multistate - Stata development version history

This page details the version history of the development version of multistate, hosted on this website, which can be installed by typing in Stata,

net install multistate, from(https://www.mjcrowther.co.uk/code/multistate)

or updated by typing adoupdate multistate.

14mar2021: version 4.3.0

• accuracy and efficiency of numerical integration based predictions from an illness-death and extended illness death model are improved
• support for family(hazard) now documented
• bug fix for numerical integration based predictions of illness-death and extended illness-death, failing to pass the entry time appropriately would return a vector of 0s instead of appropriate probabilities; now fixed
• bug fix; if reset or tsreset() was used with an illness-death model, then numerical integration was used to calculate predictions instead of simulation, giving (obvious) incorrect results - this has been fixed
• bug fix when using a relative survival model within predictms, would attempt to call bhazard() when it should only simulate from the relative survival function; now fixed
• warning added when passing a relative survival model to predictms, to ensure the expected rate has been modelled.
  

09feb2021: version 4.2.1

• bug fix; check on transmatrix() skipped final row, so could erroneously exit telling the user the number of transitions and models didn’t match; now fixed

26jan2012: version 4.2.0

• interactive and freqat() options added to msboxes for use with interactive multi-state visualisations

16dec2020: version 4.1.0

• family(pwexponential) now available as a merlin/stmerlin model, which has been synced & documented with for use with predictms

28oct2020: version 4.0.0

• delta method is now the default for ci calculation for methods that use numerical integration (survival, competing risks or illness-death), or the hybrid AJ estimator for predictions
  • this is substantially faster than the bootstrap
  • includes contrasts
  • bootstrap is still the only method available for large-sample simulation based predictions
• bootstrap option added to override the default delta method
• latent option now documented (added in 3.0.0 to invoke latent times simulation, overriding the default total hazard + multinomial draws method)
• log normal, log logistic and general log cumulative hazard scale models added
• memory leak fixed, resulting in speed improvements

11sep2020: version 3.0.3

• msaj option enter() renamed to ltruncated() for consistency with survsim and predictms

08sep2020: version 3.0.2

• rmst missed from standardise calculation; now fixed
• ci with visit or rmst error’d; now fixed
• storage improved, so bootstrap cis slightly faster

01sep2020: version 3.0.1

• left over internal print statements when simulating; now removed

24aug2020: version 3.0.0

• predictms rewrite:
  • Stata version 15.1 now required
  • all models must be fitted using stmerlin or merlin
  • transprob renamed to probability
  • survival renamed to singleevent for transmat() shorthand
  • ltruncated() introduced, which defines the starting time for all predictions
  • rmst added to predict restricted mean survival time, i.e. total time spent in non-absorbing states
  • hazard and survival options added to predict transition specific functions
  • standard survival, competing risks, illness-death and extended illness-death predictions are now calculated using numerical integration as the default method, rather than large-sample simulation (can be over-written with simulate option)
  • the default method of simulation to obtain predictions now uses method of Beyersmann et al., with a new -latent- option to use the previous default of latent times
  • each simulated dataset can now be saved
  • ci now displays a bootstrap output so it can be monitored
  • ci’s now calculated on log/atanh/logit scale and back transformed to ensure appropriate bounds
• graphms added, with help file, to produce stacked transition probability plots
• help files substantially improved, new main multistate help file added

29mar2020: version 2.4.0

• msaj updated to version 1.0.0 by Micki Hill:
  • `msaj rewritten, all calculations now in Mata for speed
  • bug fix: confidence intervals caused a conformability error and kronecker delta function incorrectly defined; now fixed
  • removal of gen() option
  • id() option no longer required due to restructure
  • enter() time option added
  • exit() time option added
  • from() state option added
  • los option added to calculate restricted length of stay
  • se option added to return standard errors of probabilities
  • bi-directional/reversible transitions now supported
  • additional error checks
    

17mar2020: version 2.3.5

• bug fix; userfunction() with diff or ratio failed -> now fixed

03mar2020: version 2.3.4

• internal tidying

29nov2019: version 2.3.3

• bug fix: reset not working for strcs and stpm2 models due to indexing change; now fixed

13aug2019: version 2.3.2

• bug fix: missed an edit in 2.3.1 fix preventing aj to work in predictms (merlin function names); now fixed
• error check added for streg, d(ggamma) combined with aj - not supported

25jul2019: version 2.3.1

• bug fix: edits in merlin function names (1.2.5) caused predictms to fail; now fixed
• error check added for bhtime with strcs models - not currently supported

03jun2019: version 2.3.0

• bug fix: stpm2 and strcs models failed when simulating, introduced in 2.2.0; now fixed
• streg, dist(lognormal) now allowed with clock-forward models
• bug fix: streg, dist(lognormal) with reset incorrectly assumed loglogistic; now fixed
  

28may2019: version 2.2.0

• bug fix: with novcv() when ci not specified introduced in 2.1.0; now fixed
• bug in error check for log normal requiring reset; now fixed
• Generalised gamma streg model now supported when used in the models() syntax
• bug fix in msaj when number of transitions was not equal to dimension of transmatrix()

15apr2019: version 2.1.0

• novcv() option added to use mean vector instead of draws in CI calculations, i.e. assumes a particular transition is estimated without uncertainty

19mar2019: version 2.0.1

• help file improved for msaj

23feb2018: version 2.0.0

• bug fix: los was incorrectly scaled to 1, introduced in 1.2.0. Now fixed.
• normal approximation for CIs is now the default. percentile option replaces normal.
• bug fix: level() was ignored in CI calculations, always assumed 95%. Now fixed.
• gen() removed
• bug fix: in some cases mm_root_vec produced an error when solving the root. Now fixed.
• all predictions now available in one call
• infinite at#()s added (limited to 50 for error check reasons)
• at() changed to at1()
• difference option added
• added atref(#) - default 1
• standardise option added to calculate standardised (sample-averaged) predictions
• bug fix: with Stata 15 and streg (apart from dist(exp)), with at2(), the ancilary parameter constant was ignored. Now fixed.
• _time = 0 now allowed with normal ci’s
• userfunction() added for user-defined prediction function, subroutines for probs and los
• userlink() added for normal cis of userfunction() - default identity, can also be log or logit
• cr option added to avoid specifying a transition matrix. For use with models() only.
• bug fix: reset now synced with root-finding updates for stpm2 and strcs models
• outsample added
• stms added
• enter() removed, now min(timevar or _t)
• aj added for Aalen-Johansen estimator with Markov models
• reversible transitions now allowed

17nov2017: version 1.2.1

• bug fix with simulation from strcs models with delayed entry

13nov2017: version 1.2.0

• tscale2() and time2() added for transition-specific timescales
• probabilities scaled to 1

21nov2016: verison 1.1.1

• bug fixes

16nov2016: version 1.1.0

• some re-writes of source code for massive speed improvements when root-finding required
• trans#() syntax removed, now parsing is done automatically and using covariates expanded using msset
• model#() changed to models()
• improvements to help files
• covariates() added to msset to create transition-specific dummies, which must be used in model fitting (if required)
• default n() now 100,000, or 10,000 with ci
• survival added to calculate all predictions on a standard single event model

09aug2016: version 1.0.1

• gen() option to allow a stub for created variables, default is pred

12jul2016: version 1.0.0

• Initial release