Compare candidate specifications with smooth (TV / NL / TVNL) effects

Superseded by rem(), which fits the same smooth (TV / NL / TVNL) effects on preprocessed case-control data. compare_models_smooth() remains fully supported.

Mirrors compare_models() but lets each statistic in a specification take one of four effect types instead of a single linear coefficient: linear, time-varying (TV), non-linear (NL), or jointly time-varying non-linear (TVNL). The smooth machinery follows Boschi, Lerner & Wit (2025); the matrix-of-event-vs-non-event trick is documented in their Section 3.3.

For each specification:

One case-control sample is drawn from event_log with n_controls = 1 (paired event / non-event design).
For every requested statistic, both the case (event) and the control (non-event) features are computed via endogenous_features().
The mgcv design uses the case-vs-control matrix trick:
- linear -> a single coefficient on case - control (column d_stat).
- tv -> s(time, by = d_stat) — smooth in time, multiplied by d_stat.
- nl -> s(stat_mat, by = I_mat) where stat_mat is a two-column matrix cbind(case, control) and I_mat is cbind(1, -1).
- tvnl -> te(time_mat, stat_mat, by = I_mat) tensor product smooth, with time_mat both columns equal to the event time vector.
The model is fitted with mgcv::gam and a degenerate logistic likelihood: response = rep(1, n), formula = one ~ -1 + ..., family = binomial. This matches Boschi et al. equation 8.

AIC values are directly comparable across specifications because every fit uses the same case-control sample. Returns the same tidy data.frame as compare_models().

Usage

compare_models_smooth(
  event_log,
  models,
  scope = c("all", "appearance", "citation"),
  mode = c("one", "two"),
  half_life = NULL,
  k = NULL,
  seed = NULL,
  keep_fits = FALSE
)

Arguments

event_log

Data frame with sender, receiver, time columns.

models

Named list of specifications. Each entry is itself a named character vector (or named list) mapping statistic names to effect types: "linear", "tv", "nl", or "tvnl". Example:


    list(
      linear = c(reciprocity_count   = "linear",
                 transitivity_count  = "linear"),
      nl    = c(reciprocity_time_recent  = "nl",
                 transitivity_time_recent = "nl"),
      tvnl  = c(reciprocity_time_recent  = "tvnl",
                 transitivity_time_recent = "tvnl"))

scope, mode

Passed through to sample_non_events().

half_life

Required when an exp-decay statistic is requested.

k

Optional integer: knot count for s() and te() terms. Default NULL lets mgcv choose (-1).

seed

Integer seed for the case-control sample.

keep_fits

Logical; when TRUE, the returned table carries the fitted model objects (one per spec, named by model, NULL for specs that failed) as attr(result, "fits"), e.g. for plotting estimated effects. Defaults to FALSE.

Value

Data frame with one row per specification and columns model, n_terms, n_obs, log_lik, AIC, delta_AIC.

References

Boschi M, Lerner J, Wit EC (2025). Beyond Linearity and Time-Homogeneity: Relational Hyper Event Models with Time-Varying Non-Linear Effects. arXiv:2509.05289.

Examples