NEWS

disclapmix 1.7.5

New vignette "Adaptive fitting"
‘disclapmix_adaptive()': - Allowing to take 'label' and generators for 'init_y' ('init_y_generator') and 'init_v' ('init_v_generator') - refer to the "Adaptive fitting" vignette. - Allowing 'disclapmix_adaptive()' to take multiple 'criteria'’s - Not necessary to supply '_marginal' suffix for 'criteria'
Add 'cluster' argument to 'simulate()' to force simulating from a particular cluster
'disclapmix()' for 'glm_method' 'glm.fit': 'covmat' calculated more efficiently
Parameters for 'disclapmix_robust()' controlling retries with random subsets
Include check for if data contains ‘NA'’s
New 'init_y_method': 'hclust'
'predict' can now accept NA values, in which case that locus is ignored (corresponding to marginalising over that locus).

'disclapmix()' takes new agument: 'init_v' (initial posterior probabilities to get EM started)
'disclapmix()' takes new agument: 'ret_x' of whether to return data or not -- the default is not to (as opposed to earlier behaviour)
Added function 'disclapmix_robust()' which is a wrapper around 'disclapmix()' that tries to avoid errors. Can sometimes avoid errors with svd problems.
Added function 'disclapmix_adaptive()' which is a wrapper around 'disclapmix_robust()' that instead of fitting one model for a given number of clusters, fits models until the best model (lowest marginal BIC) is in the interior (with margin 'M') of all number of clusters tried. E.g., the best model has 3 clusters and the margin 'M = 5', then this function ensures that models with 1, 2, ..., 3+5 = 8 clusters are fitted. If e.g. then 7 is better than 3, then it continues such that also models with up to 7+5 = 12 clusters are fitted.

Adding suppressWarnings(RNGversion("3.5.0")) to tests as a result of https://bugs.r-project.org/bugzilla/show_bug.cgi?id=17494.

Tag for the The Journal of Open Source Software paper: http://joss.theoj.org/papers/10.21105/joss.00748

Plot functionality for disclapmix was broken for fits with only one subpopulation/cluster.

Plot functionality for disclapmix fit takes an argument for the distances between clusters (defaults to clusterdist(x)). This can be used if the distances have been precalculated.
Minor technical changes for keeping CRAN checks happy

Added AICc (Akaike Information Criterium with finite sample correction)
Added plot functionality for a disclapmix fit
Corrected number of model observations, leading to corrected BIC values (AIC not affected)
Error in full likelihood ratio corrected (not used for model selection)

Added separation of two persons mixtures (see example in ?rank_contributor_pairs)
Added AIC (Akaike Information Criterium) for a model disclapmixfit

Added simulate from fitted model
Added haplotype diversity calculation from a fitted model
Added parameter to control number of IRLS iterations (glm_control_maxit)
Added parameter to control IRLS convergence (glm_control_eps)
Changed default iterations to 100L, glm_control_maxit to 50L and glm_control_eps to 1e-6
Fixed error with no IRLS output for verbose = 2L expect for glm.fit
Small changes in verbose output

Totally rewritten such that the algorithm is more memory and CPU efficient and can be used to analyse larger datasets
Internal IRLS algorithm implemented, much faster than glm.fit for this kind of data. glm.fit is still available. See glm_method in ?disclapmix.
NOTE: Updated API

Fixed error such that verbose = 1 outputs progress information
Number of model parameters corrected, which changes marginal and full BIC slightly, but only with a constant change meaning that for the same data, the optimal number of centers is the same