numerical
computations
mdirt() gains support for createItem()
inputs
More plotting points added to default plot() and
itemplot() generics to create smoother traceline
functions
fix simdata() bug for new ggum
itemtype
fix new grouping syntax specification in
mirt.model() when combining START and FIXED (reported by
Garron Gianopulos)
fix IRTpars = TRUE input when itemtype was
Rasch (reported by Benjamin Shear)
mod2values() and passing
pars = 'values' now return data.frame objects
without any factor variables (previously the defaults to
data.frame() were used, which created factors for all
categorical variables by default)
Add monopoly itemtype to fit unidimensional
monotonic polynomial item response model for polytomous data (see Falk
and Cai, 2016)
Add ggum itemtype to fit
unidimensional/multidimensional graded unfolding model (e.g., Roberts
& Laughlin, 1996). Special thanks to David King for providing the
necessary C++ derivative functions and starting values
Square brackets can now be included in the
mirt.model() syntax to indicate group-specific constraints,
priors, starting/fixed values, and so on. These are all of the general
form "CONSTRAIN [group1, group2] = ..." or
"FIXED [group1] = ..."
Added delta method for several classical IRT parameterization
(via coef(model, IRTpars = TRUE)) when a suitable
information matrix was previously estimated
numDeriv dependency removed because
numerical_deriv() now supports a local Richardson
extrapolation type. For best accuracy, this is now used as the default
throughout the package
createItem() and lagrange() now use
Richardson extrapolation as default instead of the less accurate
forward/central difference method
estfun() function added to extract gradient
information directly from fitted objects (contributed by Lennart
Schneider)
simdata() gains an equal.K argument to
redraw data until \(K\) categories are
populated for a given item
Fix initialization of fscores() when using ‘MH’
plausible value imputations (reported by Charlie Iaconangelo)
Various other small bug fixes and performance improvements, fixes for Solaris compatibility, and run a small number of examples during R CMD check
mdirt() now supports latent regression covariate
predictors. Associated function (e.g., fscores()) also
include the latent regression information for discrete models by
default
SIBTEST() replaced with the asymptotic sampling
distribution version of CSIBTEST described by Chalmers
(accepted)
calcNull set to FALSE by
default
Sandwich ACOV estimate now uses the Oakes estimate in the
computations rather than the intensive Louis form (which require
low-level coding of the item-level Hessian terms). Added a new
SE.type = 'sandwich.Louis' for the original sandwich VCOV
estimate in the previous version of mirt
fix latent regression models with QMCEM and MCEM algorithms (reported by Seongho Bae)
fscores() gains a max_theta argument to
apply upper/lower bounds to iterative searching algorithms (issue
reported by Sebastian Born), and a start input to set the
starting values as well (primarily useful in mirtCAT to reduce
iterations)
alabama package optimizer no longer used. Replaced
with generic interface from nloptr package to support
numerous optimizers with greater control instead. Associated inputs
(e.g., alabama_args) replaced as well
Export missing S4 methods for external R packages to import
MDIFF and MDISC no longer in normal ogive metric (1.702 scaling value removed)
added QMC option to residuals() for
LD and LDG2 methods. Also globally set the
number of QMC points to 5000 throughout the package for
consistency
info_if_converged and
logLik_if_converged added to technical list to
indicate whether the information matrix and stochastic log-likelihood
should be computed only when the model converges. Default is now
TRUE for both
added 'MCEM' method for Monte Carlo EM. An
associated MCEM_draws function added to the
technical list as well to control the number of draws
throughout the EM cycles
support for information matrix computations for QMCEM method added (e.g., Oakes, crossprod, Louis)
globally improve numerical efficiency of QMC methods, including the QMCEM estimator
include missing data values in itemfit() for
parametric bootstrap methods to replicate missing data pattern
ensure that nest-logit models have at least 3 categories (reported by Seongho Bae)
convergence set to FALSE if any g > u is found in
the 4PL model
in verbose console output the log-posterior is printed when priors are included in the EM (previously was only the marginal likelihood)
various bug fixes to SIBTEST, particularly for very small sample sizes
anova() LRT comparison gains a bounded
logical to indicate whether a bounded parameter is being compared, as
well as a mix argument to indicate the mixture of
chi-squared distributions
MH-RM estimation optimizer argument can now be
modified to BFGS, L-BFGS-B, and
NR instead of the default NR1
a distinction between the NR optimizer in the EM and
MH-RM applications is included, where the MH-RM now defaults to
NR1 to indicate a single Newton-Raphson update that uses an
RM filtered Hessian term
method = 'SEM' added to perform the stochastic EM
algorithm (first two stages of the MH-RM algorithm setup).
Alternatively, setting technical = list(NCYCLES = NA) when
using the MH-RM algorithm now returns the stochastic EM results
added multidim_matrix option to
iteminfo() to expose computation of information
matrices
bounded parameter spaces handled better when using the NR optimizer
various bug fixes and performance improvements
SE.type = 'Oakes' set as the new default when
computing standard errors via the ACOV matrix when using the EM
algorithm
new SE.type = 'Oakes' to compute Oakes’ 1999 form of
the observed information matrix using central difference approximation.
Applicable for all IRT models (including customized IRT types)
added support for gpcmIRT and rsm
itemtypes for traditional generalized partial credit model and Rasch
rating scale model (which may be modified for a generalized rating scale
model by freeing the slope parameters)
SE.type = 'Fisher' now supports the inclusion of
latent distribution hyper-parameters. Officially, all SE-types now
provide proper hyper-parameter influence in the information
matrices
wrapped various output objects as mirt_df,
mirt_matrix, and mirt_list class to avoid the
need for passing a digits argument for rounding output in
the console. Now, returned objects are never rounded, which makes
writing Monte Carlo simulation code safer in that rounded results will
not appear in the results
added Stone’s (2000) fit statistics and forthcoming PV-Q1 fit
statistics to itemfit()
fscores() when EAP estimates
were used in extremely long (1000+ item) tests. Error now reported when
this happens. Using MAP estimates in these extreme situations is
essentially equivalent and now recommendedadd information about the number of freely estimated parameters
to print() generic
in plot(), auto.key is only disabled
when facet_items = FALSE for dichotomous items. Also,
adjusted ordering of plot(mod, type = 'itemscore') to
reflect actual item ordering in the data
Stretched the theoretical bounds of the y-axis for score-based
functions in plot() and itemplot() (e.g., 3PL
models will now always stretch to S(theta) = 0)
plot(mod, type = 'score') not supports the
which.items input to make expected score plots for bundles
of items
penalized term added to EM algorithm estimation subroutines to
help keep the covariance matrix of the latent trait parameters positive
definite in the M-step (helps convergence properties of the optimizers,
especially ‘L-BFGS-B’). To turn this penalized term off use
technical = list(keep_vcov_PD = FALSE)
added type = 'itemscore' to plot()
generic to plot faceted version of the item scoring functions.
Particularly useful when investigating DIF with
multipleGroup()
better support for splines itemtype in
multiple-group models
fix problem with ‘EAPsum’ in fscores() when
response.pattern input supplied (reported by Eva de
Schipper)
plot(mod, type = 'rxx') now uses the latent variance
in the computations (reported by Amin Mousavi)
fix syntax input when customized IRT models are supplied
df adjustment for the S_X2 item-fit
statistic for models where the latent trait hyper-parameters have been
estimated
itemfit() and personfit() properly
detect dichotomous Rasch models which have been defined with the
constrained slopes approach
argument 'fit_stats' now used in
itemfit() to replace longer list of logicals (e.g.,
itemfit(mod, S_X2 = FALSE, X2 = TRUE, infit = FALSE, ...)).
Now fit stats are explicitly requested through a character vector input.
Default still uses the S_X2 statistic
when using 'lnorm' prior lower bound automatically
set to 0, and with 'beta' prior the lower and upper bounds
are set to [0,1]
mdirt() now uses optimizer = 'nlminb'
by default
revert using default ‘penalized version of the BFGS algorithm’ instead of L-BFGS-B when box-constraints are used (introduced in version 1.19)
Neale & Miller 1997 approximation added to
PLCI() (default still computes exact PL CIs)
type = 'score' supported for multiple group models
in itemplot()
added poly2dich function to quickly change
polytomous response data to a comparable matrix of dichotomous response
data
a penalized version of the BFGS algorithm is now used instead of the L-BFGS-B when upper and lower bounds are included (provides more robust estimates)
the variances of the orthogonal factors in bfactor()
can now be freely estimated. This allows modeling of designs such as the
testlet response model (example included in the documentation)
new spline itemtype to model B-spline response
functions for dichotomous models. Useful for diagnostic purposes after
detecting item-misfit. Additional arguments can be passed to the
spline_args list input to control the behaviour of the
splines for each item. Currently limited to unidimensional models
only
fscores() gains a plausible.type
argument to select between normal approximation PVs or
Metropolis-Hastings samples (suggested by Yang Liu)
mdirt() has been modified to support DINA, DINO,
located latent class, and other diagnostic classification models.
Additionally, the customTheta input required to build
customized latent class patterns has been changed from the previously
cumbersome
mdirt(..., technical = list(customTheta = Theta)) to simply
mdirt(..., customTheta = Theta)
simdata() gains a prob.list input to
supply a list of matrices with probability values to be sampled from
(useful when specialized response functions outside the package are
required)
simdata() supports ‘lca’ itemtypes for latent class
model generation
improved M2 accuracy when latent trait variances are estimated
corrected behaviour of M2() when linear constraints
are applied (M2 test was previously too conservative when constraints
were used). This affects single as well as multiple-group models
(reported by Rudolf Debelak)
add plausible values for latent class and related models
estimated from mdirt()
multipleGroup() throws proper error when vertical
scaling is not identified correctly due to NAs
S-X2 itemfit statistic fix when very rare expected categories appear (reported by Seongho Bae)
mdirt() function now includes explicit parameters
for the latent class intercepts (in log-form). This implies that correct
standard errors can be computed using various methods (e.g., SEM,
Richardson, etc)
new customGroup() function to define hyper-parameter
objects for the latent trait distributions (generally assumed to be
Gaussian with a mean and covariance structure)
new boot.LR() function to perform a parametric
bootstrap likelihood-ratio test between nested models. Useful when
testing nested models which contain bounded parameters (e.g., testing a
3PL versus a 2PL model)
adjust the lagrange() function to use the full
information matrix (was previously only a quasi-lagrange
approximation)
greatly improved speed in simdata(), consequently
changes the default seed
fix crash error in mirtmirt() for multidimensional
models with lr.random effects (reported by Diah Wihardini)
expbeta prior starting values fix by setting to the
mean of the prior rather than the mode (reported by Insu Paek)
itemfit() function reworked so that all statistics
have their own input flag (e.g., Zh = TRUE,
infit = TRUE, etc). Additionally, only S-X2 is computed by
default and X2/G2 (and the associated graphics and tables) are computed
using 10 fixed bins
added empirical.table argument to return tables of
expected/observed values for X2 and
G2
group.bins and group.fun argument added
to itemfit() to control the size of the bins and the
central tendancy function for X2 and G2
computations
'expbeta' option added to implement a beta prior
specifically for the g and u parameters which
internally have been transformed to logits (performes the back
transformation before computing the values)
check whether multiple-group models contain enough data to estimate parameters uniquely when no constraints are applied
set the starting values the same when using parprior
list or mirt.model() syntax (reported by Insu
Paek)
empirical_ES() function added for effect size
estimates in DIF/DBF/DTF analyses (contributed by Adam Meade)
standardized loadings not correct when factor correlations included in confirmatory models (reported by Seongho Bae)
MDISC and MDIFF values were missing the
1.702 multiplicitive constant (reported by Yi-Ling Cheng)
fix information trace-lines in multiple-group plots (reported by Conal Monaghan)
suppress standard errors in exploratory models when
rotate != 'none' (suggested by Hao Wu)
sequential schemes in DIF() generated the wrong
results (reported by Adam Meade)
M2() was not properly accounting for latent variance
terms (reported by Ismail Cuhadar)
enable lr.random input to mixedmirt()
for multilevel-IRT models which are not from the Rasch family
add common vcov() and logLik()
methods
latent regression EM models now have standard error computation supporte with the ‘complete’, ‘forward’, ‘central’, and ‘Richardson’ methods
new areainfo() function to compute the area under
information curves within specified ranges (suggested by Conal
Monaghan)
method = 'BL' supported for multiple-group models.
As well, SE.type = 'numerical' included to return the
observed-data ACOV matrix from the call to optim() (can
only be used when the BL method is selected)
new SE.type = 'FMHRM' to compute information matrix
based on a fixed number of MHRM draws, and an associated
technical = list(MHRM_SE_draws) argument has been added to
control the number of draws
added lagrange (i.e., score) test function for
testing whether parameters should be freed in single and multiple group
models estimated with the EM algorithm
numerical_deriv function made available for simple
numerical derivatives, which may be useful when defining fast custom
itemtype derivative terms
SE.type used to compute the ACOV matrix gained three
numerical estimates for the forward difference (‘forward’), central
difference (‘central’), and Richardson extropolation
(‘Richardson’)
added SIBTEST and crossed-SIBTEST procedures with the new
function SIBTEST()
added empirical_plot function for building empirical
plots (with potential smoothing) when conditioning on the total
score
more low-level elements included in extract.mirt()
function
added grsmIRT itemtype for classical graded rating
scale form (contributed by KwonHyun Kim)
added missing analytic Hessian terms when gpcm_mats
are used (contributed by Carl Falk)
technical = list(removeEmptyRows = TRUE) (reported by Aaron
Kaat)Data, Model, Fit, and so
on. Additionally, the information matrix has slot has been removed in
favour of providing the asymptotic covariance matrix (a.k.a., the
inverse of the information matrix)added extract.mirt() function to allow more
convenient extracting of internal elements
crossprod SE.type now incorporates latent variable
information (replaces NA placeholders)
changed the default full.scores = FALSE argument to
TRUE in fscores()
added profile argument to plot() for
mdirt() objects so that profile plots can be
generated
converge_info option added to fscores()
to return convergence information
add removeEmptyRows option to technical
list
return a vector of NAs when WLE estimation has a
Fisher information matrix with a determinant of 0 (reported by
Christopher Gess)
fix df in multiple-group models with crossed between/within constrains (reported by Leah Feuerstahler)
compute residuals when responses are sparse, and return
NaN when residual could not be computed (reported by Aaron
Kaat)
adjust plausible values format for multiple group objects
simdata() gains a model input to impute
data from pre-organized models (useful in conjunction with mirtCAT or to
generate datasets from already converged models). Also gains a
mins argument to specify what the lowest category should be
for each item if model is not supplied (default is
0)
number of SEMCYCLES increased from 50 to 100 in the
MH-RM algorithm, and RM gain rate changed from c(.15, .65)
to c(.1, .75)
further improve item fit statistics when using imputations
facet plots now try to keep the items in their respective order
panel theme for lattice plots changed from default to a lighter blue colour, and legend now automatically placed on the right hand side rather than the top
when using prior distributions, starting values now automatically set equal to the mode of the prior distribution, and appropriate lower and upper parameter bounds are supplied
added NEXPLORE term to mirt.model() to
specify exploratory models via the syntax
add itemGAM() function to provide a non-linear
smoother for better understanding mis-functioning items (and without
loosing established precision by reverting to purely non-parametric IRT
methods)
category scores are now automatically recoded to have spaces of 1, and a message is printed if/when this occurs
added MDISC() and MDIFF()
functions
the inclusion of prior parameter distributions will now report
the log-posterior rather than the log-likelihood. Functions such as
anova() will also report Bayesian criteria rather than the
previous likelihood-based model comparison statistics
impute argument in itemfit() and
M2() now use plausible values instead of point
estimates
START syntax element in mirt.model()
now supports multiple parameters, and FIXED argument added
to declare parameters as ‘fixed’ at their staring values
added LBOUND and UBOUND syntax support
in mirt.model()
report proper lower and upper bounds in starting values data
frame and from mod2values()
invariance argument to bfactor() now
automatically indexes the second-tier factors to make multiple-group
testing with bfactor() easier
remove rotate and Target arguments from
model objects, and pass these only through axillary functions such as
summary(), fscores(), etc
model based arguments now can be strings, which are
passed to mirt.model(). This is now the preferred method
for defining models syntactically, though the previous methods will
still work
integration range (theta_lim) globally set to
c(-6, 6), and number of default quadrature nodes have
systematically increased in parameter estimation functions. This will
slightly change some numerical results, but provides more consistence
throughout the package
add theta_lim arguments to various
functions
better control of QMC grid, and more effective usage for higher dimensions
internal code organization now makes it easier to add user
defined itemtypes (which can be natively added into the
package, if requested)
fix conservative imputation standard errors in
itemfit() and M2() (reported by Irshad
Mujawar)
fixed plausible value draws for multidimensional latent regression models (reported by Tongyun Li)
don’t allow crossprod, Louis, or sandwich information matrices when using custom item types (reported by Charlie Rutgers)
when using coef(mod, printSE=TRUE) the
g and u parameters are relabeled to
logit(g) and logit(u) to represent the
internal labels
added various facet plots for three dimensional models to
plot() generic
support optimizer = 'nlminb', and pass optimizer
control arguments to a contol list
added fixef() function to extract expected values
implied by the fixed effect parameters in latent regression
models
added gpcm_mats argument to estimation functions for
specifying a customize scoring pattern for multidimensional generalized
partial credit models
added custom_theta input to fscores()
for including customized integration grids
add a suppress argument to residuals()
and M2() to suppress local dependence values less than this
specific value
print a message in DIF() and DTF() when
hyper-parameters are not freely estimated in focal groups
constraits for hetorogenous item names added to
mirt.model() syntax
WLE support for multidimensional models added
added 'SEcontour' argument to plot()
generic
use NA’s in fscores() when response patterns contain
all NA responses (suggested by Tomasz Zoltak)
S-X2 in itemfit() now returns appropriate values for
multiple-group models
multidimensional plausible value imputation fix (reported by KK Sasa)
plot(..., type = 'infotrace') for multiple group
objects fixed (reported by Danilo Pereira)
fscores() nows accepts
method = "plausible" to draw a single plausible value
set
plot() default type is now score, and
will accept rotation arguments for exploratory models (default rotation
is 'none')
imputeMissing() supports a list of plausible values
to generate multiple complete datasets
new custom_den input to fscores() to
use custom prior density functions for Bayesian estimates
more optimized version of the ‘WLE’ estimator in
fscores()
empirical reliability added when method = 'EAPsum'
in fscores()
new START argument in mirt.model() for
specifying simple starting values one parameter at a time
fix carryover print-out error in summary() when
confirmatory models were estimated
bound contraints not were not included for group hyper-parameters (reported by KK Sasa)
improved estimation efficiency when using MH-RM algorithm. As a result, the default seed was changed, therefore results from previous versions will be slightly different
objects of class ‘ExploratoryClass’ and ‘ConfirmatoryClass’ have
been merged into a single class ‘SingleGroupClass’ with an
exploratory logical slot
the technical = list(SEtol) criteria for
approximating the information matrix was lowered to 1e-4 in
mixedmirt() to provide better standard error
estiamtes
boot.mirt now uses the optimizer used to estimate
the model (default previously was EM)
mixedmirt now supports interaction effects in random
intercepts, including cross-level interactions
added averageMI() function to compute multiple
imputation averages for the plausible values methodology using Rubin’s
1987 method
plausible value imputation now available in
fscores() using the new plausible.draws
numeric input
add return.models argument to DIF() to
return estimated models with free/constrained parameters
latent regression models added to mixedmirt() for
non-Rasch models using the new lr.formula input
mirt.model() syntax can now define within individual
item equality constraints by using more than 1 parameter specification
name in the syntax
latent regression models added to mirt() function by
using the new covdata and formula
inputs
added confidence envelope plots to PLCI.mirt, and
throw warnings when intervals could not be located
coef() now accepts a simplify logical,
indicating whether the items should be collapsed to a matrix and
returned as a list of length 2 (suggested by Michael Friendly)
bias correction in variance estimates mixedmirt when
random effects are included (reported by KK Sasa)
fix missing data imputation bug in itemfit()
(reported by KK Sasa)
M2 statistic for bifactor/two-tier models was overly conservative
better checks for numerical underflow issues
use triangle 0’s for identifying exploratory IFA models. As such, standard errors/condition numbers for exploratory models can be estimated again
sirt package added to suggests list. Special thanks
to Alexander Robitzsch (author of sirt) for developing
useful wrapper functions for mirt such as
mirt.wrapper.coef(), tam2mirt(), and
lavaan2mirt(). As well, many examples in sirt
demonstrate the possibility of estimating specialized IRT models with
mirt, such as the: Ramsay quotient, latent class, mixed
Rasch, located latent class, probabilistic Guttman, nonparametric,
discrete graded membership, and multidimensional IRT discrete traits,
DINA, and Rasch copula models.
exploratory IRT models are no longer rotated by default in
coef(), and now requires an explicit rotate
argument
computation of S_X2 statistic in
itemfit now much more stable for polytomous item
types
support for the plink package now unofficially
dropped because it was removed from CRAN
data inputs are now required to have category spacing codings exactly equal to 1 (e.g., [0, 1, 2, …]; patterns such as [0, 2, 3] which are implicitly missing spaces are now invalid)
mdirt function added to model discrete latent
variables such as latent class analysis for dichotomous and polytomous
items. Can be used to model several other discrete IRT models as well,
such as the located latent class model, multidimensional IRT with
discrete traits, DINA models, etc. See the examples and documentation
for details
axillary support for DiscreteClass objects added to
itemfit(), M2(), fscores(),
wald(), and boot.mirt()
the S-X2 statistic available in itemfit() has been
generalized to include multidimensional models
the method 'QMCEM' has been added for quasi-Monte
Carlo integration in mirt() and
multipleGroup() for estimating higher dimensional models
with greater accuracy (suggested by Alexander Robitzsch). Several
axillary function such as fscores(),
itemfit(), and M2() also now contain an
QMC argument (or will accept one through the … argument) to
use the same integration scheme for better accuracy in higher
dimensional models
nonlinear parameter constraints for EM estimation can be
specified by using the Rsolnp and alabama
packages by passing optimizer = 'solnp' and
optimizer = 'alabama', as well as the relevant package
arguments through the solnp_ags and
alabama_ags list inputs
itemnames argument added to
mirt.model() to allow model specifications using raw item
names rather than location indicators
accelerate argument changed from logical to
character vector, now allowing three potential options: ‘Ramsay’
(default), ‘squarem’, and ‘none’ for modifying the EM acceleration
approach
fixed bug in bfactor() starting values when NAs were
specified in the model argument
adjust overly optimistic termination criteria in EM algorithm
for efficiency, the Hessian is no longer computed in
fscores() unless it is required in the returned
object
estimation with method = 'MHRM' now requires and
explicitly SE=TRUE call to compute the information matrix.
The matrix is now computed using the ML estimates rather than
approximated sequentially after each iteration (very unstable), and
therefore a separate stage is performed. This provides much better
accuracy in the computations
new extract.group() function to extract a single
group object from an objects previously returned by
multipleGroup()
return the SRMSR statistic in M2() along with the
residual matrix (suggested by Dave Flora)
accept Etable default input in
customPriorFun (suggested by Alexander Robitzsch)
vignette files for the package examples are now hosted on Github
and can be accessed by following the link mentioned in the vignette
location in the index or ?mirt help file
E-step is now computed in parallel (if available) following a
mirtCluster() definition
run no M-step optimizations by passing TOL = NaN.
Useful to have the model converge instantly with all parameters exactly
equal to the starting values
confidence envelope plots in itemplot() generate
shaded regions instead of dotted lines, and confidence interval plots
added to plot() generic through the MI
input
passes to fscores() slightly more optimized for
upcoming mirtCAT package release
method = 'EAPsum' argument to fscores()
support for multidimensional models
fix forcing all SEs MHRM information matrix computations to be positive
imputeMissing() crash fix for multiple-group
models
fix divide-by-0 bug in the E-step when number of items is large
fix crash in EM estimation with
SE.type = 'MHRM'
calculating the information matrix for exploratory item factor analysis models has been disabled since the rotational indeterminacy of the model results in improper parameter variation
changed default theta_lim to c(-6,6)
and number of quadrature defaults increased as well
@Data slot added for organizing data based
arguments. Removed several data slots from estimated objects as a
consequence
removed ‘Freq’ column when passing a
response.pattern argument to
fscores()
increase number of Mstep iterations proportionally in quasi-Newton algorithms as the estimation approaches the ML location
‘rsm’ itemtype removed for now until optimized version is implemented
link to mirt vignettes on Github have been
registered with the knitr package and are now available
through the package index
optimizer argument added to estimation function to
switch the default optimizer. Multiple optimizers are now available,
including the BFGS (EM default), L-BFGS-B, Newton-Raphson, Nelder-Mead,
and SANN
new survey.weights argument can be passed to
parameter estimation functions (i.e., mirt()) to apply
so-called stratification/survey-weights during estimation
returnList argument added to simdata()
to return a list containing the S4 item objects, Theta matrix, and
simulated data
support custom item type fscores() computations when
response.pattern is passed instead of the original
data
impute option for itemfit() and
M2() to estimate statistics via plausible imputation when
missing data are present
multidimensional ideal-point models added for dichotomous items
M2* statistic added for polytomous item types
Bock and Lieberman ('BL') method argument added (not
recommend for serious use)
large bias correction in information matrix and standard errors for models that contain equality constraints (standard errors were too high)
drop dimensions fix for nested logit models
default SE.type changed to crossprod
since it is better at detecting when models are not identified compared
to SEM, and is generally much cheaper to compute for larger
models
M-step optimizer now automatically selected to be ‘BFGS’ if there are no bounded parameters, and ‘L-BFGS-B’ otherwise. Some models will have notably different parameter estimates because of this, but should have nearly identical model log-likelihoods
better shiny UI which adapts to the itemtype specifically, and allows for classical parameter inputs (special thanks to Jonathan Lehrfeld for providing code that inspired both these changes)
scores.only option now set to TRUE in
fscores()
type = 'score' for plot generics no longer adjusts
the categories for expected test scores
M-step optimizer in EM now deters out-of-order graded response model intercepts (was a problem if the startvalues were too far from the ML estimate in graded models)
return.acov logical added to fscores()
to return a list of matrices containing the ACOV theta values used to
compute the SEs (suggested by Shiyang Su)
printCI logical option to summary() to
print confidence intervals for standardized loadings
new expected.test() function, which is an extension
of expected.item() but for the whole test
mirt.model() syntax supports multiple * combinations
in COV = for more easily specifying covariation blocks
between factors. Also allows variances to be freed by specifying the
same factor name, e.g., F*F
full.scores.SE logical option for
fscores() to return standard errors for each
respondent
multiple imputation (MI) option in fscores(), useful
for obtaining less biased factor score estimates when model parameter
variability is large (usually due to smaller sample size)
group-level (i.e., means/covariances) equality constrains are now available for the EM algorithm
theta_lim input to plot(),
itemplot(), and fscores() for modifying range
of latent values evaluated
personfit() crash for multipleGroup objects since
itemtype slot was not filled (reported by Michael Hunter)
fix crash in two-tier models when correlations are estimated (reported by David Wu)
R 3.1.0 appears to evaluate List objects differently at the c level causing strange behaviour, therefore slower R versions of some internal function (such as mirt:::reloadPars()) will be used until a patch is formed
behaviour of mvtnorm::dmvnorm changed as of version
0.9-9999, causing widely different convergence results. Similar versions
of older mvtnorm functions are now implemented instead
fitIndices() replaced with M2()
function, and currently limited to only dichotomous items of class
‘dich’
bfactor() default SE.type set to ‘crossprod’ rather
than ‘SEM’
generalized partial credit models now display fixed scoring coefs
TOL convergence criteria moved outside of the
technical input to its own argument
restype argument to residuals() changed
to type to be more consistent with the package
removed fitted() since
residuals(model, type = 'exp') gives essentially the same
output
mixedmirt has SE set to TRUE by default
to help construct a more accurate information matrix
if not specified, S-EM TOL dropped to
1e-6 in the EM, and SEtol = .001 for each
parameter to better approximate the information matrix
two new SE.type inputs: ‘Louis’ and ‘sandwich’ for
computing Louis’ 1982 computation of the observed information matrix,
and for the sandwich estimate of the covariance matrix
as.data.frame logical option for coef()
to convert list to a row-stacked data.frame
type = 'scorecontour' added to plot()
for a contour plot with the expected total scores
type = 'infotrace' added to itemplot()
to plot trace lines and information on the same plot, and
type = 'tracecontour' for a contour plot using trace lines
(suggested by Armi Lantano)
mirt.model() support for multi-line inputs
new type = 'LDG2' input for residuals()
to compute local dependence stat based on G2 instead of X2, and
type = 'Q3' added as well
S-EM computation of the information matrix support for latent
parameters, which previously was only effective when estimation
item-level parameters. A technical option has also been added to force
the information matrix to be symmetric (default is set to
TRUE for better numerical stability)
new empirical.CI argument in itemfit()
used when plotting confidence intervals for dichotomous items (suggested
by Okan Bulut)
printSE argument can now be passed to
coef() for printing the standard errors instead of
confidence intervals. As a consequence, rawug is
automatically set to TRUE (suggested by Olivia
Bertelli)
second-order test and condition number added to estimated objects when an information matrix is computed
tables argument can be passed to
residuals() to return all observed and expected tables used
in computing the LD statistics
using scores.only = TRUE for multipleGroup objects
returns the correct person ordering (reported by Mateusz
Zoltak)
read.mirt() crash fix for multiple group analyses
objects (reported by Felix Hansen)
updated math for SE.type = 'crossprod'
facet_items argument added to plot() to control
whether separate plots should be constructed for each item or to merge
them onto a single plot
three dimensional models supported in itemplot() for
types trace, score, info, and
SE
new DIF() function to quicky calculate common differential item functioning routines, similar to how IRTLRDIF worked. Supports likelihood ratio testings as well as the Wald approach, and includes forward and backword sequential DIF searching methods
added a shiny = TRUE option to
itemplot() to run the interactive shiny applet. Useful for
instructive purposes, as well as understanding how the internal
parameters of mirt behave
type = 'trace' and type = 'infotrace'
support added to plot generic for multiple group
objects
fscores(..., method = 'EAPsum') returns observed and
expected values, along with general fit statistics that are printed to
the console and returned as a ‘fit’ attribute
removed multinomial constant in log-likelihood since it has no influence on nested model comparisons
SE.type = 'crossprod' and Fisher added
for computing the parameter information matrix based on the variance of
the Fisher scoring vector and complete Fisher information matrix,
respectively
customPriorFun input to technical list now available
for utilizing user defined prior distribution functions in the EM
algorithm
empirical histogram estimation now available in
mirt() and multipleGroup() for unidimensional
models. Additional plot type = 'empiricalhist' also added
to the plot() generic
re-implement read.mirt() with better consistency
checking between the plink package
starting values for multipleGroup() now returns
proper estimated parameter information from the invariance
input argument
remove as.integer() in MultipleGroup df
slot
pass proper item type when using custom pattern calles in
fscores()
return proper object in personfit when gpcm models used
GenRandomPars logical argument now supported in the
technical = list() input. This will generate random
starting values for freely estimated parameters, and can be helpful to
determine if obtained solutions are local minimums
seperate free_var and free_cov
invariance options available in multipleGroup
new CONSTRAIN and CONSTRAINB arguments
in mirt.model() syntax for specifying equality constraints
explicitly for parameters accross items and groups. Also the
PRIOR = ... specification was brought back and uses a
similar format as the new CONSTRAIN options
plot(..., type = 'trace') now supports polytomous
and dichotomous tracelines, and type = 'infotrace' has a
better y-axis range
removed the ‘1PL’ itemtype since the name was too ambiguous. Still possible to obtain however by applying slope constraints to the 2PL/graded response models
plot() contains a which.items argument to specify
which items to plot in aggregate type, such as 'infotrace'
and 'trace'
fitIndicies() will return CFI.M2 and
TLI.M2 if the argument calcNull = TRUE is
passed. CFI stats also normed to fall between 0 and 1
data.frame returned from mod2values() and
pars = 'values' now contains a column indicating the
internal item class
use ginv() from MASS package to improve accuracy in
fitIndices() calculation of M2
fix error thrown in PLCI.mirt when parameter value
is equal to the bound
fix the global df values, and restrict G2 statistic when tables are too sparse
PLCI.mirt() function added for computing profiled
likelihood standard errors. Currently only applicable to unidimensional
models
prior distributions returned in the pars = 'values'
data.frame along with the input parameters, and can be edited and
returned as well
full.scores option for residuals() to compute
residuals for each row in the original data
bfactor() can include an additional model argument
for modeling two-tier structures introduced by Cai (2010), and now
supports a 'group' input for multiple group
analyses
added a general Ramsey (1975) acceleration to EM estimation by
default. Can be disable with accelerate = FALSE (and is
done so automatically when estimating SEM standard errors)
renamed response.vector to response.pattern in
fscores(), and now supports matrix input for computing
factor scores on larger data sets (suggested by Felix Hansen)
total.info logical added to iteminfo() to return
either total item information or information from each category
mirt.model() supports the so-called Q-matrix input
format, along with a matrix input for the covariance terms
MH-RM algorithm now accessible by passing
mirt(..., method = 'MHRM'), and confmirt()
function removed completely. confmirt.model() also renamed
to mirt.model()
support for polynomial and interaction terms in EM estimation
lognormal priors may now be passed to parprior
iterative computations in fscores() can now be run
in parallel automatically following a mirtCluster()
definition
mirtCluster() function added to make utilizing
parallel cores more convenient. Globally removed the cl argument for
multi-core objects
updated documentation for data sets by adding relevant examples, and added Bock1997 data set for replicating table 3 in van der Linden, W. J. & Hambleton, R. K. (1997) Handbook of modern item response theory
general speed improvements in all functions
WLE estimation fixed and now estimates extreme response patterns
exploratory starting values no longer crash in datasets with a huge number of NAs, which caused standard deviations to be zero
math fix for beta priors
support for random effect predictors now available in
mixedmirt(), along with a randef() function
for computing MAP predictions for the random parameters
EAPsum support in fscores() for mixed item
types
for consistency with current IRT software (rather than TESTFACT and POLYFACT), the scaling constant has been set to D = 1 and fixed at this value
nominal.highlow option added to specify which categories are the highest and lowest in nominal models. Often provide better numerical stability when utilized. Default is still to use the highest and lowest categories
increase number of draws in the Monte Carlo calculation of the log-likelihood from 3000 to 5000
when itemtype all equal ‘Rasch’ or ‘rsm’ models the latent variance parameter(s) are automatically freed and estimated
mixedmirt() more supportive of user defined R
formulas, and now includes an internal ‘items’ argument to create the
item design matrix used to estimate the intercepts. More closely mirrors
the results from lme4 for Rasch models as well (special thanks to Kevin
Joldersma for testing and debugging)
drop.zeros option added to extract.item and itemplot
to reduce dimensionality of factor structures that contain slopes equal
to zero
EM tolerance (TOL argument) default dropped to .0001 (originally .001)
type = 'score' and type = 'infoSE'
added to plot() generic for expected total score and joint
test standard error/information
custom latent mean and covariance matrix can be passed to
fscores() for EAP, MAP, and EAPsum methods. Also applies to
personfit() and itemfit() diagnostics
scores.only option to fscores() for returning just
the estimated factor scores
bfactor can include NA values in the model to omit the estimation of specific factors for the corresponding item
limiting values in z.outfit and z.infit statistics for small sample sizes (fix suggested by Mike Linacre)
missing data gradient bug fix in MH-RM for dichotomous item models
global df fix for multidimensional confirmatory models
SEM information matrix computed with more accuracy (M-step was not identical to original EM), and fixed when equality constrains are imposed
new '#PLNRM' models to fit Suh & Bolt (2010)
nested logistic models
'large' option added to estimation functions. Useful
when the datasets being analysed are very large and organizing the data
becomes a computationally burdensome task that should be avoided when
fitting new models. Also, overall faster handling of datasets
plot(), fitted(), and
residuals() generic support added for MultipleGroup
objects
CFI and X2 model statistics added, and output now includes fit stats w.r.t. both G2 and X2
z stats added for itemfit/personfit infit and outfit statistics
supplemented EM (‘SEM’) added for calculating information matrix from EM history. By default the TOL value is dropped to help make the EM iterations longer and more stable. Supports parallel computing
added return empirical reliability (returnER) option
to fscores()
plot() supports individual item information trace
lines on the same graph (dichotomous items only) with the option
type = 'infotrace'
createItem() function available for defining item
types that can be passed to estimation functions. This can be used to
model items not available in the package (or anywhere for that matter)
with the EM or MHRM. Derivatives are computed numerically by default
using the numDeriv package for defining item types on the fly
Mstep in EM moved to quasi-Newton instead of my home grown MV Newton-Raphson approach. Gives more stability during estimation when the Hessian is ill-conditioned, and will provide an easier front-end for defining user rolled IRT models
small bias fix in Hessian and gradients in mirt()
implementation causing the likelihood to not always be increasing near
maximum
fix input to itemplot() when object is a list of
model objects
fixed implementation of infit and outfit Rasch statistics
order of nominal category intercepts were sometimes backwards. Fixed now
S_X2 collapsed cells too much and caused negative df
response.vector input now supports NA inputs
(reported by Neil Rubens)
S-X2 statistic computed automatically for unidimensional models via itemfit()
EAP for sum-scores added to fscores() with method = ‘EAPsum’. Works with full.scores option as well
improve speed of estimation in multipleGroup() when latent means/variances are estimated
multipleGroup(invariance = ’’) can include item names to specify which items are to be considered invariant across groups. Useful for anchoring and DIF testing
type = ‘trace’ option added to plot() to display all item trace lines on a single graph (dichotomous items only)
default estimation method in multipleGroup() switched to ‘EM’
boot.mirt() function added for computing bootstrapped standard errors with via the boot package (which supports parallel computing as well), as well as a new option SE.type = ’’ for choosing between Bock and Lieberman or MHRM type information matrix computations
indexing items in itemplot, itemfit, and extract.item can be called using either a number or the original item name
added probtrace() function for front end users to generate probability trace functions from models
plotting item tracelines with only two categories now omits the lowest category (as is more common)
parallel option passed to calcLogLik to compute Monte Carlo log-likelihood more quickly. Can also be passed down the call stack from confmirt, multipleGroup, and mixedmirt
Confidence envelopes option added to itemplot() for trace lines and information plots
lbound and ubound parameter bounds are now available to the user for restricting the parameter estimation space
mod2values() function added to convert an estimated mirt model into the appropriate data.frame used to determine parameter estimation characteristics (starting values, group names, etc)
added imputeMissing() function to impute missing values given an estimated mirt model. Useful for checking item and person fit diagnostics and obtaining overall model fit statistics
allow for Rasch itemtype in multidimensional confirmatory models
oblimin the new default exploratory rotation (suggested by Dave Flora)
more flexible calculation of M2 statistic in fitIndicies(), with user prompt option if the internal variables grow too large and cause time/RAM problems
read.mirt() fixed when objects contain standard errors (didn’t properly line up before)
mixedmirt() fix when COV argument supplied (reported by Aaron Kaat)
fix for multipleGroup when independent groups don’t contain all potential response options (reported by Scot McNary)
prevent only using ‘free_means’ and ‘free_varcov’ in multipleGroup since this would not be identified without further constraints (reported by Ken Beath)
all dichotomous, graded rating scale, (generalized) partial credit, rating scale, and nominal models have been better optimized
wald() will now support information matrices that contain constrained parameters
confmirt.model() can accept a string inputs, which may be useful for knitr/sweave documents since the scan() function tends to hang
multipleGroup() now has the logical options bfactor = TRUE to use the dimensional reduction algorithm for when the factor pattern is structured like a bifactor model
new fitIndices() function added to compute additional model fit statistics such as M2
testinfo() function added for test information
lower bound parameters under more stringent control during estimation and are bounded to never be higher than .6
infit and outfit stats in itemfit() now work for Rasch partial credit and rating scale models
Rasch rating scale models can now be estimated with potential rsm.blocks (same as grsm model). “Generalized” rating scale models can also be estimated, though this requires manipulating the starting values directly
added sample size adjusted BIC (SABIC) information statistics
new mixedmirt() function for estimating IRT models with person and item level (e.g., LLTM) covariates. Currently only supports fixed effect predictors, but random effect predictors are being developed
more structured output when using the anova() generic
item probability functions now only permit permissible values, and models may converge even when the log-likelihood decreases during estimation. In the EM if the model does not have a strictly increasing log-likelihood then a warning message will be printed
infit and outfit statistics are now only applicable to Rasch models (as they should be), and in itemfit/personfit() a ‘method’ argument has been added to specify which factor score estimates should be used
read.mirt() re-added into the package to allow for translating estimated models into a format usable by the plink package
test standard error added to plot() generic using type = ‘SE’, and expected score plot added to itemplot() using type = ‘score’
weighted likelihood estimation (WLE) factor scores now available (without standard errors)
removed the allpars option to coef() generics and only return a named list with the (possibly rotated) item and group coefficients
information functions slightly positively biased due to logistic constant adjustment, fixed for all models. Also, information functions are now available for almost all item response models (mcm items missing)
constant (D) used in estimating logistic functions can now be modified (default is still 1.702)
partcomp models recently broken, fixed now
more than one parameter can now be passed to parprior to make specifying identical priors more convenient
relative efficiency plots added to itemplot(). Works directly for multipleGroup analysis and for comparing different item types (e.g., 1PL vs 2PL) can be wrapped into a named list
infit and outfit statistics added to personfit() and itemfit()
empirical reliability printed for each dimension when fscores(…, fulldata = FALSE) called
better system to specify fixed/free parameters and starting values using pars = ‘values’. Should allow for much better simulation based work
graded model type rating scale added (Muraki, 1990) with optional estimation ‘blocks’. Use itemtype = ‘grsm’, and the grsm.block option
for multipleGroup(), optional input added to change the current freely estimated parameters to values of a previously computed model. This will save needless iterations in the EM and MHRM since these parameters should be much closer to the new ML estimates
itemplot() supports multipleGroup objects now
analytical derivatives much more stable, although some are not yet optimized
estimation bug fix in bfactor(), and slight bias fix in mirt() estimation (introduced in version 0.4.0 when multipleGroup() added)
updated documentation and beamer slide show included for some background on MIRT and some of the packages capabilities
labels added to coef() when standard errors not computed. Also allpars = TRUE is now the default
kernel estimation moved entirely to one method. Much easier to maintain and guarantees consistency across methods (i.e., no more quasi-Newton algorithms used)
Added itemfit() and personfit() functions for uni and multidimensional models. Within itemfit empirical response curves can also be plotted for unidimensional models
Wrapped itemplot() and fscores() into S3 function for better documentation. Also response curve now are all contained in individual plots
Added free.start list option for all estimation functions. Allows a quicker way to specify free and fixed parameters
Added iteminfo() and extract.item() to calculate the item information and extract desired items
Multiple group estimation available with the multipleGroup() function. Uses the EM and MHRM as the estimation engines. The MHRM seems to be faster at two factors+ though and naturally should be more accurate, therefore it is set as the default
wald() function added for testing linear constraints. Useful in situations for testing sets of parameters rather than estimating a new model for a likelihood ratio test
Methods that use the MHRM can now estimate the nominal, gpcm, mcm, and 4PL models
fscores computable for multiple group objects and in general play nicer with missing data (reported by Judith Conijn). Also, using the options full.scores = TRUE has been optimized with Rcpp
Oblique rotation bug fix for fscores and coef (reported by Pedro A. Barbetta)
Added the item probability equations in the ?mirt documentation for reference
General bug fixes as usual that were spawned from all the added features. Overall, stay frosty.
Individual classes now correspond to the type of methods: ExploratoryClass, ConfirmatoryClass, and MultipleGroupClass
plot and itemplot now works for confmirt objects
mirt can now make use of confmirt.model specified objects and hence be confirmatory as well
stochastic estimation of factor scores removed entirely, now only quadrature based methods for all objects. Also, bfactor returned objects now will estimate all the factors scores instead of just the general dimension
Standard errors for mirt now automatically calculated (borrowed from running a tweaked MHRM run)
radically changed the underlying mechanisms for the estimation functions and in doing so have decided that polymirt() was redundant and could be replaced completely by calling confmirt(data, number_of_factors). The reason for the change was to facilitate a wider range or MIRT models and to allow for easier extensions to future multiple group analysis and multilevel modelling
new univariate and MV models are available, including the 1-4 parameter logistic generalized partial credit, nominal, and multiple choice models. These are called by specifying a character vector called ‘itemtype’ of length nitems with the options ‘2PL’,‘3PL’,‘4PL’,‘graded’,‘gpcm’, ‘nominal’, or ‘mcm’; use ‘PC2PL’ and ‘PC3PL’ for partially-compensatory items. If itemtype = ‘1PL’ or ‘Rasch’, then the 1-parameter logistic/1-parameter ordinal or Rasch/partial credit models are estimated for all the data. The default assumes that items are either ‘2PL’ or ‘graded’, as before.
flexible user defined linear equality restrictions may be imposed on all estimation functions, so too can prior parameter distributions, start values, and choice of which parameters to estimate. These all follow these general 2 steps:
This is true for the parprior (MAP priors), constrain (linear equality constraints), and freepars (parameters freely estimated), each with their own little quirk. All inputs are lists with named parameters for easy identification and manipulation. Note that this means that the partial credit model and Rasch models may be calculated as well by modifying either the start values and constraints accordingly (e.g., constrain all slopes to be equal to 1/1.702 and not freely estimated for the classical Rasch model, or all equal but estimated for the 1PL model)
number of confmirt.model() options decreased due to the new way to specify item types, startvalues, prior parameter distributions, and constraints
plink package has not kept up with item information curves, so I’ll implement my own for now. Replaced plink item plots from ‘itemplots’ function with ones that I rolled
package descriptions and documentation updated
coef() now prints slightly different output, with the new option ‘allpars = TRUE’ to display all the item and group parameters, returned as a list
simdata() updated to support new item types
more accurate standard errors for MAP and ML factor scores, and specific factors in bfactorClass objects can now be estimated for all methods
dropped the ball and had lots of bug fixes this round. Future commits will avoid this problem by utilizing the testthat package to test code extensively before release
internal change in confmirt function to move MHRM engine outside the function for better maintenance
theta_angle added to mirt and polymirt plots for changing the viewing angle w.r.t theta_1
null model no longer calculated when missing data present
fixed item slope models estimated in mirt() with associated standard errors
null model computed, allowing for model statistics such as TLI
documentation changes
many back end technical details about estimation moved to technical lists
support for all GPArotation methods and options, including Target rotations
polymirt() uses confmirt() estimation engine
4PL support for mirt() and bfactor(), treating the upper bound as fixed
coef() now has a rotate option for returning rotated IRT parameters
Fixed translation bug in the C++ code from bfactor() causing illegal vector length throw
Fixed fscores() bug when using polychotomous items for mirt() and bfactor()
pass rotate=‘rotation’ from mirt and polymirt to override default ‘varimax’ rotation at estimation time (suggested by Niels Waller)
RMSEA, G^2, and p set to NaN instead of internal placeholder when there are missing data
df adjusted when missing data present
oblique rotations return invisible factor correlation matrix
degrees of freedom correctly adjusted when using noncompensatory items
confmirtClass reorganized to work with S4 methods, now work more consistently with methods.
fixed G^2 and log-likelihood in logLik() when product terms included
bugfix in drawThetas when noncompensatory items used
bugfixes for fscores, itemplot, and generic functions
read.mirt() added for creating a suitable plink object
mirt() and bfactor() can now accommodate polychotomous items using an ordinal IRT scheme
itemplot() now makes use of the handy plink package plots, giving a good deal of flexibility.
Generic plot()’s now use lattice plots extensively
Ported src code into Rcpp for future tweaking.
Added better fitted() function when missing data exist (noticed by Erin Horn)
ML estimation of factor scores for mirt and bfactor
RMSEA statistic added for all fitted models
Nonlinear polynomial estimation specification for confmirt models, now with more consistent returned labels
Provide better identification criteria for confmirt() (suggested by Hendrik Lohse)
parameter standard errors added for mirt() (1 factor only) and bfactor() models
bfactor() values that are ommited are recoded to NA in summary and coef for better viewing
‘technical’ added for confmirt function, allowing for various tweaks and varying beta prior weights
product relations added for confmirt.model(). Specified by enclosing in brackets and using an asterisk
documentation fixes with roxygenize