Skip to contents

Nestimate 1.0.12

Nestimate 1.0.11

  • Release review: TNA shortcuts keep positional start/end calls working; markov_stability() keeps recurrence times for separate closed classes; rectangular numeric sequence matrices accept build_mcml()’s exclude, trim and end.
  • Help pages rewritten: what each function does, how to read the result, and every argument’s default and errors.
  • build_mlvar(): temporal edges ran from outcome to predictor; they now run predictor to outcome. Tiny p-values are no longer 0, and Nestimate’s own print/summary/plot are used.
  • permutation(): glasso shuffles use the same estimator as the observed networks; adjust corrects over the edges tested, as tna does.
  • extract_edges(): keeps negative edges, drops zero-weight pairs, lists undirected edges once.
  • build_mmm(): classification error is 1 - mean largest posterior (it was always 0 for k = 2).
  • net_prune(): undirected networks are pruned by pair; groups take one bootstrap per network; re-pruning after net_deprune() works.
  • centrality_stability(), casedrop_reliability(): the CS-coefficient no longer depends on the order of drop_prop.
  • passage_time(), chain_structure(): unreachable targets give Inf; absorption probabilities sum to 1.
  • Also fixed: predict_links() weighted Jaccard, cooccurrence(similarity = "relative") (now directed), compare_model(scaling = "max"), extract_initial_probs() on tna models, boot_glasso() settings and failed resamples, cluster_network() dropped arguments, declared actor/time columns read as states, prepare() time zone and is_unix_time.
  • Input errors carry the class nestimate_bad_input across the package (messages unchanged), so they can be caught by class.
  • seed = no longer moves the session’s random stream after the call; results for a given seed are unchanged.
  • plot() of an entropy_trajectory() with more than nine groups draws one panel per group instead of failing.
  • The shortcut verbs name the arguments they used to hide in ...: build_tna(), build_ftna(), build_atna(), build_cna() take actor, action, time, session; build_cor(), build_pcor(), build_glasso() take actor, level; build_ising() takes actor; sequence_patterns(), sequence_metrics(), pattern_outcome() take actor, action, time; compare_mmm() takes build_mmm()’s settings (seed, n_starts, …); macro_network() takes scaling, threshold, params. Existing calls give the same results. A wrapper given method says which method it always uses.
  • Entropy: entropy_network() refuses scaling = "share" on a zero-rate chain and "chance" on one state (was a cryptic error), keeps a bare cograph_network a netobject, and warns on a non-ergodic chain as transition_entropy() does; entropy_bayes() networks declare the TNA geometry and accept wtna() output; grouped results have summary() (a data frame) and plot(); entropy_trajectory() refuses a character time column and labels its axes correctly (row number, numeric column name, nats/hartleys).
  • mosaic_plot(residuals = "asymptotic") keeps the residuals of a table with an empty row or column (it set every tile to 0). state_distribution(), plot_state_frequencies() and mosaic_plot() refuse association and one-hot networks instead of counting their values as states.
  • compare_model(): undefined edge ratios are NA (were NaN/Inf), and logarithmic_ratio no longer warns “NaNs produced”.
  • build_mcml(method = "density"): within-cluster weights are the sum, matching the macro layer’s sum / (n_i * n_j) (they were the mean). A weight matrix with exclude, trim or end is refused.
  • Persistent homology: the Vietoris-Rips print counts essential classes and shows the ascending range; bottleneck_distance() reads a (0, 0) pair as finite, as persistence_landscape() does.
  • Also fixed: evaluate_links() validates k and returns an empty table instead of NULL; net_centrality() warns when a normalised measure is constant; landscape_trend() validates iter; network_reliability() refuses a split that empties the second half; boot_glasso() refuses a correlation matrix; path_dependence() refuses a 0/1 matrix; mogen_transitions(min_count = 0) lists only contexts that can follow each other; logical flags such as combined refuse NA; build_hypa()’s message lists none once; compare_networks() errors are also nestimate_bad_input. plot() of chain_structure() names each state’s class in its axis label (it coloured the labels, which ggplot2 warns about and which encoded the class by colour alone).

Nestimate 1.0.9

Corrections to 1.0.7, from a review of that release.

  • frequencies() and magnitude_difference() on an event log with action not declared read the log as positions without a word: group_regulation_long gave 14,034 states instead of 9, with timestamps as states. As 1.0.7 said it would, a column called action that is not declared now raises an error of class nestimate_bad_input. convert_sequence_format(), extract_sequences(), sequence_patterns(), pattern_outcome(), sequence_metrics(), stability_landscape() and sequence_plot() had the same fault, from before 1.0.7, and raise the same error.
  • A frame read with no role declared must now be readable: one column or a numbered run of positions (T1, T2, …). Any other frame raises an error of class nestimate_bad_input on every sequence verb, because reading a log as positions turns its ids and dates into states: build_network(human_long, method = "frequency") gave 507 nodes instead of 9. Calls on a wide table with other column names need state_cols in build_network(), cols in frequencies(), sequence_plot(), distribution_plot(), build_mcml(), stability_landscape(), landscape_surface(), landscape_trend(), sequence_metrics(), sequence_patterns() and extract_sequences(), or a network from build_network(state_cols = ) in build_hon(), build_hypa(), build_mogen(), markov_order_test(), path_counts(), path_dependence(), state_frequencies(), sequence_compare(), association_rules(), build_clusters(), cluster_data(), build_mmm() and cluster_mmm(). The verbs that read every column also refuse a numbered run with anything else beside it, whose values would become states.
  • build_mcml() no longer reads column names. from, to and T3 … were read as an edge list and the later positions dropped: trajectories with two positions renamed went from a macro total of 1,734 to 131. An edge list is declared with the new from, to and weight, wide data with the new cols or as a numbered run. Calls with an edge list need the arguments added. A two-column frame named V1, V2 is now two positions, with inits from its first column.
  • A declared column that does not exist now raises nestimate_bad_input: frequencies(), magnitude_difference() and convert_sequence_format() read action = "misspelled" as a wide table (38 x 38 states instead of 9 x 9), extract_sequences() ignored a wrong cols beside action, and cooccurrence() ignored a wrong field or by. cooccurrence() also refuses field, by or sep without its partner, which it read as a wide table, and any of them on a list.
  • build_network() and the verbs that set covariates aside: a numeric column beside a numeric numbered run, such as grade beside T1, T2, …, became a state: on trajectories coded as numbers it gave 138 nodes instead of 3. It is now set aside as metadata, and every column set aside this way is named in a message.
  • Positions a verb has resolved read back in: the unnumbered positions that extract_sequences(), mark_first_state(), mark_terminal_state() and a network’s $data return keep their role while the frame holds exactly those columns, in the order it then holds them. A column added afterwards is refused or set aside, not read as a state, and a role declared in the later call comes first. Base column selection drops the role, so a selected frame is declared again.
  • build_network() no longer guesses one-hot data from the values. Columns of only 0 and 1 were read as one-hot codes, also with state_cols set: 15 positions recoded this way gave 15 nodes and 30,606 transitions instead of 2 states and 1,734. On a sequence method they now raise nestimate_bad_input unless declared: format = "onehot" or codes for one-hot data, state_cols for positions. Calls that relied on the guess need the argument added. The same error stops frequencies(), sequence_plot(), build_hon() and the other verbs that read positions, which read such frames as the states 0 and 1, and so does a single 0/1 column, a numeric or logical 0/1 matrix and an undeclared 0/1 run in build_mcml(). The wtna methods read 0/1 columns as before. The association methods read them as variables with the same weights, but no longer record guessed one-hot codes and windows: those made association_rules() on a cor network mine one invented transaction, and made bootstrap_network(), centrality_stability(), network_reliability(), casedrop_reliability() and permutation() refuse a binary MGM network as one sequence.
  • build_network(method = "co_occurrence") read 0/1 positions declared with state_cols or params$cols as one-hot codes: three positions gave the nodes T1, T2, T3 instead of the states 0 and 1 (also in 1.0.8; recoded trajectories gave 15 nodes and 34,419 instead of 2 and 16,250).
  • cooccurrence() and association_rules() no longer guess binary items from the values either: columns of only 0 and 1 raise nestimate_bad_input unless the new format = "binary" declares them. Calls such as cooccurrence(bin) need the argument added. cooccurrence() on a 0/1 frame with a missing value failed with subscript out of bounds; a missing cell now counts as no occurrence.
  • build_network(): format = "onehot" counted a column named in metadata_cols as a code (4 nodes and total weight 72 instead of 3 and 36). attention, ngram, gap and reverse on one-hot data counted the raw columns, with 0, 1 and the actor ids as nodes; they now raise nestimate_bad_input.
  • A one-hot network passed to build_hon(), build_hypa(), build_mogen(), markov_order_test(), path_counts(), path_dependence(), state_frequencies(), sequence_compare(), sequence_plot(), distribution_plot(), extract_sequences() or the other sequence readers was read as positions of the states 0 and 1. These raise nestimate_bad_input. build_mcml() aggregates its weights, as for a matrix. The same holds for the one-hot networks of a netobject_group (sequence_plot() drew the states 0 and 1, build_hon() read the networks’ fields as 15 states) and for a bare cograph_network, whose declaration converting it to a netobject dropped. An association network (cor, pcor, glasso, ising, mgm) passed to a sequence reader or association_rules() raises the same error: its rows are variables. build_hon(), build_hypa(), build_mogen(), path_counts() and state_frequencies() refuse a netobject_group and any list that is not a plain list of sequences, such as a fitted t.test() result, whose fields they read as states (also in 1.0.8); so does cooccurrence(). build_mmm() and cluster_mmm() refuse an association network, also as a bare cograph_network, which they fitted from zero transitions (also in 1.0.8). build_network(method = "wtna") refuses state_cols, which it ignored: wtna reads one-hot codes.
  • build_network(method = "wtna") and "wtna_cooccurrence" read logical one-hot columns as codes, as numeric 0/1 columns and wtna() already did. They were read as the states FALSE and TRUE, and wtna_cooccurrence ignored window_size (also in 1.0.8).
  • association_rules() refuses a classed list, such as packageDescription(), whose character fields it read as transactions (also in 1.0.8), and raises nestimate_bad_input for a list holding anything other than character vectors.
  • association_rules() on a netobject_group mined only its first group, without a word; it now raises nestimate_bad_input.
  • cooccurrence() on a network read its fields as transactions (24 nodes, including NULL and frequency; also in 1.0.8), and on a logical matrix gave the nodes FALSE and TRUE. A network now raises nestimate_bad_input; logical columns are binary items under format = "binary" and refused without it.
  • Duplicated column names raise nestimate_bad_input. A column added under a name the frame already had was trusted as a marked position: build_hon() read it as a state, and build_network() gave 4 transitions where 1.0.8 gave 6.
  • The advice in these errors names only routes the verb reads: the sequence readers and attention, ngram, gap and reverse no longer suggest format = "onehot", which they then refused.
  • entropy_bayes() reads counts from a network’s declared method, not from its values. A relative or scaled network is recounted from its sequences with the network’s own start and end states, concatenation, windows and one-hot units: a deterministic chain, a scaling = "max" frequency network and a relative tna model were read as counts of one where the data hold four (also in 1.0.8), and a relative one-hot network, refused before, is counted per unit. Relative co-occurrence networks (type = "relative") and weighted networks (params = list(weighted = TRUE)) are recounted as raw counts; a relative co-occurrence network was read as counts of one (also in 1.0.8). A tna model of a counting type (frequency, co-occurrence, reverse, n-gram) gives its counts; a relative, scaled or weighted one is recounted from its sequences (a weighted tna::ftna() was read as counts of one where the data hold three; also in 1.0.8), and one with windows, attention or gap weights raises nestimate_bad_input. An association network raises nestimate_bad_input: perfect correlations were read as six transitions (also in 1.0.8). A network with weights that are not counts and no counting rule (attention, ngram, gap) raises nestimate_bad_input; it was recounted as a frequency network of its data, without its settings.
  • build_mmm() and cluster_mmm() on a one-hot network fitted every unit from zero transitions and uniform components: they counted adjacent window slots, which are never both filled. They now count within the windows, as the network does, and a bare cograph_network gets the same one-hot units instead of one unit per row (also in 1.0.8). The initial state of a one-hot unit is the code of its first window (none when that window holds several); it was read from the slot of the first code, so only units starting with that code had one. build_network() of a fit, and cluster_network(cluster_by = "mmm"), on a wtna_cooccurrence network failed with “No one-hot columns found”; the clusters are rebuilt with the network’s counting method (also in 1.0.8). A bare cograph_network keeps its declared method for that rebuild; it was rebuilt as a relative network (also in 1.0.8). A windowed tna model (from tna::import_onehot()) raises nestimate_bad_input: its windows were read as plain positions, 4 of 32 transitions (also in 1.0.8). A network converted from a tna model by cograph::as_cograph() keeps neither type nor windows: build_mmm() and entropy_bayes() read its sequences only when recounting them reproduces its weights, and refuse it otherwise (the windowed example above); entropy_bayes() read its probabilities of one as counts. build_mmm() on a netobject_group fitted its edge tables as sequences, with group labels as states, and a bare association network declaring only $method was accepted (both also in 1.0.8); both raise nestimate_bad_input. Rebuilding a fit keeps the source network’s settings: an ngram network with n_gram = 3 was rebuilt with bigrams (also in 1.0.8). build_network() of a fit with an empty component failed with “Estimator ‘frequency’ must return a list”; it raises nestimate_empty_cluster.
  • build_network(method = "wtna") and "wtna_cooccurrence" with metadata_cols read the declared column beside the codes, missed the one-hot data and ignored window_size (also in 1.0.8).
  • build_clusters() and cluster_data() clustered the rows of an association network passed as a bare cograph_network (60 measurement rows in two clusters), and the rows of a bare one-hot network instead of its units. A bare network now meets the checks of a netobject; the refusal is nestimate_bad_input (also in 1.0.8).
  • centrality_stability(), casedrop_reliability() and network_reliability(): a resample that lost a state to a constant column stopped with subscript out of bounds. It counts as below the threshold, as an unbuilt draw; split halves leave it out. The warning about failed resamples said they were left out with results NA in the stability verbs, which count them as below the threshold; it now says that.
  • A numeric column called action no longer stops build_network(method = "cor") and the other association methods, which read numeric columns.
  • build_network(): the message that a time or session column is not used named columns declared under another role: session_id given as actor, group, order or time. It is shown only for a column with no role. It also said events were read in row order and sessions not split whatever was declared, also with order set; it now says only what holds for the call.
  • build_network(): the note about one long sequence followed the arguments and was shown for several sequences, with session or with time gaps and no actor. It is shown when one sequence was built.
  • build_network(mode = "overlapping") with params = list(window_size =) on sequence data now raises the warning of class nestimate_ignored_mode that the same mode in params raises.
  • boot_glasso(): a draw without a correlation counts as below the threshold in the CS-coefficient, as in centrality_stability() since 1.0.7.

Nestimate 1.0.8

Nestimate 1.0.7

Corrections from the 2026-11 package review. Each changes results for the input it names.

  • prepare() and every verb that reads a long log: time strings must match a format completely, with a four-digit year. 15/01/2024 was read as the year 15 and the event order followed that. A two-digit year now raises an error; custom_format reads it. Fractions of a second now count in every layout; 15/01/2024 10:30:00.9 lost its .9, which could reverse two events. A string that cannot be parsed raises an error of class nestimate_bad_input.

  • prepare(): day-first or month-first is decided once per column. A month-first column was read day-first wherever the day was 12 or less, which shuffled the events. A column that cannot decide is read day-first with a warning of class nestimate_ambiguous_dates.

  • summary() and the metrics in compare_networks()/compare_model(): an undirected edge counts once, and a negative edge counts. Edge Count and Network Density were doubled on undirected networks and ignored negative edges. Mean Distance now treats a signed edge’s magnitude as its cost; it used positive edges only.

  • build_clusters(), build_mmm(): multinom runs on centred and scaled columns and reports on the original scale, and a fit that does not converge warns with class nestimate_no_converge. Standard errors were far too small on covariates such as a year or an income. $covariates$model now holds the fit on the standardised columns, and $covariates$fit$converged records convergence.

  • build_clusters(), build_mmm(): columns of a wide frame take the same roles as in build_network(). An id or a covariate was read as a sequence position, which shaped the clusters and, in build_mmm(), added its values as states. The columns set aside are named in a message and kept in $metadata, and in the $metadata of each network that build_network() makes from the result.

  • build_mcml(): a wide frame named V1, V2, V3, … is read as sequences, also when trim leaves only V1 and V2. It was read as an edge list, which dropped every column after V2.

  • distribution_plot(), sequence_plot(type = "distribution"): with scale = "count" the y axis reaches the tallest stack. It stopped at the largest single cell and clipped the stacks.

  • bootstrap_network(), permutation(), network_reliability(), centrality_stability(), casedrop_reliability() on a wtna() network, or build_network(method = "wtna"): whole units are resampled. Rows were shuffled one at a time, which broke the time order transitions are counted on: in a data set with no A->A transition the bootstrap mean for A->A was about 6, and A->B fell from 18 to about 6. permutation() with an actor failed outright.

  • Column names no longer give a role, and existing calls may need changing. build_network(), its wrappers, frequencies() and magnitude_difference() read action, actor, time and session only when they are set. A column called Action used to be read as the action, in any spelling, and could replace an explicit state_cols or the columns T1, T2, …. With nothing declared it now raises an error of class nestimate_bad_input: set action for an event log, or state_cols or metadata_cols for a wide table.

    • An undeclared Time, Session or session_id beside action is not used, and a message of class nestimate_column_role says so. Events follow row order unless time is set.
    • A numeric id beside 0/1 columns on a sequence method raises the same error. Its values used to become the network’s states.
    • A character column left out of a numbered run, which could hold states, is named in a message of class nestimate_column_role.
  • A network that rests on one sequence, and one-hot data with no actor or session or with one actor, cannot be resampled: the five verbs above raise an error of class nestimate_single_sequence, naming the group if there is one. They returned zero-width intervals and p-values of 1, one-hot data without saying so. Building such a network stays legitimate. Bootstrapping groups stops if any group has one sequence, for instance a cluster of one.

  • prepare_onehot() is no longer exported. Use wtna() or build_network(codes =, window_size =, mode =), which now count overlapping windows the same way. Its interval and aggregate options have no public route.

  • wtna(mode = "overlapping"), build_network(mode = "overlapping"): the count now matches tna::import_onehot(window_type = "sliding"). Each window marks the codes active in any of its rows, and the windows are counted in blocks of window_size. Overlapping results change.

  • One-hot data and the networks built from it:

  • bootstrap_network(), permutation(), network_reliability(), centrality_stability(), casedrop_reliability(): a resample is scaled the way the network was, zeros included for the transition methods. "minmax" used two different scales, so edges could fall outside their own intervals.

  • centrality_stability(), casedrop_reliability(): a subset in which the correlation is undefined (the measure or edge vector is constant in it), or that could not be built, was dropped from the share of draws above threshold, which could raise the CS-coefficient. It now counts as a draw below the threshold, as in tna::estimate_cs() since tna 1.5.4. The prop_above column of summary() is the same share.

  • build_network(): params = list(window_size =) on sequence data was recorded in $params and otherwise ignored, so a network that said it was windowed was not. It now counts relative, frequency and co_occurrence between consecutive blocks of that many positions, as tna::tna() does since tna 1.5.4, and the resampling verbs follow. Other transition methods raise an error of class nestimate_bad_input. mode = "overlapping" raises a warning of class nestimate_ignored_mode: blocks are the only window sequence positions have.

Nestimate 1.0.6

  • sequence_plot(periods =): a fraction in the time half of trim judges each period on its own activity. It keeps the busiest periods that together hold that share of the actions, and removes the rest wherever they sit, along with periods with no action; the kept periods close up under their own labels. It used a quantile of the actors’ last active period, which cut nothing on a calendar axis most actors run to the end of.
  • sequence_plot(), distribution_plot(): the same for a fraction in the time half of trim without periods: the busiest time points that hold that share of the observed states stay, and the rest are removed wherever they sit, under their own labels. It cut everything past a quantile of sequence lengths, so fewer or more time points can remain than before. build_mcml(trim =) still cuts by the quantile, because it changes the data.
  • sequence_plot(), distribution_plot(): the row half of trim no longer removes rows. Each row ends earlier and stays in place, with the row order and cell widths it had before: a count keeps its first that many observed states, a fraction its first share of them, and it now works with tree.

Nestimate 1.0.5

Corrections to numerical results, matching tna 1.5.4. Results from the functions below can differ from earlier versions.

  • summary() of a network reported the standard deviation of the in-strength as that of the out-strength and vice versa (the means are equal by construction). Out-strength is the row sum.
  • compare_model() and compare_networks(): the distance correlation was reported squared; it is now the distance correlation (Szekely, Rizzo & Bakirov 2007), NA when one network is constant. The Canberra distance omitted the edges present in only one network (each contributes 1). reliability() and casedrop_reliability() use the same metrics.
  • BetweennessRSP (net_centrality(), centrality_stability()) counted the correction for walks absorbed at a target once for every state, also for states that cannot reach it. In a network that is not strongly connected (for example, with a state that is never entered) the values were wrong. Only sources that reach the target are counted; strongly connected networks are unchanged.
  • net_prune(method = "disparity") set the absent edges of a state with no outgoing or no incoming edges to NA. They stay 0.

Nestimate 1.0.4

  • sequence_plot() and distribution_plot(): trim takes two numbers, the time axis and then the rows. The second drops the sparsest sequences, counted over the time points each panel draws: a count keeps those with at least that many observed states (trim = c(60, 5)), a fraction keeps the most active share (trim = c(0.9, 0.9)); NA leaves an axis alone. Works for the heatmap, index, distribution and mcml views and with periods; a row cut cannot be combined with tree.
  • sequence_plot(periods =): a Date column gives one block per calendar day and a whole-number column one block per number, from the first to the last, so a day on which nobody acted stays on the axis as an empty block. A count trim bounds the range; more than 100,000 blocks is refused.

Nestimate 1.0.3

  • sequence_plot(periods =) accepts trim and group. trim counts period blocks (trim = 60 keeps the first 60 days; a fraction keeps the periods up to that quantile of the actors’ last active period). group names a column of the long data (or of the network’s metadata) that is constant within actor and draws one panel per value, one row per actor, on a shared period axis (index type; combined = FALSE gives one page per panel). The returned list adds groups and orders when grouped.
  • sequence_plot() on long data with a missing actor or period value raises a nestimate_bad_input error naming the column, instead of prepare()’s “actor/session” message.
  • Plotting a tibble no longer warns “Unknown or uninitialised column: state_colors”.

Nestimate 1.0.2

Breaking changes since CRAN 0.9.24

  • plot_mosaic() is no longer exported (it masked tna::plot_mosaic()); use mosaic_plot(), whose data-frame method takes x_var, y_var and weight (see the 0.9.29 notes).
  • sequence_plot() gained periods, fill, actor, action, time and order after group, so arguments after group must be passed by name.
  • hypergraph_centrality() returns only the clique centrality by default; $Z and $H are computed only when requested and only for uniform hypergraphs.

New features

  • stability_landscape() draws a transition network as a ball-in-cup landscape in which each state’s cup depth is the log of its expected stay, -log(1 - P_ii) (stay = e^depth), with intervals from resampling whole sequences, the share of time in each state, its exit probabilities and, for three states, the net circulation. It takes a network, a grouped network (for example build_network() of build_clusters()) or a sequence data frame or matrix; states = sets the order of the cups. print(), summary() (table = "cups", "exits", "circulation") and plot().
  • landscape_surface() draws the cup depths as a 3D surface over a per-sequence stability measure from sequence_metrics() (self-loop tendency, entropy, complexity, …), estimated from the nearest sequences at each value of the measure, with the most stable end in front. summary() compares the two ends of each measure; plot() draws one persp() surface per network and measure.
  • landscape_trend() traces the cup depths along the program phase (sliding windows of time steps) or along a numeric per-sequence covariate such as a grade (nearest-sequence slices), one line per state with intervals from resampling sequences.

Bug fixes

  • Numeric sequence positions containing an all-missing logical column stay sequence positions rather than turning a text ID into states.

  • build_clusters() excludes requested post-hoc covariates from sequence distances, preventing the covariate from determining its own clusters.

  • loading_stability() replays the composite weighting scheme and its factor-analysis options. New fits retain custom weights/functions; older custom fits that lack them request a rebuild.

  • pattern_outcome() keeps generated outcomes separate from user metadata, including an adjustment or grouping column named .outcome.

  • compare_model() aligns named nodes before comparing matrices and rejects incompatible node sets. Unnamed inputs retain positional matching.

  • Ordinary Gaussian outcome_model() intervals now use residual-df Student t limits, consistent with their coefficient tests.

  • Z/H hypergraph centralities evaluate tensor directions in log space, preventing numerical underflow for large hyperedges.

  • Landscapes replay weighting, concatenation and boundary markers; merged sequences are bootstrap units. Covariates must be constant within merged units. Transformed or modified networks that cannot represent the stored sequence chain are refused explicitly.

  • Landscape surface/trend grids use the full range when the 5th and 95th percentiles coincide for a nonconstant measure or covariate.

  • MOGen summaries preserve numeric parameter counts above the integer limit.

  • Min-max centrality normalization preserves small, nonconstant strengths.

  • Pattern discovery explicitly refuses stars-only state labels (**, …), which conflict with wildcard syntax, with nestimate_bad_input.

  • casedrop_reliability(target = "auto") assesses all cells for association, scaled and thresholded networks, including edges that appear after case removal. Raw transition graphs retain existing-edge assessment. Explicit target = "existing" and "all" select either interpretation.

  • build_network(): wide data with number-coded states and a text column (an id) built the network from the text column alone, an all-zero network with no warning. A numbered run of numeric position columns is now the sequence, and the other columns are metadata.

  • The sequence verbs (extract_sequences(), sequence_patterns(), sequence_metrics(), pattern_outcome()) read the data of a correlation, glasso, Ising, MGM or window network as states; they now raise nestimate_bad_input. A network grouped by two columns (group = c("Achiever", "Course")) no longer fails; its groups are labelled in one column, Achiever-Course.

  • sequence_metrics(): a row with no observed state has NA, not 0, for valid_proportion and cyclic_feedback_strength.

  • sequence_patterns(pattern =) accepts the spaced notation "A -> * -> B".

  • nct(): an unnamed matrix compared with a named one is matched by position again, as documented.

  • build_hon(), build_mogen() and path_counts(): a missing value inside a sequence ends a run of observed states. build_hon() stopped with an error, build_mogen() counted NA as a state, and path_counts() joined the two sides of the gap.

  • association_rules() reads factor columns of a wide data frame by their labels, not their integer codes.

  • build_simplicial(type = "pathway") leaves out HYPA paths that never occur: since 0.9.26 HYPA scores every possible path, and an unobserved one flagged “under” became a simplex of states never seen together.

  • persistent_homology(): the rows of $persistence are numbered without gaps.

  • labels = on number-coded states: every state column is relabelled together, so bootstrap_network() no longer drops the cells of columns that held no relabelled state.

  • landscape_surface() and plot(landscape_trend(), type = "surface") draw a flat surface instead of failing; a constant measure or covariate raises nestimate_bad_input; slices resting on fewer than five effective sequences warn (nestimate_thin_slice); input with no observed transition raises nestimate_landscape_unestimable; a two-state occupancy landscape reports its single ridge once.

  • plot() of a passage_time() result drew the matrix transposed: the tile in row A, column B showed the passage time from B to A. Rows are now the starting states and columns the targets, as documented, for a single network and for each panel of a grouped one.

  • pattern_outcome(): an adjust column named like one of the model’s own terms (.p1, .y, …) was replaced by that term, so the requested adjustment was silently lost; an adjust name with a space failed to parse. Every model column now has an internal name. outcome defaults to "last_obs", as in codyna and as the 0.9.29 notes stated.

  • sequence_patterns(): a gapped pattern (A->*->B) no longer spans a missing position, matching the template search; a repeated len or gap (len = c(2, 2)) no longer counts each occurrence twice; a grouped search that finds no pattern returns zero rows with all its columns and no warning; a template ending in -> is refused.

  • build_network() and extract_sequences(): an all-missing position column stored as logical NA no longer stops the numbered position names from forming a run, which had turned a character covariate into states.

  • plot() of a grouped markov_stability() result handles more than nine groups: colours cycle and shapes shift, so every group keeps its own colour-shape pair.

  • extract_sequences() and the sequence verbs raise nestimate_bad_input for a model built from a matrix, which holds no sequences.

Documentation

  • The pkgdown reference lists the new sequence-analysis and long-data topics.

  • sequence_metrics(): the emergent-state rule for a return to the first state is a spell of at least 3 that is more than half the opening spell (codyna’s rule); the help page said “over twice”.

  • Documentation revision (2026-11-01): the README examples used column names that the bundled data do not have (action, time, phase), arguments that do not exist (markov_stability(times =), build_simplicial(method =), distribution_plot(type =), evaluate_links(held_out =)) and undefined objects; every README example now runs on bundled data and uses the package’s verbs instead of $ access. Vignette prose now matches the rendered output (markov-stability, transition-entropy), the sequence-comparison vignette names its dataset correctly, and the vignettes and examples no longer reach into result objects with $. References were added to build_network(), build_mlvar(), build_gimme(), build_clusters(), cluster_choice(), build_mmm() and mosaic_plot().

Nestimate 0.9.29

Sequence patterns

  • extract_sequences() returns the sequences behind any input as a wide data frame, one row per sequence: prepared data, a long event log (actor/action/time), a wide data frame or character matrix, a netobject, a grouped netobject_group (the group column is added) or a tna model. meta = TRUE adds each sequence’s covariates.
  • sequence_patterns() discovers ordered patterns: contiguous n-grams, gapped patterns (A->*->B) and repeated runs. For each it reports the occurrences, the share of patterns of that length, the number and share of sequences that contain it (support), and its lift. The counts, support and lift agree with codyna::discover_patterns(). With group (a column, a vector, or the grouping of a netobject_group) each pattern also gets its support per group and a test of presence by group: for two groups the odds ratio, 95% CI and p-value of Fisher’s exact test, for more the chi-square and Cramer’s V; p-values are BH-adjusted. top returns the first rows.
  • sequence_patterns(pattern =) searches for a template such as "discuss->*->consensus" (each * is one position of any state) and returns every concrete instance; group = "last_obs" groups sequences by their final state. Both agree with codyna::discover_patterns().
  • sequence_metrics() computes 23 per-sequence indices (validity, spells, entropy, Simpson diversity, transitions, initial-state persistence and decay, dominant and emergent states, cyclic feedback, integrative potential, complexity), one row per sequence. They agree with codyna::sequence_indices() on 20 indices; three are corrected: dominant_max_spell is per sequence (codyna repeats the last sequence’s value), longitudinal_entropy is normalised by the number of states (codyna counts NA as a state), and complexity_index uses that entropy once (codyna normalises it twice) and is 0 for a sequence of one spell, as in TraMineR::seqici() (codyna: NA). Empty sequences give NA. A missing position is not an observation: the opening spell starts at the first observed state, transitions are pairs of adjacent observed states (self_loop_tendency + transition_rate = 1), and integrative_potential weights observed positions only, as TraMineR::seqintegr() does. The indices shared with TraMineR agree with it on sequences without internal gaps.
  • extract_sequences() and the sequence verbs read TraMineR’s void (%) and missing (*) markers as NA, as build_network() does; an stslist such as tna’s engagement no longer gains a % state.
  • pattern_outcome() reproduces codyna::analyze_outcome(), including its default outcome = "last_obs": the n patterns ranked highest by priority (default: codyna’s chi-square) enter one logistic regression of a two-valued outcome on their presence, with an optional random intercept (group). Coefficients agree with codyna’s fit. It returns a tidy table, one row per pattern, with support per outcome value, the selection statistic, estimate, Wald CI, p-value and odds ratio. It fits the selected pattern even where codyna 1.0.0 fits a data column instead (a lone pattern of one length loses its name, and a data column named V<k> takes its place).

Validation tests aligned with the current tna release

  • plot_mosaic() is no longer exported: it masked tna::plot_mosaic() whenever both packages were attached, so tna code (and apps such as Dynalytics) calling plot_mosaic(model) failed with argument "data" is missing. Its data-frame mosaic is now a method of mosaic_plot(): plot_mosaic(df, x = "group", y = "state", weight = "count") becomes mosaic_plot(df, x_var = "group", y_var = "state", weight = "count") (x_label/y_label become xlab/ylab, as in the other methods). A network, table or matrix given a data-frame-only argument (x_var, y_var, weight, fill, colors, show_labels, and label_size except with style = "flat", which uses it for its tile labels) raises nestimate_bad_input instead of ignoring it, and so does an unknown argument of the data-frame method.
  • casedrop_reliability() compares only the edges present in the original network. A subsample cannot contain a transition that is absent from the full data, so the empty cells agreed in every draw and inflated the correlation: on a sparse 8-state chain the CS-coefficient was 0.9 (the top of the grid) with a mean correlation of 0.993 after dropping 90% of the sequences. n_edges now counts the present edges, and an undefined CS-coefficient (edge weights that do not vary) is NA instead of 0. Such a result now has the usual summary columns with NA statistics, so summary() (including the grouped summary) and plot() work: the plot states that the CS-coefficient is undefined, and the grouped plot names the networks it leaves out. This case is common now, for example a fixed cycle, where every present edge has probability 1. The documentation no longer claims that case-dropping is valid regardless of within-actor correlation.
  • permutation(measures = "Diffusion") (and the node table of compare_networks()) compares Diffusion after min-max scaling to 0-1, as net_centrality() reports it and as tna does from 1.4.1.
  • The range normalization behind normalize = TRUE and Diffusion returns NaN for a numerically constant vector instead of stretching rounding error to 0-1, as tna’s does.
  • markov_order_test() splits sequences at missing values, so no window spans an NA; an internal NA previously became a context named “NA”. Results on data without internal missing values are unchanged (identical under the same seed). Its documentation no longer calls the within-context permutation exact: the windows of a sequence overlap, so the permutation distribution approximates the null (Besag & Mondal, 2013).

Bug fixes

  • plot() works on passage_time() and markov_stability() results for a grouped network: one heatmap panel per group on a shared colour scale, and a per-metric dot plot with one point per group. Both used to fall through to plot.default() and fail. The single-network stability plot uses one fill; its per-state palette left a ninth state without colour.
  • build_mmm(covariate_effect = "posthoc"): print() and summary() labelled the covariates “integrated into EM – influences cluster membership” whatever the setting. The fit records covariate_effect, and a post-hoc analysis is now labelled as one.
  • The print() methods of entropy_trajectory() and entropy_bayes() no longer tell the user to reach into the object with $; they point to summary() and plot().
  • build_network() no longer reads a character covariate of wide sequence data as states. An outcome or group label such as High/Low used to become nodes of the network and vanish from $metadata. When the position columns are named as a run (T1, T2, …; V1, …; 1, 2, …), every other character or factor column is now metadata, like numeric columns already were, even when its values coincide with state names. The network equals the one built with metadata_cols. Without such a run, or with two runs of equal size, nothing changes.
  • Association permutation() refuses inference with nestimate_permutation_failed when any permuted network cannot be estimated or has non-finite weights, reporting the number of failed draws. Failed draws no longer silently reduce p-values or invalidate null moments.
  • Transition resampling honors weighted and concat. Concatenated rows are resampled as the merged sequences used by the estimator, including their joining transitions. With actor, a merged sequence must stay within one actor; paired tests require matching merged-sequence counts.
  • Relabelling numeric sequence states, start/end markers and onehot columns keeps stored data, alphabets and column selectors consistent for resampling. Wide sequence positions retain their column names; association measurements retain their numeric values.
  • Higher-order builders and sequence_compare() preserve literal numeric states in Nestimate models and decode positional codes only for tna input or explicitly encoded converted data. Character digit alphabets such as "1" through "12" are decoded correctly.
  • Grouped wide sequences use the same resolved state columns as their child estimates. Numeric covariates do not enter the shared alphabet, and factor columns keep their labels.
  • MGM’s numeric level is validated after excluding group, actor and metadata columns. MGM actor identifiers are excluded from estimation.
  • Custom variable estimators retain their numeric cleaned data when column names identify the modeled nodes, allowing valid downstream resampling.
  • nct() rejects Inf and -Inf instead of silently isolating the affected variable; its supported missing-data and constant-variable handling is unchanged.
  • plot_mosaic(network, x_label =, y_label =) applies the labels. Explicit primitive-only styling arguments raise a classed error directing the caller to mosaic_plot().
  • association_rules() reads a character matrix of states (such as trajectories) as wide sequences, like a character data frame. It used to treat every matrix as binary: with missing cells it failed with “subscript out of bounds”, and without them it silently mined the column names. Binary input (matrix or data frame, 0/1 or logical) with a missing cell or a value other than 0/1 now raises nestimate_bad_input; logical data-frame columns are read as binary items instead of the strings "TRUE"/"FALSE".

Documentation

  • MGM’s missing-data policy is mean/modal single imputation, as documented by psychnets for its "pairwise" option; it is not pairwise deletion.
  • Permutation M/S labels describe this function’s historical total/max edge difference statistics and distinguish them from NCT’s structure/strength convention. Comparison tables no longer call total edge difference “global strength”. The NCT tutorial and local display checks use current notation and capitalization.

Nestimate 0.9.28

Higher-order methods: corrections from a review against the source papers

These change results. Each is tested against an independent value (a hand-derived oracle or a published example), and each new test fails on the previous code.

  • hypergraph_centrality() computes the clique-motif centrality (CEC) directly as the Perron vector of each connected component. The power iteration it replaces oscillated when the clique expansion was bipartite (any triangle-free hypergraph), and returned a different wrong vector depending on whether max_iter was odd or even, without a warning. A star of pairs now gives the exact vector for every max_iter.
  • hypergraph_centrality() now defaults to type = "clique". "Z" and "H" are defined by Benson (2019) for uniform hypergraphs only: with mixed hyperedge sizes the equation mixes polynomial degrees and its solution depends on the normalisation inside the iteration. Requesting them on a non-uniform hypergraph raises nestimate_nonuniform_hypergraph. On uniform hypergraphs they are unchanged and reproduce Benson’s Table 2. An unconverged Z/H iteration warns with nestimate_no_converge.
  • q_analysis() counts, at each level q, only the maximal simplices of dimension at least q (Atkin’s Q-vector). It previously counted every maximal simplex at every level, so lower-dimensional simplices appeared as extra components at high q.
  • persistent_homology() builds the filtration one dimension above max_dim (the Ripser convention), so classes of dimension max_dim can die. Previously every top-dimension cycle was reported as essential; with the default max_dim = 3, dense 7-state transition networks showed spurious never-dying bars in every case tried.
  • build_simplicial() (clique and VR) records when max_dim cut a larger clique. betti_numbers() then warns (nestimate_truncated_complex) that the top Betti number is an upper bound, and persistent_homology() warns the same way when given such a complex, whose top-dimension classes cannot die.
  • build_mogen() and markov_order_test() gain dof = c("paths", "observed"). The default "paths" counts each layer’s parameters as Scholtes (2017, Eq. 7) and pathpy do: every transition of the paths of length k that the observed first-order topology allows. The previous rule, now dof = "observed", counted only observed transitions; it can fall as the order rises and under-penalised high orders on sparse data, so AIC, BIC and the likelihood-ratio test favoured higher orders. Selected orders can change.
  • build_mogen(criterion = "lrt") selects the largest order whose test rejects (Scholtes 2017) instead of stopping at the first non-rejection, and returns every p-value as lrt_p.
  • build_hon(method = "hon", min_freq >= 2): a source whose transitions are all below min_freq is no longer an extension candidate nor a starting rule, as in the reference pyHON. Such sources could previously add HON nodes that pyHON does not create. min_freq = 1 and HON+ are unaffected.
  • The Z and H iterations of hypergraph_centrality() normalise the tensor update before adding the stabilising shift, and count as converged only when the eigen-equation residual is small. On large uniform hyperedges the unnormalised shift swamped the update, so the iteration stopped at its start and returned the uniform vector without a warning (a 3-petal sunflower of 20-member hyperedges gave a hub/petal ratio of 1 instead of Benson’s sqrt(3)).
  • Path degrees of freedom are summed as non-negative per-state terms, so they stay exact up to the double range; beyond it build_mogen() and markov_order_test() stop with nestimate_dof_overflow instead of carrying Inf/NaN into AIC and BIC.
  • Truncation detection reads the simplices already built (does a top simplex extend by one vertex?) instead of an exact maximum-clique search, which took seconds on dense graphs even at max_dim = 0; persistence on a distance matrix no longer computes it at all. The persistence warning for a truncated complex rechecks within max_scale.
  • build_hypa() classifies a path as over-represented when p_over <= alpha, i.e. HYPA >= 1 - alpha as in LaRock et al. (2020); it was < alpha. This only changes ties at exactly alpha.

Documentation

  • build_mogen() is documented as the multi-order model of Scholtes (2017), pathpy’s MultiOrderModel; it is not the multi-order generative model of Gote et al. (2023), which adds start and end states.
  • build_honem() states what it computes (decayed matrix powers of the HON transition matrix over all HON nodes) and how that differs from Saebi et al. (2020), whose order-indexed matrices over first-order nodes are not fully specified in the paper.

Nestimate 0.9.27

New features

  • sequence_plot() gains periods = and fill =. For a network built with actor = and session = "day" (or any period column), sequence_plot(net, periods = "day") draws one row per actor and splits the time axis into equal-width blocks, one per period, with ticks at the period boundaries and each label under its block. The number of steps may differ by actor and by period: fill = "stretch" (default) spreads an actor’s steps over the whole block (one step fills it, three take a third each); fill = "pad" gives every block as many fixed slots as the busiest actor-period and leaves the rest empty, so a slot is the same step position for everyone. Factor levels set the period order and an unused level shows as an empty block. Works for type = "heatmap" and "index".
  • sequence_plot() accepts long data directly: actor =, action =, time = and order = name the columns, as in build_network(), e.g. sequence_plot(log, actor = "student", action = "action", periods = "day"). With periods, each actor-period is one sequence whatever the time gaps.

Breaking changes

  • nct(): the M and S letters were the wrong way round. In van Borkulo et al. (2022) and NetworkComparisonTest, M is the network-structure test (largest absolute edge difference) and S the global-strength test; nct() stored and printed them swapped. The tests are now $structure (M) and $strength (S, with by_group, each network’s global strength); reading $M / $S raises a nestimate_renamed error naming the field to use.
  • summary.net_nct() now returns the M and S tests (columns test, statistic, observed, p_value, significant) instead of the per-edge table with m_stat / s_stat attributes (those attributes had the M/S swap too). The per-edge table is as.data.frame().

Bug fixes

  • plot_mosaic() given a network (netobject, group, mcml, htna) hands it to mosaic_plot(); it failed with “is.data.frame(data) is not TRUE”. Data-frame calls are unchanged.

  • cluster_network(cluster_by = "mmm") passes covariate_effect (and any other build_mmm() argument) to build_mmm(); it failed with “Unknown argument(s) in cluster_network(): covariate_effect”.

  • build_mcml() decodes a tna object’s integer-coded sequences (it failed with “Unknown nodes in clusters”).

  • summary.mcml() works without within-cluster layers (compute_within = FALSE, where it crashed), and within_total includes self-transitions. For raw counts it equals the macro diagonal; separately normalized micro and macro layers can have different totals.

  • entropy_trajectory() windows run through time. Transitions were pooled actor by actor, so a window could start in one actor’s late events and end in another’s early ones.

  • Markov-chain measures (transition_entropy(), entropy_network(), passage_time(), markov_stability(), chain_structure()) refuse correlation-type networks and negative weights (nestimate_not_markov) instead of row-normalising them into a meaningless chain. Co-occurrence networks (a random walk) are still accepted.

  • path_dependence() no longer drops states whose names contain “NA” (e.g. “FINAL”, “ANALYSIS”); missing cells are tested directly.

  • sequence_compare() on a network whose states are numbers (1, 3, 5) reads them as states; they were decoded as positions, relabelling one state and dropping another.

  • build_mmm(): BIC/AIC/ICL count covariate-dependent mixing by design column (a 4-level factor is 3 columns); and a sequence whose first state is missing adds no initial-state term instead of being scored as starting in the first state.

  • set_state_colors() on a netobject_group stores the palette as an attribute; it was added as a list element, after which every plot, print and summary of the group failed.

  • sequence_plot() heatmaps pin each state to its colour; when the drawn part held only some states (e.g. after trim), cells were coloured as other states. Panels with a single sequence (one row, a singleton cluster, one actor with periods) draw instead of erroring.

  • build_network() refuses a compare_mmm() table with a classed error (nestimate_bad_input) giving the call to run, with the lowest-BIC k: build_network(build_mmm(data, k = 2)). The table is a data frame, so it used to be read as sequence data and “built” into a network whose states were its log-likelihoods and BICs.

  • permutation() on cor, pcor and glasso networks pools the two datasets by column name. It pooled them by position, so the same variables in a different column order were mixed in the permutation null (one edge’s p-value moved from 0.47 to 0.71 on identical data).

  • boot_glasso() edge and centrality difference tests count a tie (both values equal, e.g. two absent edges) in both tails: p = 2 * min(P(d >= 0), P(d <= 0)), capped at 1. Ties used to count as evidence of a difference; with one edge present, 29 of 45 edge pairs were “significantly different” (now 9, the present edge against each absent one).

  • Betweenness, edge betweenness and the shortest-path counts behind them use one relative tolerance for “same length”. A tie that floating-point addition breaks (1/10 + 1/15 vs 1/6) was counted as one path but credited as a full path, so node b got 1 where igraph gives 0.5.

  • Ising networks can be resampled: bootstrap_network(), network_reliability(), centrality_stability() and casedrop_reliability() returned NaN or CS = 0 for every Ising network, because the estimator rejected the stored data matrix and every draw failed silently. The resamplers now also warn (nestimate_resample_failed) with the number of draws whose network could not be estimated, instead of dropping them without a word.

  • build_hon(), build_hypa(), build_mogen(), path_counts() and state_frequencies() read data frames column by column, so a factor column next to character columns keeps its labels instead of becoming states “1”, “2”, …

  • build_network() input handling:

    • actor = now names the ID column in wide and association data, as it already did in long data. Before, the ID values became nodes (wide sequences) or the ID column a variable (cor, pcor, glasso, Ising, which errored on a non-binary ID).
    • A numeric column in character sequence data (age, score) goes to $metadata instead of becoming states.
    • cor, pcor and glasso convert data to state frequencies only when it has at least two categorical columns and fewer than two numeric ones (sequences, possibly with one numeric covariate, which is left out). One character column among numeric variables (e.g. gender) used to turn the whole network into a network of that column’s values; the numeric variables are now the nodes. Declared metadata_cols are dropped before this check.
    • labels = now also relabels $data, $initial, $frequency_matrix and $node_groups, so bootstrap_network() and the other resamplers no longer return all-zero matrices for a relabelled network, and predictability() no longer reports NA RMSE.
    • prepare()’s unix_time_unit, is_unix_time and custom_format passed to build_network() reach prepare() (they were silently ignored), and prepare() warns (nestimate_time_unit) when numeric times read in the given unit fall after the year 2500, the sign of millisecond timestamps read as seconds.
    • level = "both" keeps actor and labels; a grouped call keeps labels; an NA in the group column is an error (nestimate_bad_input) instead of a hidden all-NA group.
  • Start/end markers (start = TRUE, end = TRUE) are now honoured by the precomputed resampling path behind bootstrap_network(), permutation(), centrality_stability(), network_reliability() and casedrop_reliability(). Before, the Start row and End column were zero in every replicate (bootstrap mean, SD and CI of 0; every Start edge “significant” in a permutation test).

  • build_mmm() / cluster_mmm() with covariates and covariate_effect = "em": the covariate-dependent mixing step is now an exact M-step, nnet::multinom() fitted to the posterior probabilities. The former hand-written Newton step had no step control; with k >= 3 it diverged (coefficients around 1e7 to 1e9, an empty cluster, a log-likelihood below the fit without covariates) while reporting converged = TRUE. The likelihood now rises at every EM iteration, and converged also requires the last mixing step to have converged. Covariate fits change; $covariates$beta columns are now named.

  • nct() compared the two groups’ variables by position: a second group whose columns arrived in another order was tested against the wrong variables with no error. Columns that differ in name or order now raise nestimate_bad_input, as psychnets’ net_compare() does.

  • nct() failed with the uninformative nearPD error “infinite or missing values in ‘x’” on any missing value, on a variable that is constant in a group, and on a permuted split in which a rare binary variable becomes constant. New na_method = c("pairwise", "listwise") (default "pairwise": pairwise correlations, median co-observation count as the sample size; listwise drops incomplete rows, whole pairs when paired = TRUE). A constant variable enters the network as an isolated node (the psychnets convention), with a nestimate_constant_variable warning when it is constant in an observed group; the statistic is then defined on every permuted split, so the test stays exact. A pair sharing fewer than two observed rows raises nestimate_degenerate_sample.

  • nct() drops non-numeric columns (an id or group label) instead of failing in cor(), validates iter as a single whole number (2.7 was silently truncated), and raises nestimate_bad_input for malformed arguments.

  • summary.net_nct() labelled every edge edge_1, edge_2, … with to = NA when the data had no column names; edges are now labelled V1, V2, … $E$edge_names is always a from/to data frame (it was an expand.grid() result with Var1/Var2, or NULL).

  • Printing an nct() result shows everything: both groups by name, M and S with each group’s global strength, and every edge (largest difference first) with both groups’ weights (print(x, n = 10) shortens the list) – the same layout as psychnets’ net_compare(). New as.data.frame.net_nct(): one row per edge (from, to, weight_1, weight_2, difference, p_value, significant).

  • nct() treats a network weight of |w| <= 1e-10 as no edge. psychnets before 0.6.0 can return an absent edge as a ~1e-17 residue, and two empty networks then tested as different (M = 5.6e-17 and a p-value below 1; 10 of 40 random n = 60, p = 12 null datasets were affected with psychnets 0.5.2).

  • On complete data every nct() number is unchanged (frozen in tests/testthat/fixtures/nct-baseline.rds).

Nestimate 0.9.26

Breaking changes

  • build_hypa() now defaults to p_adjust = "none", following LaRock et al. (2020), who classify paths against the threshold alpha without multiple-testing correction. Pass p_adjust = "BH" for the previous behaviour.

Bug fixes

  • build_hypa() used a miscalibrated null model. The propensity matrix was s_out * s_in restricted to observed paths, normalised over the whole graph and never refitted, so expected counts did not preserve node strengths: high-strength contexts were systematically under-represented and rare contexts over-represented, and memoryless (first-order Markov) data produced dozens of anomalies. Xi is now built on every possible De Bruijn edge (an observed k-gram continued by an observed first-order transition) and refitted until expected in- and out-strengths match the observed ones, following LaRock et al. (2020) and the reference tlarock/hypa code.

  • build_hypa() fits the propensity matrix by continuous iterative proportional fitting instead of a rounded, keep-best sweep. Expected counts now preserve every De Bruijn node’s in- and out-strength to numerical precision (order 2 equals the first-order Markov expectation exactly), and converged is reported only when the tolerance is met; otherwise a warning of class nestimate_hypa_no_converge is raised. Stalled fits were previously reported as converged. Expected counts can differ from the original Python implementation, which rounds inside every sweep, by a few thousandths.

  • build_hypa() validates its input: min_count must be a single whole number, alpha a single number in (0, 0.5) and order free of duplicates (duplicates previously doubled every row and anomaly). Internal missing states now split a trajectory instead of becoming a state called “NA”, and labels containing the internal separators \x01/\x02 raise an error of class nestimate_hypa_bad_label.

  • build_hypa() now uses the over-representation rule of LaRock et al. (2020): a path is over-represented when its HYPA score P(X <= f) exceeds 1 - alpha, so p_over is the strict tail P(X > f) = 1 - p_under (previously the inclusive P(X >= f)).

  • build_hypa() can now flag paths that never or rarely occur as under-represented. min_count gates over-representation on the observed count and under-representation on the expected count; previously a path observed fewer than min_count times could never be a deficit. Possible but unobserved paths are now scored (observed = 0), so $scores and n_edges contain more rows. At order 2 the expected count reduces to the first-order Markov expectation.

  • print.q_analysis() reported the fragmentation point from the highest q down, because q_vector is stored from the highest q; it now reports the lowest q with more than one component (e.g. q = 2, not q = 4, for a clique complex whose maximal simplices meet only in triangles).

Improvements

Nestimate 0.9.24

CRAN release: 2026-10-07

Documentation

  • Fixes to the 0.9.23 help-page merge, found in review:
    • Pages are named after their function again: build_mcml and plot_state_frequencies had been named after a print() method, and their usage now lists the function before its methods.
    • The descriptions of 51 methods (what a plot draws, what a print shows) are restored, as a “Methods” section on each function’s page.
    • Each page’s ... entry states what each method does with extra arguments: methods that reject them say so, plot.net_bayes() names cograph::splot(), and the rest say they are ignored.
    • compare_networks: the ... and labels entries no longer repeat; net_edge_betweenness and net_pruning_details no longer list a non-existent class.
    • Every S3 method has a documented return value.

Behaviour notes

  • The mgm estimator rejects scale = FALSE with a classed error (nestimate_mgm_unscaled); this started in 0.9.23 and was not listed there. Missing values use mean/modal single imputation under psychnets’ named "pairwise" policy, now passed explicitly.

Tests

  • Two more reference-implementation tests (a hand-written linear solve for passage_time(), a manual multinom re-computation) and unused reference helpers moved to the local suite; section headers that claimed equivalence tests are renamed. New tests: mgm on data with missing values, and the scale = FALSE error.

Nestimate 0.9.23

Internal changes

  • The mgm and ising estimators delegate to psychnets::mgm_fit() and psychnets::ising_fit() (glmnet engine), joining cor, pcor and glasso. Results are unchanged: identical to the former in-package code (max absolute difference 0 for Ising over 180 configurations, 5.6e-17 for MGM over 48). Nestimate keeps type detection, input validation and the result fields. The Ising result no longer carries the unused asymm_weights and lambda_selected fields. The unused internal moderated-MGM code is removed.

Documentation

Bug fixes

  • build_network(method = "mgm") no longer fails on data with missing values (“number of observations in y not equal to the number of rows of x”).

Tests

  • Equivalence tests against other packages (tna, igraph, mgm, stringdist, glm()) and naive-loop reference implementations are no longer shipped.
  • CRAN runs a core of 15 test files (data preparation, estimation registry, extraction, Markov, link prediction, Bayesian verbs, pruning, and the htna contract); every other file is skipped on CRAN and runs locally and in CI. The CRAN test run drops from about 300 s to 22 s.
  • Slow examples run with fewer iterations (boot_glasso(), nct(), permutation_diagnostics(), build_gimme()) and ?sequence_plot draws fewer figures.

Nestimate 0.9.21

Bug fixes

  • summary() of a predict_links() result reports NA scores for a method with no predictions (every possible link already exists). It returned NaN / -Inf / Inf with one warning per column.

Documentation

  • Every help page now has example code that runs under R CMD check. 31 pages had all of it inside \donttest{} or \dontrun{}; fast examples are unwrapped, and slow ones (nct(), permutation_diagnostics(), entropy_bayes()) run live with fewer iterations.
  • The predict_links() family of examples uses fixed data that leaves links to predict, and the loading_stability() / as_networks() examples use items with real two-factor structure, so neither emits warnings.
  • pkgdown: the cograph-tutorial-nestimate article was a 47.8 MB page (300 dpi figures inlined as base64); it is now about 70 KB with figures as separate files. The README no longer lists outdated imports.

Nestimate 0.9.20

Documentation

  • The documentation site is now https://pak.dynasite.org/Nestimate/ (DESCRIPTION URL, _pkgdown.yml, README links). saqr.me/Nestimate was a stale copy. The README lists the permutation-nested-data and compare-networks articles.

Nestimate 0.9.18

Documentation

Nestimate 0.9.17

New features

  • compare_networks() gains actor =, passed to the permutation backend: whole actors are reassigned between the networks, the printed header names the actor column, and global_differences() gains the rows ICC, Design effect (edges) and Design effect (M) (category "Nesting"). Requires test = "permutation" (error class nestimate_compare_actor_needs_permutation).

Nestimate 0.9.16

Breaking changes (development versions only)

  • The nesting argument added in 0.9.13 (permutation(), permutation_diagnostics()) and 0.9.15 (bootstrap_network()) is renamed from block to actor, the column identifying whose sequences they are: the same vocabulary as build_network(actor = ). Results carry actor and n_actors; diagnostic columns use _sequence (sequences reassigned) and _actor (actors reassigned), e.g. p_global_sequence / p_global_actor; conditions are nestimate_bad_actor, nestimate_actor_missing, nestimate_actor_misaligned, nestimate_actor_unsupported and nestimate_few_actors. The printout reads Actor: Group (200 actors) and Nesting in Group: ICC = .... Neither name was ever on CRAN.

Bug fixes

  • build_network(): columns named in metadata_cols (or left out of state_cols) are no longer read as sequence positions in wide data. They were moved to $metadata only after estimation, so their values became states and created spurious transitions, in single and grouped networks. For cor, pcor, glasso, ising and mgm, declared metadata columns are no longer entered as variables.

  • An ICC that cannot be estimated (one block, or no variation) is reported as not estimable instead of NaN [NA, NA] by permutation(), bootstrap_network() and permutation_diagnostics().

Nestimate 0.9.15

New features

  • bootstrap_network() gains block =, the column identifying the unit that sequences are nested in. Whole units are resampled with replacement, keeping their sequences together (cluster bootstrap, top level only; Davison & Hinkley, 1997; Field & Welsh, 2007). The result reports the ICC (Shrout & Fleiss, 1979) and the design effect (Kish, 1965) for the edges, printed under the bootstrap summary and returned as clustering and clustering_edges. With one sequence per unit it reproduces the ordinary bootstrap exactly. Transition networks only.

Nestimate 0.9.14

Documentation

  • New pkgdown article “Comparing two networks with compare_networks()”: high vs low achievers in group_regulation_long, every summary table and all eight plot views with a short reading of each.

Nestimate 0.9.13

New features

  • permutation() gains block =, the column identifying the unit that sequences are nested in (sessions in students, students in teams). Whole units are reassigned between the groups; units with sequences in both groups, as in repeated-measures designs, have their sequences reassigned within the unit; mixed designs combine the two (Good, 2005; Anderson & ter Braak, 2003). Supported for transition networks (relative, frequency, co_occurrence); association networks raise nestimate_block_unsupported. A warning of class nestimate_few_blocks is raised when no p-value could fall below alpha.

  • With block, permutation() reports the nesting effect: the intraclass correlation (Shrout & Fleiss, 1979) with a jackknife 95% interval, and the design effect (Kish, 1965), the blocked over the unblocked null variance, for the edges and the global statistic M.

  • permutation_diagnostics(x, block =) runs the ordinary and the blocked test on the same data and returns them side by side as a tidy data frame, one row per pair or, with level = "edges", per edge.

Improvements

  • print() of a net_permutation shows the global test (M and S, with p-values) and, with block, the ICC and design effects. A grouped result prints every pair in full.

  • The permutation() help page is reorganised into sections: what is tested, nested data and block, ICC and design effect, reading the printed output.

  • Building a transition network from a single long sequence now raises a message instead of a warning; running bootstrap_network() or permutation() on such a network still warns. Both carry the class nestimate_single_sequence.

Bug fixes

Nestimate 0.9.12

New verbs

  • set_state_colors(x, colors) attaches a palette to a netobject, netobject_group, mcml or htna, and every figure drawn from that object then uses it – sequence_plot(), distribution_plot(), plot_state_frequencies() and cograph::splot(). The cograph half goes through the documented meta$splot producer contract (defaults$node_fill, stamped in node order), so splot(net) needs no arguments and renders byte-identically to splot(net, node_fill = ...). A state_colors(x) <- replacement form is also available.

  • state_colors(x) reads the resolved palette back as a tidy data.frame: one row per key, with state, color and source ("set" or "default"). It is what the figures actually draw, not just what was passed in.

Improvements

  • States a named palette does not mention are now dealt the Okabe-Ito colours the palette has not used, one each, instead of taking their positional slot. Pinning plan = "#0072B2" no longer leaves another state defaulting to that same blue. Affects the named form only; an unnamed state_colors and the all-default palette are unchanged.

  • An attached palette resolves once, in the object’s own key order, so every figure drawn from it colours a state identically. Previously only the colours you set were carried and each figure dealt the remaining defaults in its own sort order – plot_state_frequencies() sorts states by frequency and sequence_plot() alphabetically, so one state could come out amber in one figure and black in the other.

  • In-tile labels in plot_state_frequencies() take whichever ink (dark or white) has the higher WCAG contrast ratio against the tile, instead of one fixed grey. A percentage on a dark tile – black, navy, dark wine – is now readable, whether the colour came from the default palette or from a palette you set.

Nestimate 0.9.11

Fixes

  • A named state_colors no longer has to match the figure exactly. Names the plot does not draw are dropped with a message() naming them, so one project-wide palette – states from a wider coding scheme, clusters a given call merged away – can be handed to every plot and each takes the keys that apply to it. 0.9.10 raised an error instead, which made a shared palette unusable.

Nestimate 0.9.10

Improvements

  • state_colors accepts a named vector everywhere sequence_plot(), distribution_plot() and plot_state_frequencies() draw, and the names are a lookup rather than a positional list: only the keys you name are overridden, every other key keeps its default, and the vector may be shorter than the number of states. An unnamed vector is still positional, unchanged.

  • For an mcml, that lookup covers the whole multichannel figure, not only the states. A cluster name colours its Summary band, its channel strip and its faded band in the other panels; a group merged by combine = is named by its label (the list name, or "A + B"); rest_label is a key too. So sequence_plot(fit, combine = list(Task = c("Cognitive", "Regulation")), state_colors = c(Task = "#0072B2")) recolours the combined cluster and leaves the rest of the palette alone. Previously the cluster and combined keys were drawn from a fixed internal palette with no way to set them.

  • A state_colors name that matches no key in the figure is now an error naming the available keys, instead of being silently ignored.

Nestimate 0.9.9

New verbs

  • build_mcml() gains combine = and expand =. On new input they change the partition before estimation, in any clusters form and for every input type: build_mcml(data, clusters = cl, combine = c("A", "B")) equals a build with A and B merged in cl (the merged cluster lists its states in cluster-name order). On an existing mcml they re-partition it: build_mcml(mc, combine =) merges clusters, build_mcml(mc, expand =) splits clusters into one cluster per state, and build_mcml(mc, clusters =) applies a new partition. The model (macro network, within-cluster networks, sequences) is re-estimated from the sequences the mcml carries, with its original type, method and directed. With the partition unchanged the result equals the input; expand = "all" reproduces the node-level transition network of the same type. Raises nestimate_mcml_no_sequences for an mcml built from a matrix or an edge list (which keeps only within-cluster edges), and errors when sequence-shaping arguments (trim, exclude, end, labels, actor, …) are passed with a re-partition, since the carried sequences already reflect them.
  • session_ids() names the session behind every sequence of a network built from long data, and of a build_mmm() or build_clusters() fit on such a network. It returns one row per sequence in model order: sequence, the actor and session columns under their own names, session_label, and, for a fit, cluster (plus posterior for a mixture). The fit’s assignments can then be joined to the input by those columns instead of parsing the "actor | session" label, which breaks when an id contains " | ". Raises nestimate_no_session_ids for wide-data input or for fits made before this version, and nestimate_session_ids_misaligned when the metadata and the sequences differ in number.
  • item_loadings() returns the tidy item-diagnostic table of a build_mcml_pc() fit (node, cluster, loading, weight, sign, max_cross, cross_cluster, misfit); misfit = TRUE/FALSE filters it.
  • composites() returns the per-respondent cluster scores of a build_mcml_pc() fit: one row per input row (input order and row names, NA where all of a cluster’s items are missing) and one column per cluster. Raises nestimate_no_composites for the descriptive aggregations ("average", "escoufier", "cancor"), which form no score.

Changes

  • sequence_plot() on an mcml gains combine =: named clusters are merged into one channel (a character vector for one group, a list for several; list names label the merged channels, default "A + B"). The merged group acts as one cluster across the figure (one panel, one Summary key, one faded band) and can itself be opened with expand =.

  • sequence_plot() on an mcml gains rest = c("clusters", "pooled", "none"): how a cluster’s panel shows the time spent in other clusters (one faded band per cluster, one pooled grey band, or blank, leaving only the panel’s own states). Applies to the carpet and distribution views. rest_label = (default "Other states") sets the legend text: the pooled band takes it as is, per-cluster bands read "Social (Other states)". This replaces the former "(elsewhere)" wording.

  • sequence_plot() on an mcml now honours na = in the distribution view: na = FALSE drops the NA (ended) band and shows each time point as shares of the sequences still running, as distribution_plot() already did.

  • macro_network() accepts an mcml_pc fit and returns its cluster-level network; expand = on an mcml_pc raises nestimate_no_expand, and method = or ... error (the estimator is set in build_mcml_pc()).

  • sequence_plot() errors when combine, expand, rest or rest_label is passed for input that is not an mcml, instead of ignoring them.

  • print.mcml_pc() and its build-time warnings name item_loadings() instead of pointing at $loadings.

  • prepare() (and so build_network() on long data) keeps the session column(s) in the per-sequence metadata under their own names, and returns the metadata explicitly in sequence row order (it was assembled with merge(sort = FALSE), whose order is unspecified).

  • build_mmm() and build_clusters() keep the input network’s $metadata (restricted to the fitted rows when build_mmm() drops sequences with missing covariates).

Fixes

  • sequence_plot() on an mcml with type = "heatmap"/"index" and expand = drew the other-cluster wash as blank cells: the wash was keyed by the Summary keys, which are states once a cluster is expanded. It is now keyed by cluster.
  • sequence_plot() on an mcml with type = "distribution" and expand = no longer fails with “subscript out of bounds” in the default (normalize = FALSE) view. The Summary band now opens the expanded cluster into its states, and the other panels draw it as one faded "<cluster> (<rest_label>)" band.
  • sequence_plot() on an mcml with a single channel (one cluster, or every cluster merged by combine) no longer fails in the carpet view with “replacement has 1 row, data has 0”.
  • sequence_plot() on an mcml: with expand =, each cluster’s faded band in the other panels now has its own colour (expanded clusters all shared one).

Nestimate 0.9.5

New verbs

  • macro_network() returns the macro (cluster-level) network of an mcml with one or more named clusters expanded back into their member states and every other cluster left collapsed — a network at mixed resolution. It re-counts from the recoded sequence data rather than splitting the k x k aggregate, which cannot be disaggregated; a matrix-derived mcml raises nestimate_no_expand_source. $node_groups maps each expanded state back to its parent cluster, so the result plots grouped.
  • extract_pathways() cuts a long event log into pathways and returns one row per pathway. Three cuts via type =: "unit" (one pathway per group), "segments" (one per contiguous run), "anchored" (spans around an anchor event). resolve = appends a resolution label as the closing state.
  • outcome_model() regresses a unit-level outcome on sequence or network predictors — pattern indicators, simplicial_features() output, or any numeric covariate — with family auto-detection, an optional lme4 random intercept, select = "split" hold-out selection, BH-corrected p-values, confidence intervals and odds ratios. effects_table() is the tidy accessor; summary() returns the same table.
  • simplicial_features() returns topological summaries of one or many networks as a tidy data.frame, one row per network per threshold, ready to use as regression predictors. Accepts a netobject, netobject_group, mcml, a square weight matrix, or a named list of any of these.

Extended

  • build_mcml() gains the exclude, trim, end and end_by sequence arguments, applied in that fixed order.
  • as_tna() is now a generic. as_tna.mcml(expand =) delegates to macro_network() for the macro layer only; per-cluster layers are untouched. An as_tna.default() covers everything else.
  • sequence_plot() and its mcml method gain panel = c("both", "summary", "channels").

Fixes

  • frequencies(format = "frequency") returned zero rows for any long-format data whose id column was not integer-valued. The sequence key was coerced with as.integer() before the merge, turning every character id into NA; the merge then matched nothing and failed silently. Sequence keys are now aligned in their character form.
  • sequence_plot() on an mcml errored whenever a singleton cluster was named after its own state: the shared fill scale built its levels with factor(levels = c(states, clusters)), and duplicate levels are an error. Levels and values are now de-duplicated.
  • as.data.frame() methods for net_hypergraph_transduction and net_hypergraph_cluster placed what in the generic’s row.names slot, so as.data.frame(x, "scores") bound "scores" to row.names. The generic’s arguments now come first and row.names is honoured (R CMD check “S3 generic/method consistency”).

Nestimate 0.9.4

  • pkgdown reference index repaired: the five topics added in 0.9.3 were unindexed and build_reference_index() failed. The S3 Methods section now wildcards as.data.frame. alongside print., summary. and plot., so a future result class with the house-standard accessor cannot repeat it.

Nestimate 0.9.3

New verb: compare_networks()

  • Compares two or more networks (netobject, netobject_group, tna, group_tna, mcml, cograph_network, matrices, or a list of them) in one object. summary() returns the tidy one-row-per-pair overview, as every other Nestimate summary() does; the full tables are their own verbs: edge_differences(), node_differences(), global_differences(), network_metrics(), each with a pair column. All pairs by default, or every network against one reference =.
  • Ratios are guarded (NA, never Inf/NaN); every cell is kept, absent edges appear at weight 0; higher (colour) and status (shape) columns drive the plots.
  • plot(x) draws one view per call, selected with type = (the package-wide argument name; what = is accepted as an alias): "networks" (default; each network drawn once via cograph::splot()), "difference" (the signed difference network of each pair), "edges" (ranked dumbbell), "nodes", "global", "heatmap", "scatter", and "inference" (a forest of the edge differences on the difference scale, with credible intervals when the Bayesian backend ran, a filled marker for a significant edge and the p-value in a right-hand column; it raises nestimate_compare_no_test when test = "none"). One sign-to-colour contract (#4A6FE3 = first/reference higher, #D33F6A = second higher) backed by line type and marker shape; the heatmap scale follows the data.
  • plot(x, combined = FALSE) splits a multi-pair view into a named list of single-pair plots (one per pair) instead of facetting them into one figure, matching the combined argument of plot.net_reliability(), plot.simplicial_complex() and friends. The base-graphics views draw one panel per page.
  • Optional inference on the same tables: test = "permutation", "bayes" (with an optional rope), "bootstrap", combinable. Non-significant results are faded, never deleted.
  • permutation() gains a $global data frame with NCT-style M (sum of absolute differences) and S (largest absolute difference) statistics and their permutation p-values, computed from the same null.
  • compare_model() is unchanged and will be soft-deprecated once compare_networks() has been through one release.

Hypergraph suite

Nestimate 0.9.1

tna-parity release: sequence-side gaps against tna::build_model() and tna::prepare_data() closed.

New estimators

  • build_network(method = "ngram") (params n_gram, default 2), "gap" (params max_gap, default 1) and "reverse" (params weighted) mirror the tna types "n-gram", "gap" and "reverse". Weights are numerically identical to tna on shared inputs (tested). All three accept wide or long input, start/end boundary markers, group = dispatch and scaling = "normalize" for row probabilities.
  • New aliases: "n-gram" / "n_gram" -> "ngram", "co-occurrence" -> "co_occurrence".

Timezone-safe timestamp parsing

  • prepare() and build_network() gain timezone = "UTC" (Olson name). Naive timestamps are interpreted in that zone; ISO-8601 Z/UTC/GMT markers and numeric offsets (+0200, +02:00) are converted from their offset. Parsing no longer depends on the machine’s local time zone.
  • Fixes: ...Z timestamps parsed to NA (the marker was stripped before matching), offsets were silently ignored, and wall-clock readings inside a DST gap could turn into spurious session boundaries. Unparsable values now raise an error instead of becoming session breaks.

Nestimate 0.9.0

Delegation release: Nestimate stops owning psychometric-network math and delegates it to its two clean-room home packages, psychnets (cross-sectional) and idiographic (temporal). No public API changes: every signature, default, return shape, and — verified against frozen pre-delegation baselines — every number is preserved.

Delegated to psychnets (new hard dependency, >= 0.5.2)

  • The cor, pcor, and glasso estimators keep Nestimate’s input layer, validation, and return contract; the math now runs in psychnets::cor_network() / pcor_network() / ebic_glasso(). Verified field-complete against the frozen baseline: cor exact, pcor within 5.6e-17, glasso exact on weights, precision, selected lambda, and the EBIC path (including penalize.diagonal and refit branches).
  • nct() delegates its inner EBIC-glasso solve; the NCT-specific nearPD symmetrization stays local. Seeded runs: networks and p-values exact.
  • boot_glasso() and permutation() (glasso branch) delegate the per-resample solve via psychnets::ebic_glasso(lambda_path = ), keeping the fixed-path-across-resamples semantics. Seeded permutation() is byte-identical; boot_glasso() is identical except wall-clock timing.
  • The internal pure-R glasso kernel (glasso_pure.R) and its EBIC helpers are deleted; psychnets owns that math now.
  • The ising and mgm estimators remain local for now (their delegation needs psychnets to expose per-node lambda selection and a scale passthrough) and are unchanged.

Delegated to idiographic (new hard dependency, >= 0.3.4)

  • build_mlvar() keeps its signature, S3 methods, and return object; the lmer estimation pipeline now runs in idiographic::fit_mlvar(). Verified identical at tolerance 0 (object, class, print, summary) across lag/standardize/day/beep configurations, and against mlVAR::mlVAR() at 8.8e-16 over 954 checks.
  • When the between-subjects network is not estimable (a random-intercept SD of zero), build_mlvar() now warns before returning the zero matrix; it previously returned it silently. Numbers are unchanged.
  • build_gimme() delegates its whole search to idiographic::fit_gimme(). This is the one delegation that changes results: idiographic’s search reproduces the upstream gimme package (>= 10.0) exactly, which Nestimate’s own search did not — the individual-level path search in particular under-detected person-specific paths relative to upstream. Group-level results are typically unchanged; individual-level path sets can grow. The signature and the net_gimme field contract are unchanged (the object gains idiographic’s netobject fields and now renders directly with cograph); print/summary/plot dispatch to idiographic’s methods. Treat pre-0.9.0 build_gimme() individual-level results as superseded.

Nestimate 0.8.5

CRAN release: 2026-08-21

Bug-fix release. 0.8.4 was published on r-universe but never reached CRAN.

Plotting in non-UTF-8 locales

  • plot() on net_entropy_bayes and net_sequence_comparison results no longer fails where the graphics device cannot represent the arrow glyph. Edge labels and the comparison subtitle previously used U+2192/U+2190, which a non-UTF-8 locale cannot convert; the graphics engine raised an error (conversion failure ... in 'mbcsToSbcs') rather than substituting a character, so the plot could not be drawn at all. Both now use the ASCII arrow notation (A -> B) already used by the higher-order network verbs. Console output is unaffected.

  • plot() on net_entropy_bayes results no longer emits a deprecation warning under ggplot2 4.0, which removed geom_errorbarh().

Rendering entropy networks with cograph

  • entropy_network() now renders with its intended transition-network styling on cograph 2.4.4, the current CRAN version. The object states its full style through cograph’s meta$splot producer contract instead of relying on cograph recognising the "entropy" method name, so no cograph update is required. Output is identical under cograph 2.4.4 and 2.4.5.

Documentation

Nestimate 0.8.4

Transition matrix entropy

  • New entropy suite around the entropy rate of a transition matrix (Shannon 1948; Cover & Thomas 2006, ch. 4; the transition entropy of Krejtz et al. 2015 and the real-time mobile transition matrix entropy of Krejtz et al. 2025):
    • transition_entropy() — entropy rate H (stationary distribution from the eigendecomposition at lambda = 1), stationary entropy, redundancy, per-state branching entropies, all with the normalized (scale-free) variants; print/summary/plot and netobject_group dispatch.
    • entropy_network() — the exact edge-level decomposition of H: each edge carries its term pi_i P_ij log(1/P_ij); weights sum to H. Displays the summands of the entropy-rate equation — no new quantity is estimated. scaling = "share" relabels edges as percentages of H (sum = 100). Additional weights: "surprisal" (optionally scaling = "chance") and "production" (irreversibility). The result is a regular netobject that inherits its source network’s styling and ships an entropy house style via cograph’s meta$splot producer contract, which states the styling outright rather than relying on cograph recognising the method name (cograph >= 2.4.4).
    • entropy_trajectory() — sliding-window entropy over the transition stream (the windowed design of Krejtz et al. 2025): tidy per-window table, per-group trajectories, loess-trend plot; the per-window estimator weights rows by observed occupancy (robust to non-ergodic window fragments).
    • entropy_bayes() — Dirichlet-posterior estimation: credible intervals for H, per-state entropies, and per-edge contributions; edges flagged credible when their share of H credibly exceeds min_share; $model holds the pruned stable entropy network.
  • New vignette: Transition Matrix Entropy.

Mixed Markov clustering contract

  • cluster_mmm() now returns the fitted net_mmm clustering object, retaining assignments, posterior probabilities, mixing proportions, fit criteria, and fitted component models. Network materialization remains the responsibility of build_network(fit) or the one-step cluster_network(..., cluster_by = "mmm") workflow.
  • as_htna() gains a net_mmm method. An MMM fit created from an HTNA model is materialized into an htna_group without rerunning the MMM fit, while the original actor partition and clustering diagnostics are preserved.

Nestimate 0.8.3

HTNA expansion

  • as_htna() still rebuilds one full node-level network from the original source, preserving every between-cluster transition, and now completes the canonical HTNA contract. Its result inherits from htna, netobject, and cograph_network; stores character actor labels in $node_groups$group, a factor in $nodes$groups, and actor order in $actor_levels; and retains actor-order metadata on $node_groups for lossless partition round trips, plus the legacy $nodes$cluster and "cluster_members" metadata.

Nestimate 0.8.2

Session grouping

  • prepare() now identifies sessions from the observed combinations of the actor and session columns instead of base::interaction(). Three defects are fixed:

    • Integer overflow. interaction() codes a combination over the marginal level space, which exceeds .Machine$integer.max once both columns pass 46,341 distinct values. The resulting NAs were pasted into the literal string "NA", merging unrelated events into one pseudo-session. On 50,000 actor-session pairs this silently discarded 14% of sessions and manufactured transitions that no input sequence contained, including self-loops on terminal states.
    • Separator collisions. Identifiers were pasted with " | " before being used as a grouping and metadata merge key, so ("a | b", "c") and ("a", "b | c") collapsed into a single session. Grouping now keys on the original columns; the readable label is display-only and is exposed as .session_label in meta_data.
    • Missing identifiers. Missing values in a grouping column silently changed the session count. They now raise an error naming the columns.

    Group numbering reproduces interaction()’s ordering, so prepared row order and finite same-seed bootstrap results are unchanged.

  • time_threshold = FALSE switches session-interval splitting off, so each actor (or actor-session) forms a single sequence regardless of gap length. Accepted by prepare(), build_network() and build_mcml().

Testing

  • Added a randomized grouping sweep over 1,000 seeded datasets covering actor/session column counts, identifier vocabularies including separator and UTF-8 cases, time on and off, tied timestamps, explicit order columns, shuffled input, and time-gap structures that split, do not split, or sit exactly on the threshold. Verified against ground truth and against interaction() for row-order compatibility.

Nestimate 0.8.1

HTNA interoperability

  • Distance clustering and mixed-Markov clustering now preserve HTNA inputs. build_clusters() and cluster_mmm() carry the node-to-actor partition into network materialization, while cluster_network() returns an htna_group directly. Every child remains an htna object with $node_groups, $nodes$groups, and $actor_levels; clustering assignments, posterior probabilities, fit diagnostics, and other outer attributes remain attached.

  • Added extensive randomized equivalence coverage across distance clustering, mixed-Markov clustering, all transition-network estimators, actor-absent clusters, and HTNA-versus-plain input paths.

Nestimate 0.8.0

CRAN release: 2026-07-10

Documentation

  • The Bayesian verbs get their own reference section, placed directly after Network Estimation: certainty(), bayes_compare(), subtract_networks() and as_netdifference(). They were previously buried in a fourteen-entry “Bootstrap & Inference” list. bootstrap_network() now points at certainty() as its closed-form counterpart, and permutation() points at bayes_compare() as its Bayesian complement, so each pair is reachable from either side.

  • frequencies() is no longer marked \keyword{internal}. The topic page and the exported function share a roxygen topic name, so the keyword from the topic block leaked onto the function’s own help page even though the function is exported (and called by the package). cluster_data() keeps its internal keyword: it is a deprecated alias for build_clusters() and is meant to stay out of the index.

  • Dropped the utils help page, which documented no exported object. The @importFrom directives it carried are retained.

  • audit_codex/ is no longer tracked; it holds generated audit artifacts.

Dependencies

  • Suggests: cograph (>= 2.4.4). The netdifference verbs added in 0.7.8 need cograph 2.4.x: CRAN’s cograph 2.3.6 contains no netdifference support, so cograph::plot_difference() does not exist there and cograph::splot() on a netdifference falls through to the plain netobject renderer and silently draws an unsigned network. Nestimate must not be submitted to CRAN before cograph 2.4.4 is available there.

Nestimate 0.7.8

New features

  • subtract_networks() / as_netdifference() — verbs for the difference between two networks. subtract_networks(x, y) returns the edge-wise difference as a netdifference object; as_netdifference() promotes an existing comparison result to the same class — a bayes_compare() result, or a netdifference, which passes through; anything else errors — so a difference computed by any route prints the same way. Adds print.netdifference.

  • bayes_compare() accepts two net_edge_betweenness() objects (source method "relative" only). Edge betweenness is recomputed on every posterior draw, giving the Bayesian analogue of permutation()’s edge-betweenness dispatch, with posterior mean betweenness matrices and the plug-in observed_diff.

  • permutation() gains a measures argument for centrality permutation tests, matching the tna package’s dispatch.

Enhancements

  • bayes_compare()’s probability-of-direction column is renamed pd -> p_difference in the summary() frame, and the result now carries class c("net_bayes", "netdifference", "net_permutation") so it dispatches to the difference verbs as well as the permutation ones.

  • Non-ASCII characters normalized across R sources and man pages.

Bug fixes

  • centrality_stability() no longer errors with “missing value where TRUE/FALSE needed” when a requested measure is undefined on the network (e.g. Diffusion is NaN on a small cyclic net): sd() returned NA, which poisoned if (!any(keep)). Such measures now drop like zero-variance ones.

  • centrality_stability()’s default measures is restored to c("InStrength", "OutStrength", "Betweenness"). 0.7.7 had swapped OutStrength for Diffusion, which broke the package: it calls centrality_stability() with no measures and compares the result against its own explicit trio. centrality() / net_centrality() keep the Diffusion default; only centrality_stability() reverts.

  • Suggests: cograph relaxed from (>= 2.4.4) to (>= 2.3.6), the version available on CRAN. Additional_repositories removed — every declared dependency now resolves from CRAN.

Nestimate 0.7.7

Enhancements

  • plot() on a net_centrality_group gains type = "delta", showing the between-group difference per measure, and now supports three or more groups. Zero-valued edges can be blanked with drop_zero = TRUE.

Nestimate 0.7.6

New features

Enhancements

  • plot() on centrality results gains alternative views: type = c("bar", "line", "heatmap") for a single net_centrality, and type = c("bar", "line", "delta") for a net_centrality_group. Count-like measures get integer axis labels.

Nestimate 0.7.5

Version bump only; no user-visible changes.

Nestimate 0.7.4

New features

Enhancements

  • Centrality gains the tna-parity measures. net_centrality(x, measures = "all") now returns OutStrength, InStrength, ClosenessIn, ClosenessOut, Closeness, Betweenness, BetweennessRSP, Diffusion and Clustering — previously only the strengths, Closeness and Betweenness. Adds plot.net_centrality and plot.net_centrality_group.

  • sequence_plot() and the MCML plots gain layout refinements.

Nestimate 0.7.3

New features

  • as_netobject() / validate_netobject() — the boundary layer between (which owns the psychometric-network math and emits a lean cograph_network) and Nestimate (which owns the canonical netobject schema). as_netobject() promotes a psychnet result or a bare cograph_network to the dual-class c("netobject", "cograph_network") so it dispatches to every Nestimate verb, parking psychnet-specific fields (including the GLASSO KKT certificate) under $meta$psychnet; netobjects pass through unchanged. validate_netobject() enforces the shared structural contract so schema drift on either side fails loudly. psychnet is not a declared dependency — Nestimate never calls it; the converter works by S3 dispatch on whatever psychnet object the caller supplies.

  • certainty() — analytic Bayesian counterpart of bootstrap_network() for transition networks. Models each state’s outgoing transitions as a Dirichlet-Multinomial process (Jeffreys prior) and returns posterior mean, sd, credible interval and a stability decision per edge in closed form (no resampling). Returns the exact net_bootstrap object layout and carries class c("net_certainty", "net_bootstrap"), so it is a drop-in: every net_bootstrap method works on it. Completes the assessment trio certainty / stability (bootstrap_network) / reliability (reliability).

Enhancements

  • sequence_plot() gains a multichannel view for mcml objects built from sequences. sequence_plot(fit) draws one carpet panel per cluster channel plus a macro Summary panel — each channel’s own states solid, the other clusters a faded wash, finished cells white, rows aligned by the macro sequence. sequence_plot(fit, type = "distribution") stacks the prevalence (own states + faded other clusters + an explicit NA band, to 100%), and normalize = TRUE gives a TraMineR-style seqdplot where each time point sums to 1. ggplot-based and dependency-free; returns a ggplot object.

  • bayes_compare() results are now 100% compatible with the permutation() format: the object carries class c("net_bayes", "net_permutation") with all net_permutation slots (diff_sig, p_values, effect_size, iter, alpha, paired, adjust), and its summary is a superset of summary.net_permutation (from, to, weight_x, weight_y, diff, effect_size, p_value, sig plus the Bayesian extras count_x, count_y, ci_lower, ci_upper, ci_width, pd). A bayes_compare() result is now a drop-in wherever a net_permutation is consumed.

Nestimate 0.7.2

New features

  • bayes_compare() — Bayesian Dirichlet-Multinomial comparison of two transition networks, a complement to permutation(). Models each source state’s outgoing transitions as a Dirichlet-Multinomial process (Jeffreys prior) and returns, per edge, a posterior mean difference, a credible interval, the probability of direction (pd) and its two-sided p-equivalent. Adds print/summary/plot methods and netobject_group dispatch (all-pairwise or matched). Method source: Johnston & Jendoubi (2026), How Delivery Mode Reshapes Resource Engagement: A Bayesian Differential Network Analysis, TNA Workshop 2026.

Nestimate 0.7.1

New features

  • as_networks() — promote a build_mcml_pc() result into a netobject_group (the psychometric-network counterpart of as_tna()). Singleton clusters with no within-network are dropped with a warning; an existing netobject_group passes through unchanged.

Documentation

  • Vignettes and articles now call package verbs directly instead of hand-assembled base-R subsetting rituals: markov_order_test() reads sequences straight from a fitted network (markov_order_test(net)); HYPA anomaly tables use summary(hypa, order_by = "ratio"); higher-order pathways use pathways(hon, top = ); grouped-clustering inspection uses cluster_diagnostics().

Nestimate 0.7.0

New features (experimental)

  • build_mcml_pc() — MCML aggregation for psychometric networks (cor / pcor / EBICglasso). Five aggregation methods with explicitly different statuses: "average" (descriptive block-mean; works without raw data), "composite" (cluster scores re-estimated with the chosen estimator — a genuine cluster-level network), "loadings" (composites weighted by mean within-cluster connection strength — Nestimate’s own weighting, not an EGA reimplementation), "rv" (Escoufier’s RV matrix correlation between blocks), and "canonical" (first canonical correlation — the upper bound for composite methods). Within-cluster networks re-estimated by default (within = "reestimate") since a pcor submatrix is not the subsystem’s pcor network. Item diagnostics in $loadings: signed loadings (reverse-keyed items detected via the leading eigenvector of the within-block matrix and flipped in composites), cross-cluster strengths, and a misfit flag when an item is more connected to another cluster than its own (warned). Composites tolerate missing data (row-wise renormalized weighted means); Composite item weights are selectable via weighting — ten built-in schemes spanning three views of the cluster: the scale as scored ("equal", "item_total"), the network’s view ("strength", "eigen", "closeness", "betweenness", "expected_influence", "specificity" — the misfit margin as a weighting, zeroing items that belong as much to another cluster), and the latent-variable view ("pca", "factor"); plus fully custom weighting via a named numeric vector or a function(W_block, data_block, nodes). aggregation = "loadings" is the alias for composite + strength. The "factor" weighting exposes its extraction method via fa_method: "ml" (factanal), "paf" (iterated principal axis), "minres" (ULS), or "cfa" (one-factor lavaan model; with cor_method = "polychoric" the categorical DWLS factor model) — all operating on the cor_method-consistent correlation structure. Reverse-keyed handling works under every sign-carrying scheme (item-total correlations are computed on eigen-sign-pre-oriented columns so a reversed member cannot contaminate small clusters). cor_method = "polychoric" (via lavaan) supports ordinal items; id_col drops identifier columns so convert_sequence_format(format = "frequency") actor-profiles feed the function directly (the within-person co-occurrence view of event data). Returns class mcml_pc (macro + within netobjects, all undirected) with print/summary/plot; the composite/loadings macro is a full netobject, so bootstrap_network(), vertex_bootstrap(), and vertex_compare() apply to it directly. cograph::plot_mcml() (>= 2.3.8) renders the two-layer undirected MCML view. Experimental: API and formulas may change.
  • loading_stability() — case-bootstrap stability of the build_mcml_pc() composite weights (percentile CIs, sign-flip rates), with print and forest-style plot.

Nestimate 0.6.5

New features

  • vertex_bootstrap() — Snijders & Borgatti (1999) vertex bootstrap for network-level statistics (density, mean weight, strength centralization, weighted reciprocity, plus custom statistic_fn). Needs only the weight matrix, so it works on data-less netobjects (build_mlvar() constituents, as_tna(mcml) elements, plain matrices) where bootstrap_network() cannot run. Returns a tidy one-row-per-statistic net_vertex_bootstrap with print/summary/plot. Self-loops are preserved (diagonal carries the resampled vertex’s own self-weight); undirected replicates stay symmetric.
  • vertex_compare() — the Snijders & Borgatti two-network test the vertex bootstrap was originally proposed for: z-tests and normal-approximation CIs for differences in network-level statistics between two networks (netobjects, matrices, or precomputed net_vertex_bootstrap objects). Tidy net_vertex_comparison result with print/summary/plot (forest plot of differences).
  • bootstrap_network() and vertex_bootstrap() gain ci_method = c("percentile", "basic"): basic intervals (Davison & Hinkley 1997, eq. 5.6) reflect the percentile bounds around the observed estimate, correcting first-order bootstrap bias. Default remains "percentile".

Nestimate 0.6.4

Bug fixes

  • build_mcml() (sequence and edge-list paths) now records the effective directedness in $meta$directed: FALSE when type = "cooccurrence", whose weights are symmetrized, instead of echoing the directed argument unchanged. Renderers that auto-detect directedness (e.g., cograph::plot_mcml() with directed = NULL) now draw co-occurrence MCML objects as undirected networks automatically.

Nestimate 0.6.3

New features

  • mosaic_analysis(data, var1, var2) — two-variable mosaic analysis on a data.frame: chi-square or Fisher test, Cramer’s V (df-adjusted effect size) and a flat mosaic plot. Returns class mosaic_analysis with a tidy one-row-per-cell $counts, a one-row $stats, and print/summary/plot. Distinct from mosaic_plot(), which draws from a fitted network object.

Enhancements

  • mosaic_plot() gains style = c("classic", "flat"). The flat style uses variable-width columns, white gutters and in-tile or side labels, sharing the classic style’s geometry and diverging palette; values = TRUE prints residuals inside the tiles.

Nestimate 0.6.2

Bug fixes

  • Corrected stale rank-scaling assertions in the test suite. No user-visible change.

Nestimate 0.6.1

New features

  • build_network() and the transition wrappers (build_tna(), build_ftna(), build_atna(), build_cna()) gain start and end boundary markers: FALSE (default), TRUE (labels "Start" / "End") or a custom string. start prepends a source state to every sequence; end places a sink in the single cell after each sequence’s last non-NA state (not absorbing — see mark_terminal_state() for that). Honoured by the relative, frequency, co_occurrence and attention estimators; other methods error.

  • build_mmm() gains covariate_effect. "em" (default) folds covariates into the EM as covariate-dependent mixing, changing the fit; "posthoc" fits a plain mixture and uses covariates only for the after-fit multinomial logit, leaving the clustering bit-identical to a no-covariate fit.

Nestimate 0.6.0

CRAN release: 2026-05-31

New features

Documentation

Packaging

  • Removed the Remotes: field; cograph and tna are available from CRAN, so no non-CRAN source pin is needed.

Nestimate 0.5.1

Audit follow-up (clustering + MCML)

Followed codex_docs/audit_clustering and codex_docs/audit_mcml recommendations across two modules. Eleven of thirteen findings addressed; two deferred pending design decisions on numeric semantics (directed = FALSE raw-data MCML, MMM first-non-NA initial state).

Bug fixes

  • cluster_network() now forwards distance-clustering arguments (na_syms, weighted, lambda, seed, q, p, covariates) to build_clusters() instead of silently passing them to build_network(). The split runs on caller ... only — netobject build_args continue to flow only to the build_network() step, protecting attention-method (atna) network history from being re-routed to weighted Hamming. (audit_clustering item 1)
  • .auto_detect_clusters() (used by build_mcml() and cluster_summary()) now requires node_groups to carry a node identifier column when shaped as a data.frame, or be a named atomic vector keyed by node label. Previously, a bare cluster-only data.frame was read positionally — silently mis-assigning nodes whenever node_groups rows were in a different order than x$nodes. (audit_mcml item 1)
  • build_clusters() now rejects all-missing input early with a clear message instead of failing indirectly downstream in pam/hclust. (audit_clustering item 4)

New parameters

  • compare_mmm(return_fits = FALSE) — when TRUE, the fitted net_mmm models are attached as attr(result, "fits") keyed by k, so users can pick the chosen model without re-running EM. Default behaviour unchanged. (audit_clustering item 6)

Improvements

  • build_clusters() validation messages now name the offending argument ("'k' must be at least 2 (got k = 1)") rather than dumping the failing predicate. Top-level type checks switched to named-condition stopifnot() for the same reason. (audit_clustering item 2)

Documentation

  • summary.mcml() roxygen corrected — was claiming a printing side effect that doesn’t exist. (audit_mcml item 5)
  • build_mcml() clusters = "<col>" mode now documents its narrow contract: assigns each row’s group label to both endpoints, so it only makes sense for within-group edge lists. (audit_mcml item 2)
  • build_mcml() method parameter doc now steers raw sequence / event-log inputs to "sum", since the function counts observed transitions. Other methods are for weighted edge lists or pre-existing matrices. (audit_mcml item 4)
  • as_tna.mcml() “Excluded Clusters” section corrected — drop emits a warning() (was claimed silent) and only fires for relative method (was claimed unconditional). (audit_mcml item 6)
  • build_clusters() na_syms doc adds an explicit “Missing-value distance rule” subsection: NA becomes a comparable sentinel state, not pairwise deletion. (audit_clustering item 3)
  • build_mmm() adds an “Initial states” section explaining first-column-verbatim init and that build_mmm does NOT honor build_clusters-style na_syms — only actual NA cells become NA inits. (audit_clustering item 5, doc-only path)

Tests

  • +12 new tests pinning the corrected contracts and the documented edge cases (misordered node_groups alignment, label propagation through state_distribution(), as_tna.mcml() drop-warning fixture, MMM first-column NA behaviour, and the four-way cluster_network() arg-routing contract). Full sweep: 1628 / 1628 pass, 0 fail.

Nestimate 0.5.0

Bug fixes

  • .extract_edges_from_matrix() no longer drops the diagonal. Netobjects built via .wrap_netobject() (and therefore everything from build_network(), build_mcml(), bootstrap_network(), build_mmm(), wtna(), as_tna()) now have $edges containing every non-zero matrix entry, including self-loops. Previously $weights and $edges were silently inconsistent on any matrix with a non-zero diagonal, causing downstream consumers (e.g. cograph::centrality() on an MCML macro) to under-count node degree by 2.

New features

  • plot_state_frequencies() — native S3 generic for state-frequency plots across netobject, netobject_group, mcml, and htna. Defaults to a marimekko (mosaic) layout where column widths reflect per-group totals and segment heights reflect within-group state proportions; also supports a colored-bars style and a per-group faceted marimekko. Uses the package Okabe-Ito palette throughout.
  • plot_mosaic() — exported low-level marimekko primitive built on geom_rect() with cumulative-width / cumulative-height geometry. Reusable for any tidy data.frame(group, state, weight) input.

Nestimate 0.4.4

Bug fixes

  • passage_time() and markov_stability() now raise an explicit error naming the dead state when a transition-matrix row sums to zero, instead of silently propagating NaN through eigen/solve. Zero rows mean the chain is not ergodic; mean first passage times are undefined. Shared helper .mpt_normalize_rows() factored out of both entry points.
  • .prepare_association_input() no longer hard-rejects non-square numeric matrices. For association methods (glasso, pcor, cor) the netobject’s $data slot is a numeric matrix (not a data.frame). Any downstream caller that row-subsetted $data and re-invoked the estimator (centrality_stability(), bootstrap_network(), reliability()) was silently producing NULL centralities caught by tryCatch, which surfaced as an “all centrality measures have zero variance” warning or all-NaN correlations. The matrix branch now recognises non-square input as raw observation data and recursively re-enters through the data-frame branch. Square symmetric matrices (pre-computed correlation / covariance) still go through the symmetric-matrix path with the symmetry check intact.

New parameters

  • build_network() gains state_cols and metadata_cols parameters (both default NULL). Explicit overrides for the state-vs-metadata column classifier, which previously used a “values-in-nodes” heuristic that silently misclassifies metadata columns whose values coincide with node labels (e.g. a condition column with levels "A","B","C" when nodes are "A","B","C"). Validation: error on overlap between the two vectors, error on column names not present in the input data. Forwarded through the group = ... recursive dispatch so per-group calls honour the override.

Removed

  • plot.net_link_prediction() and plot.mcml() removed. Nestimate is a computation engine — visualization is the user’s concern. Previously both methods called cograph:: directly, violating the stated dependency invariant (Nestimate -> cograph direction forbidden). Users call cograph::splot(net) or cograph::plot_mcml(fit) directly.

Documentation

  • wtna() @param type now flags that type = "relative" combined with method = "cooccurrence" produces an asymmetric matrix (conditional co-occurrence given row state), not a symmetric undirected weight matrix. Use type = "frequency" if symmetric counts are required.

Testing infrastructure

  • New numerical-equivalence tests (gated by NESTIMATE_EQUIV_TESTS=true): test-equiv-permutation.R (vs. stats::p.adjust + hand-coded base-R permutation loop), test-equiv-mlvar.R (vs. mlVAR::mlVAR at machine precision), test-equiv-association-rules.R (vs. arules::apriori), test-equiv-link-prediction.R (vs. clean-room matrix algebra + igraph::similarity), test-equiv-centrality-stability.R (vs. bootnet::corStability). Total ~162k per-value comparisons; all within machine precision except centrality-stability which uses a documented drop-grid tolerance because bootnet uses igraph path-based centrality and Nestimate uses Floyd-Warshall.
  • New HON-family equivalence tests under local_testing_and_equivalence/ validating HON, HONEM, HYPA, MOGen, and hypergraph against pathpy 2.2.0 (via reticulate), BiasedUrn, RSpectra, and HyperG. Not shipped in the R-package tests/ directory; added to .Rbuildignore.
  • Branch-matrix coverage added for the four many-mode APIs (wtna, bootstrap_network, build_clusters, sequence_plot) — systematic cross-product tests over all combinations of mode parameters to catch regressions where one branch silently diverges.

Nestimate 0.4.3

CRAN release: 2026-04-20

CRAN resubmission (addresses incoming-check NOTEs on 0.4.2)

  • Tarball now ships build/vignette.rds (the vignette index). Previous 0.4.2 build used R CMD build --no-build-vignettes, which preserved pre-built inst/doc/*.html but stripped the index — CRAN flagged “VignetteBuilder field but no prebuilt vignette index.”
  • test-gimme.R now skip_on_cran(). GIMME tests fit a lavaan SEM per subject and took ~50s locally (2-3× on Windows), pushing total check time to 11 min on win-devel. Full test suite still runs in CI and local dev.

Nestimate 0.4.2

CRAN resubmission

  • Full --as-cran --run-donttest audit pass.
  • Purged stale .Rcheck/ and Meta/ build artifacts from working tree; added explicit ^Nestimate\.Rcheck$ and ^\.\.Rcheck$ entries to .Rbuildignore as belt-and-suspenders against repeat-submission contamination.

Nestimate 0.4.1

CRAN resubmission

  • Pre-built vignettes included in inst/doc/ as required by CRAN.
  • Fixed 301-redirect URLs in README.
  • Added skip_on_cran() to slow test block to keep check time under 10 minutes.

Nestimate 0.4.0

New functions

New datasets

  • human_long, ai_long — canonical long-format human–AI pair programming interaction sequences (10,796 turns, 429 sessions).
  • chatgpt_srl — ChatGPT-generated SRL scale scores for psychological network analysis.
  • trajectories — 138-student engagement trajectory matrix (15 timepoints, 3 states).

API

Bug fixes

  • build_mmm() no longer crashes on platforms where parallel::detectCores() returns NA (macOS ARM64 CRAN check failure).
  • gimme convergence filter now correctly handles all typed NA variants (NA_character_, NA_real_, etc.).
  • NaN values in numeric metadata aggregation (all-NA sessions) normalized to NA_real_.
  • HYPA p-values corrected; hypa_score column renamed to p_value.

CRAN compliance

Performance

Nestimate 0.3.4

  • HYPA: Renamed hypa_score column to p_value for clarity. Added $over, $under, $n_over, $n_under fields to net_hypa objects. Scores are now pre-sorted with anomalous paths first.
  • HYPA: summary.net_hypa() now shows over/under-represented paths separately with a configurable n parameter.
  • pathways.netobject(): New S3 method to extract higher-order pathways directly from a netobject (builds HON or HYPA internally).
  • path_counts(): Now handles NAs in trajectories by stripping them before k-gram counting.

Nestimate 0.2.15

  • Preparing for publication

Nestimate 0.2.0

  • Reduced hard dependencies from 6 to 4 Imports (ggplot2, glasso, data.table, cluster).
  • Removed igraph from Imports — centrality_stability() and boot_glasso() now accept a centrality_fn parameter for external centrality computation.
  • Removed tna from Imports — moved to Suggests (only used for input class detection).
  • Implemented graphical_var() from scratch using coordinate descent lasso + graphical lasso with EBIC model selection, eliminating the graphicalVAR dependency.
  • Dropped ml_graphical_var() — users should use mlvar() for multilevel VAR.
  • Removed cograph plot wrappers — plot.netobject(), plot.net_bootstrap(), plot.net_permutation(), plot.net_hon(), plot.net_hypa() and as_cograph() removed. Users call cograph plotting functions directly on netobjects.
  • Added attention estimator for decay-weighted transition networks.
  • Increased test coverage from 84.5% to 96.1% (2780 tests).
  • Passes R CMD check with 0 errors, 0 warnings, 0 notes.

Nestimate 0.1.0

  • Initial release. Split from Saqrlab v0.3.0.
  • Core estimation via build_network() with 8 built-in estimators.
  • Bootstrap inference (bootstrap_network()), permutation testing (permutation()), EBICglasso bootstrap (boot_glasso()).
  • Higher-order networks: HON, HONEM, HYPA, MOGen.
  • GIMME, MCML, multilevel VAR, graphical VAR.
  • Temporal network analysis and velocity TNA.
  • Dual-class c("netobject", "cograph_network") output for cograph compatibility.