lagseq belongs to the Dynalytics ecosystem alongside tna (transition network analysis), Nestimate (network estimation), cograph (rendering), and TraMineR. The ecosystem separates estimation from rendering and analysis, and the packages share a common data grammar and object schemas. In practice this means two things: lsa() can read sequences straight out of those packages’ objects, and a fitted lsa model converts back into them in one call.

Into lagseq: fit from another package’s object

lsa() recognises sequence-bearing objects from tna (tna, tna_data, tna_seq_data), Nestimate, and TraMineR (stslist). It pulls the sequences out and fits, so no manual extraction is needed.

A tna object

tna and lsa() share the same long-format grammar (actor / action / time / order / session). Prepare the data with tna, build a model, and hand either object to lsa().

log <- tna::group_regulation_long
prepared  <- tna::prepare_data(log, actor = "Actor",
                               action = "Action", time = "Time")
tna_model <- tna::tna(prepared)

lsa(prepared)     # fit from the prepared sequence object
#> Lag Sequential Analysis  —  classical  (lag 1, directed)
#>   9 states | 25533 transitions | 27533 events | 2000 sequences
#>   states: adapt, cohesion, consensus, coregulate, discuss, emotion, monitor, plan, synthesis
#>   independence: G² = 13203.8, df = 64, p <2e-16
#> 
#>   Significant transitions (p < 0.05): 72 of 81
#>   strongest over-represented (of 23):
#>     emotion -> cohesion      z =  +58.2  ***
#>     discuss -> synthesis     z =  +48.0  ***
#>     synthesis -> adapt       z =  +38.8  ***
#>     consensus -> coregulate  z =  +35.3  ***
#>     consensus -> plan        z =  +32.6  ***
#>     ... and 18 more
#> 
#>   Initial states:
#>     consensus  0.214  ████████████████████████
#>     plan       0.204  ███████████████████████
#>     discuss    0.175  ████████████████████
#>     emotion    0.151  █████████████████
#>     monitor    0.144  ████████████████
#>     cohesion   0.060  ███████
#>     synthesis  0.019  ██
#>     coregulate 0.019  ██
#>     adapt      0.011  █
lsa(tna_model)    # or from a built tna model -- same sequences, same fit
#> Lag Sequential Analysis  —  classical  (lag 1, directed)
#>   9 states | 25533 transitions | 27533 events | 2000 sequences
#>   states: adapt, cohesion, consensus, coregulate, discuss, emotion, monitor, plan, synthesis
#>   independence: G² = 13203.8, df = 64, p <2e-16
#> 
#>   Significant transitions (p < 0.05): 72 of 81
#>   strongest over-represented (of 23):
#>     emotion -> cohesion      z =  +58.2  ***
#>     discuss -> synthesis     z =  +48.0  ***
#>     synthesis -> adapt       z =  +38.8  ***
#>     consensus -> coregulate  z =  +35.3  ***
#>     consensus -> plan        z =  +32.6  ***
#>     ... and 18 more
#> 
#>   Initial states:
#>     consensus  0.214  ████████████████████████
#>     plan       0.204  ███████████████████████
#>     discuss    0.175  ████████████████████
#>     emotion    0.151  █████████████████
#>     monitor    0.144  ████████████████
#>     cohesion   0.060  ███████
#>     synthesis  0.019  ██
#>     coregulate 0.019  ██
#>     adapt      0.011  █

A TraMineR state sequence object

A TraMineR stslist (the standard state-sequence object) is read directly.

sts <- TraMineR::seqdef(engagement)
lsa(sts)
#> Lag Sequential Analysis  —  classical  (lag 1, directed)
#>   3 states | 1734 transitions | 1870 events | 136 sequences
#>   states: Active, Average, Disengaged
#>   independence: G² = 618.3, df = 4, p <2e-16
#> 
#>   Significant transitions (p < 0.05): 7 of 9
#>   strongest over-represented (of 3):
#>     Active -> Active          z =  +21.7  ***
#>     Disengaged -> Disengaged  z =  +15.4  ***
#>     Average -> Average        z =  +12.5  ***
#> 
#>   Initial states:
#>     Active     0.382  ████████████████████████
#>     Average    0.368  ███████████████████████
#>     Disengaged 0.250  ████████████████

Out of lagseq: hand a fit to another toolkit

A fitted model is a directed weighted network, so it converts to the native object of either network package. Estimation stays in lagseq; the downstream analysis runs in the toolkit built for it.

fit <- lsa(engagement)

To a tna network

lsa_to_tna() returns a tna object. Choose the edge weight with weights ("prob" for a transition-probability network, "count" for a frequency one). From there, tna’s analysis verbs apply.

tn <- lsa_to_tna(fit, weights = "prob")
tna::centralities(tn) |> head()
#> # A tibble: 3 × 10
#>   state    OutStrength InStrength ClosenessIn ClosenessOut Closeness Betweenness
#>   <fct>          <dbl>      <dbl>       <dbl>        <dbl>     <dbl>       <dbl>
#> 1 Active         0.302      0.324      0.0811       0.0779     0.100           0
#> 2 Average        0.390      0.664      0.160        0.0973     0.160           2
#> 3 Disenga…       0.517      0.221      0.0691       0.101      0.114           0
#> # ℹ 3 more variables: BetweennessRSP <dbl>, Diffusion <dbl>, Clustering <dbl>

Shared rendering: the transition network is a tna model

Because estimation and rendering are separate layers, the transition network view of an lsa fit is a tna model: lsa_to_tna() builds it on the fly and tna’s own plot method renders it (coloured nodes, initial-probability arcs, weighted directed edges).

plot(fit, type = "network", weights = "prob")

The residual network, by contrast, is rendered by cograph and keeps the lag-sequential meaning (each edge a tested departure from independence). One fit, two rendering layers, one shared object model.

One grammar across the ecosystem

The practical payoff is a single mental model. The arguments that sequence a raw log are the same in lsa() and in tna::prepare_data():

tna::prepare_data(log, actor =, action =, time =, order =, session =)
lsa(log,               actor =, action =, time =, order =, session =)

So an analysis can move between transition-network tooling and lag-sequential testing without reshaping the data, and an object built in one package is valid input to the other.