nwsaom and RSiena

nwsaom follows the Siena project closely and exists because of it. The stochastic actor-oriented model is due to Tom Snijders. RSiena (Ripley, Snijders, Boda, Voros, Preciado and many contributors) is the reference implementation, and multiSiena (Snijders, Koskinen, Ripley) supplies the Bayesian estimator that nwsaom bayes follows. nwsaom is not affiliated with or endorsed by the Siena project, and errors in nwsaom are the author’s. If you can use R, RSiena remains the reference and the more complete program.

What is the same

• The model and the effects. The effect library follows RSiena’s table of effects. In the effects table of RSiena 1.6.6 (getEffects() over twelve data configurations) 1,121 rows are planned for nwsaom; 1,058 of them (94%) are implemented, and every implemented effect was compared with siena07. The effects that RSiena 1.6.7 to 1.6.14 added or changed (for example WXZX and the two-mode toU) are implemented and compared with RSiena 1.6.14 but are not part of these counts; the counts change with RSiena releases.
• The algorithm and the defaults. Method of moments by Robbins-Monro stochastic approximation in three phases, conditional estimation for one network, the same starting values, the same convergence rule (all t-ratios below 0.1, overall maximum ratio below 0.25), centered covariates.
• The output. Coefficients, rates and convergence statistics as in RSiena; the option detail prints RSiena’s names of the effects.
• The postestimation tools: goodness of fit, score tests, time-heterogeneity test, relative importance.

What differs

Not implemented in nwsaom:

• the generalised method of moments estimator;
• the settings model;
• a hierarchical Bayesian analysis of several groups (the joint estimation of one model in several groups is the option groups(), see Tutorial 14; groups fitted separately can be combined with nwsaom meta, the counterpart of siena08). With groups(), maximum likelihood and Bayesian estimation, estat gof, estat simulate and estat importance, two-mode and non-directed networks, several networks and a few effects are not yet available;
• the continuous (stochastic differential equation) behavior;
• inside nwsaom multinet: composition change, missing ties, estat gof, estat importance and estat simulate.

Maximum likelihood and Bayesian estimation do not cover non-directed or two-mode networks, several networks, missing data or composition change. Only binary ties are supported. A few implementation details differ and are disclosed in the help: for example estat gof simulates fresh networks at the estimates instead of reusing those stored during the fit.

How the estimates were compared

For each implemented effect the observed statistic was compared with the one siena07 computes, and the estimates were compared over at least two seeds; for a few effects, where the estimates are degenerate in both programs, only the statistic could be compared. The comparisons were made with RSiena 1.6.14 (see nwsaom version). Because both programs estimate by simulation, results agree within simulation noise, not to the last digit. The manual (PDF) gives the validation in detail.

Speed

nwsaom runs its simulations in a compiled, multithreaded plugin and evaluates the change statistics of an actor in batches. In our benchmark on a single thread, nwsaom was faster than RSiena in all 28 models (median 2.1 times, 1.5 to 9.0; version 1.1.1). The tables, the cases where it is not faster and the method are on the page Speed compared with R.

Translation table

RSiena	nwsaom
sienaNet, sienaDataCreate	nwset for the networks; waves and covariates are named in the nwsaom options
print01Report	nwsaom describe, waves(...)
coCovar, varCovar	variables in the dataset; tvcov() for changing covariates
getEffects, includeEffects	effect options of nwsaom (outdegree, reciprocity, transtrip, gwesp(), egox(), altx(), samex(), simx(), …); detail shows RSiena’s names
includeInteraction	interact()
effect type endow, creation	endow(), creation()
includeEffects(fix, test), score.Test	fix() and estat score
sienaAlgorithmCreate (n1, n3, seed, firstg, cond, maxlike, cores)	k0(), k3(), seed(), firstg(), unconditional, ml, cores()
siena07	nwsaom
siena07 with simOnly = TRUE	nwsaom simulate, estat simulate
sienaGOF, plot.sienaGOF	estat gof
sienaTimeTest	estat timetest
sienaRI	estat importance
sienaGroupCreate and siena07	groups() (and gwaves()) on a data set with the groups stacked
siena08	nwsaom meta
sienaCompositionChange	present(), joinat(), leaveat()
missing data in sienaDataCreate	missnet(), missbeh()
multi_siena / sienaBayes	nwsaom bayes (one directed network, with or without one behavior)
several dependent networks	nwsaom multinet
netmediate (MEMS)	estat mems

RSiena has renamed several of its functions in recent versions (for example siena for siena07, test_gof for sienaGOF); the older names are used above.

About the software

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 RSiena.

References

Snijders, T.A.B. (2001). The statistical evaluation of social network dynamics. Sociological Methodology 31. Ripley, R.M., Snijders, T.A.B., Boda, Z., Voros, A., Preciado, P. Manual for RSiena. Lospinoso, J., Snijders, T.A.B. (2019). Goodness of fit for stochastic actor-oriented models. Methodological Innovations 12(3).