Skip to contents

amorem amorem logo

amorem — end-to-end simulation, sampling, feature engineering, model selection, and inference for relational event models (REMs) in R, in one consistent API.

The package consolidates four lines of recent methodological work:

Installation 

install.packages("amorem")                              # CRAN release
# remotes::install_github("franciscorichter/amorem")    # development version

30-second tour 

library(amorem)

# 1. Bundled datasets, ready to load
data(classroom_events)        # 691 events,    20 actors  (McFarland 2001)
data(social_evolution_calls)  # 439 events,    54 actors  (Madan et al. 2012)
data(radoslaw_email)          # 82,927 emails, 167 actors (Nurek & Michalski 2020)

# 2. Compute endogenous features post-hoc from any (sender, receiver, time) log
feat <- endogenous_features(
  classroom_events, stats = c("reciprocity", "recency"))

# 3. Simulate a ready-to-fit case-1-control table with known structure ...
w <- simulate_relational_events(
  n_events = 1500, senders = LETTERS[1:8], receivers = LETTERS[1:8],
  n_controls = 1, endogenous_stats = "reciprocity_count",
  endogenous_effects = c(reciprocity_count = 0.4), wide = TRUE)

# 4. ... and fit it with the default "gam" backend
fit <- rem(~ reciprocity_count, data = w, method = "gam")
summary(fit)

What’s inside

  • rem() — the unified fitter for preprocessed case-control data, with a conditional-logit backend (case-k-control) and a gam backend (case-1-control) supporting linear / tv / nl / tvnl smooth effects and re() random effects, plus summary() / coef() / plot(). widen_case_control() reshapes a long case-control log into the wide form rem() expects.
  • simulate_relational_events() — exact Gillespie or τ-leap; composes exogenous, endogenous, and time-varying global covariates; with wide = TRUE emits a ready-to-fit case-1-control table.
  • endogenous_features() — the endogenous catalogue computed post-hoc from any (sender, receiver, time) log. On case-control data only the observed events build the network history; the sampled controls are scored against it without entering it.
  • sample_non_events() — nested case-control sampling with appearance / citation / remove risk-set rules, and time-shifted sampling (shift = "exponential") that keeps global covariates estimable.
  • Datasets — event logs from published interaction networks, and enso_nino34, a reconstructed El Niño index (1301–2005) for use as a global covariate.
  • Goodness of fit — gof_rem() tests a fitted rem() model (time, functional-form and score processes, bootstrap null that accounts for the estimated coefficients); gof_univariate(), gof_multivariate(), gof_global(), gof_auxiliary() refit a linear model from the event log; pointwise martingale_residuals().
  • The earlier compare_models(), compare_models_smooth() and compare_models_global() remain available (superseded by rem()).

Guides

Guide What you’ll find
Quick start install + simulate-and-recover in a few lines
Simulation the dyadic mechanisms, Gillespie vs τ-leap
Sampling options the risk-set options of sample_non_events(): n_controls, mode, scope, risk = "remove", exclude_concurrent, exclude_pairs, shift
Endogenous catalogue the statistic catalogue and its variant axes
Estimation case-control sampling, the three rem() backends — clogit (linear), gam (smooth tv/nl/tvnl/re effects), nn (neural conditional logit) — model comparison, GOF
Hyperedge models the (I, J, time) data model, subset repetition, multi-recipient e-mails as hyperevents, and the RHEM simulators
Datasets the event logs and covariates that ship with amorem: what each one looks like, its source and licence
Data catalogue 78 relational event datasets and covariates, bundled or not: what each one is, its licence, code to download it, and pictures of its network and timeline; no models are fitted
Analyses of bundled event logs models fitted to three bundled logs (classroom, company email, CollegeMsg): timing vs count statistics with and without a sender frailty, smooth effect curves
Validation experiments recovery, smooth, scaling, parity, and the neural backend (gradient check + interaction recovery)
GoF calibration type-I error of the martingale-residual GoF tests under correct specification
Case study: email and the clock time of day and day of week as global covariates in the Manufacturing e-mail log, time-shifted sampling, cyclic smooths, tv() on shifted data, gof_rem(), gof_multivariate(), gof_auxiliary()
Case study: Social Evolution data intake, actor attributes and survey friendship as covariates, rem(method = "clogit"), a sender random effect with re()
Case study: estimation backends clogit, neural (nn, MLP and additive splines, nn_uncertainty()) and gam with tv() / tvnl() on Email-EU-core, compared on held-out e-mails
Case study: cosponsorship in the 118th Congress bills as directed hyperevents (sponsor → cosponsor set): hypergraph, party-coloured network, chord and adjacency views; a sender- and size-matched case-control fit of repetition, reciprocity, pair co-cosponsorship, party and state, and a tv() party-homophily curve
Case study: calls and text messages two-layer (multiplex) REM on the Copenhagen Networks Study: cross-layer history as a covariate, reciprocity and repetition per layer, gender homophily, the daily cycle as a global effect with time-shifted sampling and cyclic smooths
Case study: leader visits to and from the US a star-shaped event network, 1945–2019: per-direction case-control sampling (mode = "two", scope = "appearance"), reciprocity and inertia by direction with rem(method = "clogit"), the nl() decay of reciprocity and tv() inertia over 75 years
Case study: armed-conflict dyads yearly tied events scored from earlier years only (prior_log), a risk set of dyads with a past; recurrence, escalation, the decay of a conflict’s pull with nl(), and a tv() effect
Case study: international football matches since 1872 as undirected events: rivalry network with regional groups, recency smooth, common opponents (triadic closure), friendlies against competitive matches
Case study: first records of alien species observed species-to-region first records with mode = "two", risk = "remove", prior_log and time-shifted sampling; shared-invader network, timing heatmap, taxonomy by group, ENSO (ONI) as a global smooth
Species invasions a one-shot (risk = "remove") process with ENSO as a global driver, fitted by time-shifted sampling

Full per-function documentation is under Reference.

References

  • Juozaitienė R., Wit E.C. (2024). It’s about time: revisiting reciprocity and triadicity in relational event analysis. JRSS-A 188(4), 1246–1262. doi:10.1093/jrsssa/qnae132.
  • Boschi M., Lerner J., Wit E.C. (2025). Beyond Linearity and Time-Homogeneity: Relational Hyper Event Models with Time-Varying Non-Linear Effects. arXiv:2509.05289.
  • Lembo M., Juozaitienė R., Vinciotti V., Wit E.C. (2025). Relational Event Models with Global Covariates. JRSS-C.
  • Boschi M., Wit E.C. (2025). Goodness of fit in relational event models. Statistics and Computing 36(4).