nwrem and relevent

nwrem implements the relational event framework of Carter Butts (2008), and relevent, Butts’ R package, is the program it was checked against. nwrem is not affiliated with or endorsed by the relevent project, and errors in nwrem are the author’s. If you can use R, relevent remains the reference and the more complete program; nwrem is for people who work in Stata, or prefer it.

Where the package comes from

In plain words, as in the notice that ships with the package: nwrem implements the statistical model of Butts (2008). It was developed with the R package relevent as the reference implementation and as the target of the cross-checks. Because the package follows the structure of relevent, it is distributed under the GNU General Public License, version 3, which is compatible with relevent’s license (GPL, version 2 or later). The package exists because of that work.

The software was developed with Claude Code (Anthropic), an AI coding assistant, working under the author’s direction; its correctness rests on numerical comparison with the R package.

What is the same

• The model. The ordinal partial likelihood of Butts (2008) with maximum likelihood estimation, the analogue of rem.dyad() with ordinal = TRUE and fit.method = "MLE".
• The effects of the table below, with relevent’s definitions: the degree statistics are normalised shares, and the degree-accumulation semantics were checked against the output of relevent’s acl.deg().
• The comparison. Coefficients and log likelihoods were compared with rem.dyad() of relevent 1.2-1 on identical simulated event sequences; they agree within the tolerances of the package’s test scripts. See Tutorial 3 for how the optimizer’s starting point changes the fifth digit at most.

What differs

Not implemented in nwrem:

• triadic and shared-partner effects;
• fixed effects;
• the full likelihood that uses the waiting times (the non-ordinal likelihood);
• Bayesian estimation.

There is no goodness-of-fit command and no intercept (an intercept cancels out of the ordinal likelihood, so it is not estimable there). nwrem has no compiled plugin; the fits run in Mata (see Speed below). Data preparation differs as well: instead of an R edgelist matrix you declare the events with nwset ..., eventtime(), add actors that never appear in the log with nwaddnodes, and read actor attributes from the dataset of the actors (nwload ..., xvars).

Translation table

relevent	nwrem
the event list passed to rem.dyad(edgelist, n, ...)	nwset sender receiver, directed eventtime(t) name(net) (actors absent from the log: nwaddnodes)
rem.dyad(..., ordinal = TRUE, fit.method = "MLE")	nwrem net, <effects>
NODSnd	nodsnd
NIDRec	nidrec
NIDSnd	nidsnd
NODRec	nodrec
NTDegSnd, NTDegRec	ntdegsnd, ntdegrec
FrPSndSnd	frpsndsnd (inertia)
FrRecSnd	frrecsnd (reciprocity of receipt)
RSndSnd, RRecSnd	rsndsnd, rrecsnd
CovSnd, CovRec, CovInt with covar = list(...)	covsnd(var), covrec(var), covint(var), the variable holding one value per actor
CovEvent	covevent(netname), a network with the same actors holding the pair values
summary() of the fit	the table printed by nwrem; e(b), e(V), e(ll), e(N), e(nodes)

Speed

nwrem runs in Mata (there is no compiled plugin). On 26 and 27 October 2026 we timed it against rem.dyad() of relevent 1.2-1 on simulated event streams of 200 to 20,000 events among 20 to 100 actors (Apple M5 Pro, Stata 18, R 4.6.0). nwrem always computes the covariance matrix, so the like-for-like setting in relevent is hessian=TRUE; relevent’s default is hessian=FALSE, and both were timed. nwrem was faster than relevent with degree, inertia and covariate effects (up to 41 times with the Hessian) and slower than relevent without the Hessian with the recency effects (rsndsnd rrecsnd). The tables, the cases with a time-out and the method are on the page Speed compared with R. The coefficients of the two programs agree to about three significant digits, not to six, because each stops at its own convergence tolerance. Timings depend on the machine, the data and the effects.

Acknowledgment

nwrem rests on the work of Carter T. Butts, who developed the relational event framework and the relevent package. If you use nwrem, please cite the paper and the package, as on the page How to cite.