Functions for fitting continuous-time Markov and hidden Markov multi-state models to longitudinal data. Designed for processes observed at arbitrary times in continuous time (panel data) but some other observation schemes are supported. Both Markov transition rates and the hidden Markov output process can be modelled in terms of covariates, which may be constant or piecewise-constant in time.
-- text --
(For detailed changes see https://github.com/chjackson/msm from Nov 2016 onwards, and the ChangeLog in the source package before that)
o New function updatepars.msm() to overwrite the estimates in a fitted model object to a given vector of values.
o Fix of bug in pearson.msm, triggered by r-devel.
o Fix of random memory crashes for models with censoring, revealed by asan testing.
o New feature viterbi.msm(..., normboot=TRUE) to return Viterbi results for a parameter estimate randomly sampled from the distribution of the MLEs.
o Bug fix to prevalence.msm with factor subject IDs.
o Bug fix to observed state prevalences in prevalence.msm for "ematrix"-style misclassification models with censoring - censored states were not being imputed correctly.
o Bug fix to plot.survfit.msm, which had been assuming that everyone starts at time zero.
o plot.survfit.msm gets a speed-up for bigger datasets, and "from" is now handled properly in the empirical curve.
o Bug fix to qpexp, and new "special" argument to qgeneric.
o CRAN release. Vignette source included in vignettes directory, on request of CRAN.
o r-forge release only. Fix of bug for qtnorm with vectorised arguments. Thanks to James Gibbons for the report.
o r-forge release only. Fix of bug for Pearson test with censored states. Thanks to Casimir Sofeu for the report.
o Fix of bug introduced in 1.5.2 for models with "obstrue" and "ematrix". This affected the first misclassification model presented in the PDF manual. Documented behaviour of "obstrue" clarified: with "ematrix" models, the state data are assumed to contain the true state if "obstrue" is turned on at the corresponding observation, and with "hmodel" models, the state data are generated from the HMM outcome model conditionally on the true state.
o Fix of minor bugs in draic.msm and output printing.
o CRAN release. Includes the changes from versions 1.5.1 - 1.5.3, plus also:
o Analytic derivatives for HMMs with multiple outcomes.
o Bug fix for printing model output when only one transition rate is affected by covariates. Thanks to Jordi Blanch for the report.
o More underflow correction for probabilities of hidden states in viterbi.msm. Thanks to Hannah Linder for the report.
o "death" argument in msm() is deprecated and renamed to "deathexact".
o censor.states now defaults to all transient states if not supplied, instead of complaining, even if there is no absorbing state. Thanks to Jonathan Williams for the report.
o HMMs can now have multiple observations at each time generated from different distributions. See new function hmmMV().
o obstrue can now contain the actual true state, instead of an indicator. This allows the information from HMM outcomes generated conditionally on this state to be included in the model.
o R-forge only release.
o HMMs can now have multiple observations at each time from the same distribution. The "state" in the "formula" argument of msm() is supplied as a matrix.
o Up-to-date version of the vignette included in the package.
o CRAN release. Includes the changes from versions 1.4.1 - 1.4.3.
o R-forge only release.
o Phase type models now allow an extra hidden Markov model on top.
o R-forge only release.
o viterbi.msm now returns the "posterior" probability of each hidden state at each time, given the full data.
o Bug fixes to misclassification models where some states were misclassified as other states with probability 1, for both ematrix and hmmCat specifications. Thanks to Li Su.
o R-forge only release.
o Experimental facility for two-phase semi-Markov models.
o Memory leaks in C code fixed. Thanks to Brian Ripley.
o Don't print CIs for fixed parameters.
o Documented that factors are allowed as the state variable as long as their levels are called "1", "2",...
o Bug fixes for covariates on initial state occupancy probabilities with structural zeros. Thanks to Jeffrey Eaton and Tara Mangal.
o Bug fixes for drlcv.msm. Thanks to Howard Thom.
o Give warning that polynomial contrasts aren't supported.
o Three and four-state versions of the BOS data provided.
o CRAN release. Includes the major changes from versions 1.3.2 and 1.3.3 below, previously only released on R-forge, plus:
o Default confidence interval method for pnext.msm changed to "normal", since delta method may not respect probability <1 constraint.
o R-forge only release.
o C interface changed from .C to .Call, giving a slight speed improvement.
o Probabilities of passage, see ppass.msm.
o R-forge only release.
o The new compact format for printing results from fitted models is now the default. The underlying numbers can be accessed from the functions msm.form.qoutput or msm.form.eoutput, or from the object returned by the print function, in the same tidy matrix form. The old print method is still available as "printold.msm".
o Analytic derivatives available for most hidden Markov models and models with censored states (excluding unknown initial state probabilities, constraints on misclassification / categorical outcome probabilities and their covariates, and truncated or measurement error distributions). This should speed up optimisation with the BFGS or CG methods. The corresponding Fisher information matrix is also available for misclassification (categorical/identity outcome) and censored state models.
o The BFGS optimisation method is now the default, rather than Nelder-Mead.
o The internal code that deals with reading the data and passing it to models has been rewritten to use formulae, model frames and model matrices more efficiently. As a result the "data" component of msm objects now has a different structure. The data can be extracted with the new model.frame() and model.matrix() methods for msm objects. Also see help(recreate.olddata) for a utility to get the old (undocumented) format back, but this will not be supported in the long term.
o New methods (draic.msm, drlcv.msm) for comparing models with differently-aggregated states. Thanks to Howard Thom.
o Parallel processing supported for bootstrapping and bootstrap confidence intervals (ci="boot"), if the "doParallel" package is installed.
o If msm is called with hessian=FALSE, then the Fisher (expected) information is used to obtain standard errors and CIs, though this is only available for non-hidden and misclassification models. This may be preferable if the observed Hessian is very intensive to approximate.
o Optimisation code tidied, making it easier to add new methods. As an example, the "bobyqa" algorithm, a fast derivative-free method, is now supported if the "minqa" package is installed.
o Give informative warning for initial outcomes in HMMs which are impossible for given initial state probabilites and outcome models.
o Internal centering of "timeperiod" covariates around their means for inhomogeneous models specified with "pci" is now done consistently with other covariates, by omitting subjects' last observations before calculating the mean, since they don't contribute to the likelihood. Therefore for these models, the initial values (with "covariates centered around their means in the data") and outputs for covariates="mean", have a very slightly different meaning from previous versions.
o When calculating the likelihood for hidden Markov or censoring models, P matrices are not recalculated when the same one occurs more than once. This may speed up some models.
o Test suite tidied up and converted to use "testthat" package.
o Data consistency check added to crudeinits.msm().
o Bug fix for misclassification models with constraints on baseline misclassification probabilities and fixed parameters.
o Bug fix for bootstrap CIs with efpt.msm
o Miscellaneous minor bug fixes, see Changelog.
o R-forge only release.
o Time-dependent covariates supported in totlos.msm.
o New function envisits.msm() for expected number of visits to each state over a period, calculated as a corollary of totlos.msm().
o New utility "msm2Surv" to export data from msm format to counting process format for use with the survival and mstate packages. This assumes the exact transition times of the process are known.
o More informative messages from model fits which have not converged. In particular, a warning is now given when the optimiser iteration limit was reached without convergence, which previously happened silently.
o Miscellaneous minor bug fixes, see Changelog.
o CRAN release. Includes the changes below from R-forge versions 1.2.1 up to 1.2.7, plus:
o Fix of bug introduced in 1.2.3 which broke models with non-standard state ordering.
o Datasets now lazy loaded so data() not required.
o New "start" argument to efpt.msm, allowing averaging over a set of starting states.
o R-forge only release
o Fix of bug for logLik.msm with by.subject=TRUE.
o R-forge only release
o Row numbers reported in error message about different states at the same time corrected to account for missing data. Thanks to Lucy Leigh for the report.
o An informative error is now shown if trying to use gen.inits with a hidden Markov model, and it is now documented that this is not supported.
o R-forge only release
o Analytic formula for totlos.msm implemented, which is vastly more efficient than the numerical integration used previously. Debugging outputs left in 1.2.3 also removed.
o Matrix exponentials, in MatrixExp and non-analytic likelihood calculations, are now calculated using expm from the expm package by default. As a result msm now depends on the expm package.
o R-forge only release
o Range constraints can now be given for HMM outcome parameters, through a new argument "hranges" to msm. This may improve HMM identifiability.
o R-forge only release
o New interface for easily specifiying different covariates for each transition intensity, through a named list in the "covariates" argument to msm. Previously this required "fixedpars".
o Major restructuring of the internal code, mainly so that parameters are adjusted for covariates in R rather than C. There should be no differences visible to the user.
o Initial state occupancy probabilities are estimated on the multivariate logit scale, not univariate, and confidence intervals are calculated using a simulation-based method (with 10000 simulations, so there will be a small Monte Carlo error).
o When centering covariates around their means for the default likelihood calculation, the means used are now after dropping missing values and subjects with one observation, not before. Thanks to Howard Thom for the report.
o Relatedly, the covariate values for subjects' last observations are not included in this mean, since they don't contribute to the likelihood, so interpretation of initial values for the qmatrix, and outputs for covariates="mean", will now be very slightly different.
o Bug fix in totlos.msm: calculations were wrong for fromt > 0.
o Memory bug in Viterbi, which could crash R, fixed.
o R-forge only release
o Can now examine subject-specific -2 log likelihoods at the maximum likelihood estimates, via logLik.msm().
o The state can now be a factor with levels (1:nstates), as well as numeric. Previously supplying a factor state led to unpredictable behaviour and potential crashes.
o R-forge only release
o A matrix of fixed patient-specific initial state distributions can now be supplied as "initprobs" in hidden Markov models.
o Implemented accurate p-value for the Pearson-type test from Titman (Lifetime Data Analysis, 2009). Non-hidden Markov models for pure panel data only.
o A Fisher scoring algorithm can now be used to maximise the likelihood for panel data without censored / hidden states. Thanks to Andrew Titman for help with this.
o New function efpt.msm for expected first passage times for time-homogeneous models.
o prevalence.msm now produces expected values by integrating model predictions over the covariate histories observed in the data, if 'covariates="population"' is supplied. This is the default, but the old behaviour is available by supplying fixed covariates in the "covariates" argument.
o In prevalence.msm and plot.prevalence.msm, subjects reaching the absorbing state can be removed from the risk set after they have reached an optional censoring time. Thanks to Andrew Titman.
o Newly user-accessible function simfitted.msm for simulating from a model defined by the estimates from a model fitted in msm.
o Subjects with only one observation are dropped from the data stored in fitted model objects. This gives more accurate numbers at risk in prevalence.msm.
o Arguments can be passed through summary.msm to prevalence.msm.
o pmatrix.piecewise.msm allows time-homogeneous models with change point vector "times" of length 0.
o Fixes for bugs in the the Pearson test introduced in 0.9.5.
o Misclassification models where some off-diagonal misclassification probabilities are 1 are now handled properly. Thanks to Howard Thom for uncovering this.
o Bug fix for interp="midpoint" method in calculation of observed prevalences (prevalence.msm). Thanks to Erica Liu.
o Bug fix for Viterbi algorithm with obstrue. Thanks to Linda Sharples.
o Minor modification of package tests to enable R CMD check to pass with the forthcoming release of mvtnorm.
o Bug fix: qmatrix.msm and ematrix.msm were returning inaccurate delta method standard errors / CIs with center=FALSE, covariates and user-supplied covariate values. Thanks to Vikki O'Neill for the report.
o Use BFGS method for one-parameter optimisation unless method supplied explicitly, avoiding warning about unreliability of Nelder-Mead.
o New Student t distribution for hidden Markov model outcomes. Thanks to Darren Gillis.
o Removed debugging browser which had been inadvertently left in pearson.msm. Thanks to Chyi-Hung Hsu.
o Corrected equation 5 in the PDF manual for the likelihood under exact transition times. The code was unaffected. Thanks to Simon Bond.
o Fix of bug in calculation of confidence intervals using "ci=normal". Affected models were those with fixed parameters or HMMs. Users are advised to check their results with the corrected package - apologies.
o If user supplies an ematrix with all misclassification probabilities zero, this degrades gracefully to a non-misclassification model. Thanks to Sharareh Taghipour for the report.
o Bug fix for error messages when model inconsistent with data, and when subject IDs not adjacent. Thanks to Kelly Williams-Sieg for the report.
o Bug fix in pearson.msm for models where transitions are only allowed from one state. Thanks to Gavin Chan for the report.
o qtnorm fixed for p=0 or 1 and upper < lower. Thanks to Art Owen for the report.
o New function "pnext.msm" to compute a matrix of probabilities for the next state of the process.
o New "[" method to intuitively extract a row and column of matrix-based estimates and confidence intervals, for example qmatrix.msm(x)[1,2]
o Miscellaneous doc and minor bug fixes, see Changelog.
o Fix of a bug which made pmatrix.msm break for time-inhomogeneous models with non-integer time cut points "pci". Thanks to Christos Argyropoulos for the report.
o Return -Inf in dtnorm when outside truncation bounds and log=TRUE.
o 1.0 release to accompany the forthcoming Journal of Statistical Software paper about msm.
o Line types, colours and widths can be configured in plotprog.msm, plot.survfit.msm and plot.prevalence.msm.
o Added warning for multiple observations at the same time on the same person with different states, which leads to zero likelihood and the dreaded "cannot be evaluated at initial values" message.
o If center=FALSE, the $Qmatrices$baseline, $Ematrices$baseline and $sojourn components of msm objects are evaluated with covariate values of 0, for consistency with "logbaseline". Documentation and printed output corrected accordingly. These issues caused problems with viterbi.msm. Thanks to Kenneth Gundersen for the report.
o Bug fixes for bootstrapping with totlos, covariates on HMM outcomes and fixedpars. Thanks to Li Su for the report.
o Fix of a bug which caused occasional wrong likelihood calculations for models with "exacttimes". Thanks to Brian Tom for the report.
o Fix for "NA in probability vector" error in pearson.msm. Thanks to Wen-Wen Yang for the report.
o Fix for a bug in pearson.msm triggered by a change in R version 2.10.0, which caused all expected values to be returned as zero. Thanks to Brian Tom for the report.
o Bug fix for calculation error in scoreresid.msm. Thanks to Aidan O'Keeffe for the report.
o Options to MatrixExp for calculating the matrix exponential can be passed through from pmatrix.msm and pmatrix.piecewise.msm. Thanks to Peter Adamson for the suggestion.
o Missing data handling bug fixes, in particular, crudeinits.msm and gen.inits no longer give errors if there are missing values in the subject, time or state variable.
o Other minor bug fixes, see ChangeLog.
o Bug fix - estimates of covariate effects in matrices outputted by msm were ordered wrongly in models with "qconstraint". Thanks to Brian Tom for the report.
o Bug fix - "gradient in optim evaluated to wrong length" was still affecting certain models with fixed parameters. Thanks to Aidan O'Keeffe for the report.
o Fix to pearson.msm for R versions >= 2.9.1 ("replacement has 0 rows" error)
o Bug fix for models with fixed parameters fitted using optimisation methods with derivatives ("BFGS"), which failed with the error "gradient in optim evaluated to wrong length". Thanks to Isaac Dinner for the report.
o Minor update to the test suite to allow build on Fedora / Red Hat Linux.
o Time-inhomogeneous models fitted with the "pci" argument to msm are now fully supported in all output functions.
pmatrix.msm can now compute transition probabilities over any given time interval for time-inhomogeneous models fitted with "pci". A new argument "t1" to pmatrix.msm specifies the starting time, while "t" still specifies the interval length.
All functions which call on pmatrix.msm, such as plot.msm, plot.survfit.msm, prevalence.msm and totlos.msm, now account for time-inhomogeneity in models fitted using "pci".
o Extractor functions are now more tolerant. If a list of covariate values is supplied, unknown covariates are ignored and covariates with unspecified values are set to zero. Factor values can be specified either by factor levels or by 0/1 contrasts.
o Bug fix - score residuals were being calculated wrongly for models with covariates.
o Derivatives are now used in the optimisation by default (use.deriv=TRUE) for optimisation methods such as BFGS which employ them.
o Licence clarified as GPL-2 or later, to enable packaging of msm for Fedora/Red Hat Linux.
o Bug fix - extractor functions were not being calculated for models with interactions between covariates.
o Sources for the PDF manual included in the source package, to enable inclusion of msm in Debian GNU/Linux.
o New option "pci" to msm, which automatically constructs a model with piecewise-constant transition intensities which change at the supplied times.
o The HMM outcome model is assumed to apply to censored states in HMMs, unless obstrue = 1.
o totlos.msm now calculates total length of stay for all states, not just transient states. New argument "end" added.
o Bug fix in the likelihood calculation for data containing a mixture of obstype = 1 and obstype = 2. Thanks to Peter Jepsen for uncovering this.
o New function "pearson.msm" implementing the Pearson-type goodness-of-fit test for multi-state models fitted to panel data (Aguirre-Hernandez and Farewell, Statistics in Medicine 2002; Titman and Sharples, Statistics in Medicine 2007). Thanks to Andrew Titman for his work on this.
o New function "scoreresid.msm" to compute and plot score residuals for detecting influential subjects.
o New function "plotprog.msm" to plot Kaplan-Meier estimates of time to first occurrence of each state.
o New function "plot.survfit.msm" to plot Kaplan-Meier estimate of survival probabilty compared with the fitted survival probability from a model.
o New convenience function "lrtest.msm" for comparing a set of models with likelihood ratio tests.
o logLik method returns the log-likelihood, not the minus log-likelihood, for consistency with methods in other packages. Thanks to Jay Rotella.
o msm now depends on the "survival" package.
o Data "heart" renamed to "cav" to avoid clashing with the dataset in the "survival" package.
o Covariates on misclassification probabilities can now be specified in simmulti.msm. Simulation bug introduced in 0.7.5 fixed.
o quantile functions (qtnorm,qmenorm,qmeunif,qpexp) made more robust for small probabilities.
o The Viterbi algorithm can now be used to impute the most likely true state for censored states, as well as for HMMs
o prevalence.msm now handles models with censored states correctly, using the Viterbi algorithm to determine the observed states.
o Bug fix: account for extra arguments supplied to "prevalence.msm" when producing the plot of prevalences against time. Thanks to Peter Jepsen for the report.
o Bug fixes involving factor covariates in bootstrapping and qratio.msm. Thanks to Peter Jepsen.
o New beta outcome distribution for hidden Markov models.
o Minor changes to satisfy the package-building tools in the new R version 2.6.0.
o Confidence intervals in various output functions can now be calculated by simulating from the asymptotic normal distribution of the maximum likelihood estimates of the Q matrix and transforming. The "ci.boot" argument in these functions has been replaced by the "ci" argument, which can take values "none", "normal" and "bootstrap". This is implemented for qmatrix.msm, ematrix.msm, sojourn.msm, qratio.msm, pmatrix.msm, pmatrix.piecewise.msm, totlos.msm and prevalence.msm. Such CIs are expected to be more accurate than the delta method, but less accurate than bootstrapping. There is a similar compromise in computation time. Thanks to Andrew Titman for the suggestion.
o As a result, msm now depends on the mvtnorm package.
o In prevalence.msm, observed and expected prevalences can now be plotted against time. Thanks to Andrew Titman for the suggestion.
o In prevalence.msm, observed states can be interpolated using the assumption that they change at the midpoints between observation times.
o Matrix exponential routines now handle matrices with complex eigenvalues. Thanks to V�ronique Bouchard for uncovering the bug.
o Bug fix to surface.msm for HMMs. Thanks to Michael Sweeting.
o Bug fix for bootstrapping - now handles models with obstype and obstrue. Thanks to Peter Jepsen for the report.
o An error in the calculation of multinomial logistic regression probabilities has been fixed. This will change the results of misclassification models where there were both a) three or more possible classifications for a particular underlying state and b) covariates on the corresponding classification probabilities. Any changes are not expected to be substantial.
o Misclassification probabilities are now estimated on a different scale during the optimisation: log relative to baseline probability, instead of on a univariate logit scale. Therefore maximum likelihood estimates for misclassification models may be very slightly different from previous versions.
o Confidence intervals for probabilities are now more appropriately calculated using a delta method approximation to the variance of logit(p), instead of log(p).
o New argument "initcovariates" and "initcovinits" to msm, to allow covariate effects on initial state probabilities in hidden Markov models to be estimated through multinomial logistic regression.
o Initial state probabilities initialised to zero are now fixed at zero during optimisation, if initprobs is being estimated ("structural zeroes").
o New argument "obstrue" to msm, to allow some observations to be observed without error in misclassification models.
o Constraints on covariate effects on transition intensities are now allowed such that some effects are equal to other effects multiplied by -1.
o New option "ci.boot" to prevalence.msm. This is a helper to calculate bootstrap confidence limits for the expected prevalences using "boot.msm".
o rtnorm() for sampling from the truncated normal distribution now uses the efficient rejection sampling methods by Christian Robert.
o Maintainer's email address is now [email protected]
o msm now gives a warning when the standard errors cannot be calculated due to the Hessian at the converged "solution" being non-positive-definite. This issue had been causing a lot of user confusion.
o prevalence.msm can now calculate expected prevalences for models with piecewise-constant intensities, in the same manner as pmatrix.piecewise.msm. Intensities must still be common to all individuals.
o Bug fix for presentation of intensity matrices in print.msm and qmatrix.msm when center = FALSE. These had been returning matrices with covariates set to zero, when they should have been set to their means. Thanks to Ross Boylan.
o Covariates on transition process which are missing at an individual's last observation are not dropped, because they are not used in the analysis. Thanks to Jonathan Williams. This has the consequence that output from prevalence.msm may be different from earlier versions (0.7 or earlier) if there are missing values in the data. Users are advised to deal with missing values in their data appropriately before using msm.
o Miscellaneous other bug fixes, see ChangeLog.
o Initial state occupancy probabilities in hidden Markov models can now be estimated. See new argument "est.initprobs" to msm.
o Bootstrap resampling is implemented. This may be used to calculate confidence intervals or standard errors for quantities such as the transition probability matrix where this was previously not possible with msm, or as an alternative to Hessian-based standard errors or the delta method for other quantities. See new function boot.msm.
o Bootstrap confidence intervals can be calculated directly from pmatrix.msm and totlos.msm.
o Bug fix in estimation of observed prevalences at maximum observed time. Thanks to Jeremy Penn for the report. The function has also been rewritten so that the calculation of these prevalences is now much faster.
o prevalence.msm is adapted sensibly to handle data where not all individuals start at a common time.
o The values of categorical (factor) covariates in output functions, such as qmatrix.msm, are now specified in an intuitive way. For example, to calculate a statistic with the categorical covariate "smoke" at the level "CURRENT", just supply list(smoke="CURRENT") as the "covariates" argument to the output function.
o Bug fix to rtnorm for vector parameters. Thanks to Jean-Baptiste Denis for the report.
o Bug fix to sim.msm: multiply covariates by baseline intensities in the correct order. Thanks to Stephan Lenz for the report.
o Correction to version 0.6.2 with the references reinstated in the manual.
o The likelihood for certain transient 2, 3, 4 and 5 state models is now calculated using analytic expressions for the transition probability matrix, instead of by numerically calculating the matrix exponential. This can give big speed improvements.
o Various bug fixes, including support for character subject IDs.
o Bug fix release. In Viterbi algorithm, don't ignore initial state occupancy probabilities. Thanks to Melanie Wall for reporting this. For other bug fixes see the ChangeLog.
o New argument "use.deriv" to msm. If TRUE, then analytic derivatives are used in the algorithm to maximise the likelihood, where an appropriate algorithm is being used, such as optim's BFGS. These derivatives are also used to calculate the Hessian at the maximum. Not supported for hidden Markov models or models with censoring. This may substantially speed up convergence, especially for larger models.
o The Newton-type algorithm (Dennis and Schnabel) from the R function "nlm" can also be used to maximise the likelihood, as an alternative to the algorithms in "optim".
o New function "surface.msm" to plot likelihood surfaces, for example, in the region of a suspected maximum. Includes methods for the generic R functions contour(), persp() and image(), to produce each respective type of surface plot for a "msm" object.
o Bug fix in Viterbi algorithm. It didn't handle underlying Markov models with progressive and regressive states properly. Thanks to Rochelle Watkins.
o Negative binomial hidden Markov output distribution added.
o Miscellaneous other bug fixes, see ChangeLog.
o Bug fix in simulation functions (sim.msm, simmulti.msm). Models with time dependent covariates were not being simulated properly, the covariate changes were not fully accounted for. Thanks to Mike Sweeting for the report.
o New functions dpexp, ppexp, qpexp, rpexp for the exponential distribution with piecewise-constant rates.
o Bug fix. covariates with the same names as internal msm variable names, such as "subject", "time" and "state", are now allowed.
o Argument "hessian" added to msm, to avoid calculating standard errors, for example when bootstrapping.
o Miscellaneous internal edits and fixes, see ChangeLog.
o Major update. Much of the internal R and C code has been re-written.
o General continuous-time hidden Markov models can now be fitted with msm, as well as misclassification models. Allowed response distributions conditionally on the hidden state include categorical, normal, Poisson, exponential and others. See the new "hmodel" argument. Misclassification models can either be fitted in the old style using an ematrix, or using a general HMM with a categorical response distribution. Covariates can be fitted to many of the new hidden response processes via generalized regressions. See "hcovariates", "hcovinits" arguments.
o Per-observation observation schemes, generalising the "exacttimes" and "death" concepts. An optional new variable in the data can specify whether each observation is a snapshot of the process, an exactly-observed transition time, or a death state. Observations are allowed to be at identical times, for example, a snapshot followed instantly by an exact transition time.
o Various syntax changes for cleaner moder specification.
Instead of 0/1 indicators, qmatrix and ematrix should contain the initial values for the transition intensity / misclassification matrix. These matrices can be named with names for the states of the Markov chain.
The inits argument is abolished. Initial values are estimated automatically if the new argument to msm "gen.inits = TRUE" is supplied. This uses the initial values calculated by crudeinits.msm.
misc no longer needs to be specified if an ematrix is supplied.
fixedpars=TRUE fixes all parameters, or specific parameters can be fixed as before.
crudeinits.msm takes a state ~ time formula instead of two separate state, time arguments, for consistency with the msm function.
Initial values for covariate effects on transition rates / misclassification probabilities are assumed to be zero unless otherwise specified by the new "covinits" / "misccovinits" argument.
o Support for 'from-to' style data has been withdrawn. Storing data in this format is inadvisable as it destroys the longitudinal nature of the data.
o Speed improvements. The algorithm for calculating the likelihood for non-hidden multi-state models has changed so that the matrix exponential of the Q matrix is only calculated once for each time difference / covariate combination. Therefore, users should see speed improvements for data where the same from-state, to-state, time difference, covariates combination appears many times.
o Confidence intervals are now presented instead of standard errors for uncertainty in parameter estimates.
o New method of calculating matrix exponentials when the eigenvector matrix is not invertible. It now uses the more robust method of Pade approximants with scaling and squaring, instead of power series. Faster LAPACK routines are now used for matrix inversion.
o covmatch argument to msm has been abolished. To take a time-dependent covariate value from the end of the relevant transition instead of the default start, users are expected to manipulate their data accordingly before calling msm, shifting the positions of the covariate back by one within each subject.
o Syntax changes for simmulti.msm.
o The likelihood is now calculated correctly for individuals with censored intermediate states, as well as censored initial and final states. Thanks to Michael Sweeting for reporting this.
o hazard.scale and odds.scale were interpreted wrongly in hazard.msm and odds.msm respectively.
o time-dependent covariate values now taken from the start instead of end of the transition under hidden Markov models.
o Censored outcomes in misclassification models are assumed to be not subject to misclassification.
o A couple of bug fixes for exact transition times.
o Censored observations are now supported, via new "censor" and "censor.states" arguments. A censored observation is unknown, but known to be one of a particular set of states.
A major update to msm is under development, for release in the first half of 2005. This will support hidden Markov models with general response distributions.
o Maintenance release with minor fixes and enhancements ready for R-2.0.0.
o More than one death state is now permitted, through the "death" argument. Death states are those whose exact entry time is known, but the state at the previous instant before death is unknown.
o The "tunit" argument has been abolished. Death times are now assumed to be exact rather than known within one day. This makes more sense since for longitudinal studies, all observations are usually recorded to within one basic time unit, not just death times.
o Cleanups of the manual and minor fixes, as detailed in ChangeLog.
o Bug fix. The likelihood was being wrongly calculated in cases when both the data represent exact transition times and the transition intensity matrix had repeated eigenvalues.
o The "death" argument is no longer ignored when exacttimes=TRUE, as it is reasonable to have the entry time into one state accurate to within one day, and all other times exactly accurate.
o More memory problems should be fixed.
o Two errors in the calculation of the likelihood for a multi-state model have been corrected. These bugs affect only models with reversible transition matrices, that is, models which allow progression and regression between states.
o The first bug occurred when death times were known to within one time unit (death = TRUE) - the likelihood calculation did not account for reversible states.
o The second bug occurred when the data represent exact transition times (exacttimes = TRUE). The likelihood calculation did not properly account for reversible states.
o Baseline transition intensities, or misclassification probabilities, can now be constrained to be equal to each other, in the same manner as covariate effects. Specified by new arguments "qconstraint" or "econstraint".
o The memory allocation problems of version 0.2 have been fixed.
o Fixed some minor bugs, as detailed in ChangeLog.
o New function, pmatrix.piecewise.msm, for calculating transition probability matrices for processes with piecewise-constant intensities.
o Fixed a handful of minor bugs, as detailed in ChangeLog.
o Minor edits and additions to the manual.
o The subject ID can now be factor or character.
o A full manual in PDF format is included in the doc directory. This gives the mathematical background behind multi-state modelling, and a tutorial in the typical use of the functions in the msm package.
o Many more methods for extracting summary statistics from the fitted model are included. These are generally called with the fitted model as the argument, plus an optional argument indicating the assumed covariate values. The functions include qmatrix.msm, ematrix.msm, pmatrix.msm, qratio.msm, sojourn.msm, totlos.msm, hazard.msm, odds.msm, prevalence.msm.
o New function statetable.msm to calculate frequencies of transitions between pairs of states observed in the data.
o New function crudeinits.msm to estimate transition intensities assuming the data represent the exact transition times of the Markov process. These can be used as initial values in the msm function for fitting the model.
o prevalencemisc.msm has been removed, as its methodology was overcomplicated and confusing. The methods used in prevalence.msm have been extended naturally to deal with misclassification models.
o Fix of a bug in the likelihood calculation for misclassification models (the number of non-death states was assumed to be the same as the number of states that could be misclassified, leading to failure to calculate the likelihood for models where some states are observed without error, but are not death states. ) Thanks to Martyn Plummer for reporting this.
o Fix of a bug in the simulation routines (getobs.msm, called by simmulti.msm), where for models with absorbing states, the absorbing state is not retained in the simulated data.
o New heart transplant example data set, as used in the manual, so that all the examples given in the manual can be run by the user.
o Tidying of the help pages.
o First release.