MEDseq: Mixtures of Exponential-Distance Models with Covariates

MEDseq v1.4.2 - (13^th release [patch update]: 2025-03-10)

Improvements, Bug Fixes & Miscellaneous Edits

• Many adjustments to plot.MEDseq:
  • MEDseq_clustnames gains the MAP=FALSE arg., for use when size=TRUE: this is now used by
    plot.MEDseq (where soft=FALSE corresponds to MAP=TRUE) when both SPS & size are TRUE.
  • TraMineR type y-axis labels now properly account for all combinations of soft & weighted,
    particularly when subset is invoked &/or type="ms", with additional minor fixes to subset arg.
  • type="ms" can now show noise component’s modal sequence by setting subset appropriately,
    but not by default as the model does not estimate modal sequences for noise components.
  • The MEDseq_clustnames arg. cluster can now also be passed via ... when SPS=TRUE.
  • Allowed new TraMineR arg. col.entr to be passed via ... when type="Ht" or type="dH".
• Minor fixes in relation to MEDseq_control arg. tau0:
  • tau0 can now always be supplied as a vector (previously allowed only with noise.gate=TRUE).
  • Related bug fix when tau0 is supplied as a vector with noise.gate=TRUE.
  • Related initialisation bug fix when tau0 != 0.5 (the implied default) for G=2 models with noise.
• MEDseq_control gains new init.z option "soft.random":
  • The "random" option has been renamed to "random.hard", but
    init.z="random" will work as before due to partial matching.
  • Related bugs when a ‘soft’ z.list is used with algo != "EM" are also fixed.
• MEDseq_fit now checks for and terminates in the presence of both types of missingness
  as per new TraMineR function seqhasmiss, i.e. now also accounts for void values.
• plot.MEDcriterion added as a wrapper to plot.MEDseq with related type, for convenience:
  for example, plot(x$BIC) is now equivalent to plot(x, type="bic").
• Experimental MEDseq_control arg. dist.mat no longer governs ASW calculations: it still defaults
  to a Hamming distance matrix but now allows only for initialisation to be based on other distances.
• Detecting modal sequence ties is now handled properly for alphabets of size greater than 9.
• Minor E-step speed-ups for "CU", "CUN", "UU", & "UUN" models for clusters with only one observation.
• Additional minor speed-ups for unweighted models with G=1.
• print.MEDseq now works again for all models with G=1.
• Further minor speed-ups to various utility functions using vapply in place of tapply.
• Ensured TraMineR (>= 2.2-10) in DESCRIPTION Imports: field due to col.entr & seqhasmiss.
• Many additional minor documentation improvements.

MEDseq v1.4.1 - (12^th release [patch update]: 2023-12-12)

Improvements, Bug Fixes & Miscellaneous Edits

• Minor fixes to plot.MEDseq:
  • sortv options "from.start" and "from.end" borrowed from TraMineR when
    seriated is "observations" or "both" for the "clusters", "i", & "I" type plots.
  • Seriation is now explicitly prohibited when type="gating" when:
    • models contain no gating covariates: previously an error was returned in such cases.
    • x.axis is supplied via the ... construct.
  • Removed spurious warnings in TraMineR type plots when using extra args. via ....
  • Remedied labelling issues introduced in previous update when seriated="none" is supplied.
• MoE_entropy and MoE_AvePP both gain the arg. group for computing the average entropies
  and posterior probabilities of each component, respectively: defaults to FALSE, i.e. old behaviour.
• Now following proper convention for reexporting TraMineR::seqdef.
• Extensive edits to avoid overheads introduced in matrixStats (>= 1.0.0) + related minor speed-ups.
• Now using newer CITATION commands & updated License: GPL (>= 3).

MEDseq v1.4.0 - (11^th release [minor update]: 2022-12-20)

New Features & Improvements

• Function seqdef added as an exact copy of TraMineR::seqdef, to enable experienced
  users of MEDseq & TraMineR to use the former without needing to explicitly load the latter.
• MEDseq_clustnames gains the arg. weighted=FALSE for use when size=TRUE:
  this is now respected by the weighted arg. to plot.MEDseq where relevant.
• New function dist_freqwH added for calculating pairwise dissimilarity matrix associated with
  wKModes(..., freq.weighted=TRUE) for subsequent use (e.g. silhouettes).
• The plot.MEDseq function’s type arg. gains the option "dH",
  provided version 2.2-4 or later of the TraMineR package is installed.
• plot.MEDseq also gains the "similarity" option for its type argument.
• New function MEDseq_AvePP added.

Bug Fixes & Miscellaneous Edits

• wKModes now also returns x$tot.withindiff (i.e. sum(x$withindiff)).
• Minor speed-ups to wKModes when freq.weighted=TRUE.
• Minor cosmetic changes to type="dbsvals" & type="aswvals" in plot.MEDseq.
• Minor speed-ups to plot.MEDseq related to its seriated arg. in G=1 settings.
• Fixed rare bugs in tie-breaking for modal sequence estimate in MEDseq_fit & wKModes.
• Fixed documentation typos.

MEDseq v1.3.3 - (10^th release [patch update]: 2022-03-28)

Improvements, Bug Fixes & Miscellaneous Edits

• Major speed-ups to E-steps for all model types when G>1.
• Minor speed-ups to distance calculations for all model types when G>1.
• MEDseq_meantime gains the map.size arg. and a related print method.
• Added summary (and related print) methods for MEDCriterion objects.
• New function MEDseq_entropy added.
• Fixed mismatched plotting symbols for models with noise in model-selection criteria plot legends.
• Minor fix to handle (rare) empty components.
• Minor edits for compatibility w/ latest TraMineR release w.r.t. "mt" & "ms" plots.

MEDseq v1.3.2 - (9^th release [patch update]: 2021-12-19)

Bug Fixes & Miscellaneous Edits

• Modifications to WKModes (& thus related MEDseq_control init.z
  options "kmodes"/"kmodes2"), by further altering klaR::kmodes:
  • Ties for modes now broken randomly, using new wKModes arg. random (defaults to TRUE).
  • All tie-breaks for cluster assignments now biased towards previous iteration’s assignments.
  • Fixed rare bug when modes is supplied as a number with aggregated data, e.g. "kmodes2".
• MEDseq_fit & other functions now work for sequence alphabets of any size;
  previously, only sequences with fewer than 10 states/categories were accommodated.
• Minor fix to dbs function when supplying clusters with a noise component.
• sapply replaced with vapply, with other negligible speed-ups.
• Updated citation info after final publication in JRSSA.

MEDseq v1.3.1 - (8^th release [patch update]: 2021-10-14)

Bug Fixes & Miscellaneous Edits

• Fixes for init.z options "kmodes" & "kmodes2" in MEDseq_control, with new function wKModes
  provided for running the k-modes algorithm on weighted data: previously, k-modes initialisation
  was only available for unweighted sequences via the now-replaced klaR::kmodes function
  (consequently, the klaR package has been removed from the DESCRIPTION Suggests: field).
• plot.MEDseq gains the arg. subset, for use with the TraMineR type plots:
  allows plotting some but not all components, e.g. only the noise component (see documentation).
• Fixed minor bug causing MEDseq_fit to crash when weights are supplied and unique=FALSE.
• Fixed ASW calculation when unweighted sequences are aggregated (i.e. unique=TRUE, the default).
• Fixed small bug for type="ms" plots for models with a noise component when SPS=TRUE.
• Fixed printing of noise.gate in MEDseq_compare for G=2 models w/ noise & gating covariates.
• Improved checks on G in MEDseq_fit.

MEDseq v1.3.0 - (7^th release [minor update]: 2021-07-15)

New Features & Improvements

• plot.MEDseq gains a number of new arguments:
  • soft allows soft cluster membership probabilities to be used for the "d", "f", "Ht", "ms",
    & "mt" type plots (default: soft=TRUE) + the "i" & "I" plots (default: soft=FALSE), in a
    manner akin to WeightedCluster::fuzzyseqplot(): previously, all but the "ms" plot used the
    hard MAP partition and discarded the soft assignment information (i.e. soft=FALSE, implicitly).
  • sortv allows overriding the smeth arg. to instead order observations in certain plots
    (where seriated is one of "observations" or "both") by the "dbs" or "asw" values;
    additionally, and for consistency with WeightedCluster::fuzzyseqplot(),
    sortv="membership" is provided for soft=TRUE type="I" plots.
  • weighted (TRUE, by default) allows control over whether the weights (if any) are used;
    relevant only for "d", "f", "Ht", "i", "I", "ms", & "mt" type plots.
• Exported MEDseq_clustnames & MEDseq_nameclusts functions and added SPS arg. to plot.MEDseq,
  MEDseq_meantime, MEDseq_stderr, & various/more print/summary methods: now certain plots &
  outputs can be (or are by default) labelled with the central sequences in SPS format, as per the paper.
• seriated options "observations" & "both" can now be used for "i" type plots,
  with related minor fixes for "i" & "I" type plots for weighted data with seriated observations.
• Added predict, fitted, & residuals methods for "MEDgating" objects, i.e. x$gating.
• MEDseq_meantime gains the arg. wt.size (defaults to FALSE).
• Minor speed-ups to model-fitting for modtype="CU".

Bug Fixes & Miscellaneous Edits

• A warning message is now printed if the gating network’s MLR ever fails to converge, prompting users to
  modify the itmax arg. to MEDseq_control: the 2^nd element of this arg. governs the maximum number of
  MLR iterations — consequently, its default has been modified from 100 to 1000, which is liable to slow
  down internal calls to nnet::multinom, but generally reduces the required number of EM iterations.
• Changes to default colour palettes & plotting symbols for certain plot types:
  Suggests: package viridisLite now only invoked if available.
• Minor fixes to returned x$gating object, especially for equalPro models
  with a noise component and weighted models without any gating covariates at all.
• Stronger checks to ensure weights arg. is explicitly supplied to MEDseq_fit
  in cases where the "stslist" object passed via seqs has the "weights" attribute.
• Added error message to MEDseq_fit when the number of states exceeds 9,
  to better inform of this bug which will be rectified in future updates.
• Fixed bug preventing inclusion of higher-order terms in gating formulas when there are duplicates.
• Minor fixes to get_MEDseq_results and how its optional args. are internally handled by plot.MEDseq.
• Stronger checks for variables in gating formula which are not found in covars.
• type="mean" option renamed to type="central" in plot.MEDseq.
• type="ms" plots now work properly when seriated="clusters" or seriated="both".
• Removed some superfluous warnings for all but the "mt" TraMineR type plots.
• Fixed small bug in MEDseq_meantime when MAP=FALSE.
• Further robustifications to handle empty components.
• Minor fixes to print.MEDseq for models where DBS &/or ASW statistics weren’t computed.
• Minor vignette edits and documentation clarifications.
• Updated citation info after online publication in JRSSA.

MEDseq v1.2.1 - (6^th release [patch update]: 2020-12-29)

Bug Fixes & Miscellaneous Edits

• The "d", "f", "Ht", "i", & "I" plot types now properly account for sampling weights.
• Layout and legend-placement has been improved for these same types of plots.
• Mimicking TraMineR further, plot.MEDseq also gains the type options "ms" & "mt".
• Minor speed-ups associated with the opti="medoid" setting.
• Added ORCID iDs to DESCRIPTION.
• Minor CRAN compliance edits to the vignette.

MEDseq v1.2.0 - (5^th release [minor update]: 2020-11-20)

Significant User-Visible Changes

• Corrected the parameter count penalty for the BIC, ICL, and AIC model selection criteria,
  specifically, the count is now greater for the central sequence position estimates.
• Hence, criterion="bic" is now the default for MEDseq_control, MEDseq_compare, and
  get_MEDseq_results (previously "dbs"), with modifications to print & summary functions.
• Non-noise components’ central sequence positions associated with precision parameters of zero
  are now printed (print.MEDseqtheta) & plotted (plot.MEDseq(..., type="mean")) always:
  the preczero argument has thus been removed from both functions.

New Features & Improvements

• MEDseq_meantime gains two new arguments (see documentation for more details):
  • weighted (default: TRUE, old: FALSE) allows the sampling weights to be used,
    with or without the cluster assignment probabilities, in the computation of the weighted averages.
  • prop (default: FALSE) divides the output when norm=TRUE by the sequence length.
• MEDseq_control gains the arg. random=TRUE, governing tie-breaking of estimated central sequence
  positions: old behaviour (always choosing the first candidate state) recoverable via random=FALSE.
• plot.MEDseq arg. quant.scale=FALSE replaces old arg. log.scale: quantiles now used
  to determine non-linear colour breakpoints when invoked with type="precision".
• Sped-up init.z="kmedoids" initialisation via pam for unweighted sequences, by using the
  highest available value for the pamonce option, based on the cluster package’s version number.
• init.z gains the options "kmodes" & "kmodes2", though only for unweighted sequences:
  both require the newly suggested klaR (>= 0.6-13) package.
• plot.MEDseq gains the arg. smeth, governing the seriation method to be used ("TSP", by default).
• For weighted sequences, init.z="kmedoids" is now itself initialised by Ward’s hierarchical clustering.
• Significant speed-ups to computation of central sequences for all opti settings (esp. "mode").
• Added SPS arg. (default=FALSE) to print.MEDtheta & summary.MEDseq.
• dbs gains the optional/experimental arg. clusters - no change to default.

Bug Fixes & Miscellaneous Edits

• Various fixes to the seriated arg. to plot.MEDseq:
  • Arg. name changed from seriate to avoid conflict with function seriation::seriate.
    
  • Fixed seriated options "clusters"/"both" for models with no noise component.
    
  • seriated="observations" (the default) now also works for type="I" plots.
    
  • seriated="clusters" now also works for type="dbsvals" & type="aswvals" plots.
• MEDseq_fit now always internally normalises the weights to sum to the sample size.
• Minor fixes to properly account for weighted sequences &/or duplicates when noise.gate=FALSE.
• Minor fix to gathering of results to account for noise.gate=FALSE when G=2.
• MEDseq_stderr now respects the algo, opti, & noise.gate settings of the original model.
• MEDseq_compare now returns & prints opti info where relevant.
• Fixes to print & summary methods for MEDgating objects if equalPro=TRUE.
• MEDseq_fit now coerces "character" covariates to "factor".
• Minor fixes to print method for MEDlambda objects also.
• Additional minor edits to plot.MEDseq(..., type="gating").
• print.MEDseqCompare gains the args. maxi & rerank=FALSE.
• Minor speed-ups for G=1 models.
• Added viridisLite (>= 0.2.0) to Suggests: (for plot.MEDseq(..., type="precision")).
• Ensured matrixStats (>= 0.53.1) and TraMineR (>= 1.6) in Imports:.
• Package startup message now checks if newer version of package is available from CRAN.
• Significant vignette edits.
• Updated maintainer e-mail address.
• Minor documentation, examples, and CRAN compliance edits.

MEDseq v1.1.1 - (4^th release [patch update]: 2020-05-12)

Bug Fixes & Miscellaneous Edits

• Maintenance release for compatibility with R 4.0.0 - minor edits.
• summary.MEDseq gains the printing-related arguments
  classification=TRUE, parameters=FALSE, and gating=FALSE.
• x$params$lambda now inherits the MEDlambda class,
  with its own print method as per x$params$theta.
• x$params$tau now has informative dimnames.
• Minor changes when supplying x.axis to plot.MEDseq(..., type="gating").
• Documentation, vignette, examples, and references improvements.
• Added rmarkdown to Suggests:.
• Reformatted package startup message.

MEDseq v1.1.0 - (3^rd release [minor update]: 2020-03-30)

New Features, Improvements, & Bug Fixes

• Significant efficiency gains when ignoring duplicates in the presence of weights:
  • before, unique cases were defined as unique sequence/covariates/weight combinations,
    
  • now, cases with different weights that are otherwise duplicates are treated as duplicates.
• MEDseq_stderr is provided for computing the standard errors of the
  coefficients for the covariates in the gating network via either the
  weighted likelihood bootstrap or jackknife methods.
• Small robustifications in the presence of empty components.
• Fixed get_MEDseq_results when what="MAP" and non-noise models are chosen.
• Fixed bug related to the colours used in the vignette plots.
• Odds ratios now returned (and printed) when calling summary on x$gating.
• Cosmetic fixes to plot.MEDseq when type="clusters" for small sample sizes.
• Other small cosmetic plotting & reference-formatting changes.
• Spell-checking of documentation and fixes to donttest examples.

MEDseq v1.0.1 - (2^nd release [patch update]: 2019-12-10)

Bug Fixes & Miscellaneous Edits

• Speed-ups to E-step, especially for models with a noise component.
• Clarifications and improvements to documentation and examples.

MEDseq v1.0.0 - (1^st release: 2019-08-24)