nwdynam and goldfish

nwdynam follows the R package goldfish and exists because of it. The dynamic network actor model is due to Christoph Stadtfeld and Per Block (2017); goldfish (Stadtfeld, Hollway, Block and many contributors) is the reference implementation, and nwdynam takes its effect definitions, its likelihoods and its way of estimating from that work. nwdynam is not affiliated with or endorsed by the goldfish project, and errors in nwdynam are the author’s. If you can use R, goldfish remains the reference and the more complete program; nwdynam is for people who work in Stata, or prefer it.

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.

nwdynam is distributed under the GNU General Public License, version 3, as goldfish is (GPL version 3 or later).

What is the same

• The model. A rate sub-model (which actor acts next) and a choice sub-model (whom the actor chooses), plus the coordination sub-model for undirected events (submodel(choice_coordination)).
• The effects. The effects of the choice and rate models that are listed in the table below, with the same definitions.
• The estimation. Maximum likelihood with the ordinal partial likelihood (only the order of events); with intercept, the continuous-time hazard of the rate model.
• The comparison. Coefficients and log likelihoods were compared with goldfish 1.6.12 on the Social Evolution data that goldfish ships (84 actors, 439 phone calls) and on small hand-built examples. On those data the choice model with inertia, recip and indeg gives 5.3656, 1.3894 and -0.1735 in both programs (see Tutorial 1). The compiled plugin was checked against the Mata implementation of the package to floating-point round-off.

What differs

Things that goldfish does and nwdynam does not (yet) do, as documented in help nwdynam:

• the effect tertiusDiff, the mixed variants of the closure effects that use two networks, and closure effects through an exogenous network;
• a second network in tie() that itself changes over time (nwdynam accepts a static network only);
• the ignoreRep argument and, for tie(), the arguments weighted and window;
• windowed rate effects (indegwindow(), outdegwindow()) in combination with intercept;
• windows on tertius(), and weighted or windowed versions in choice_coordination;
• attribute effects (same, diff, sim, ego, alter, egoalterint, tertius) and closure effects in two-mode networks;
• opportunities() in the same call as an attribute effect (the two need different data layouts);
• time-varying covariates: attributes are fixed over the log.

The help file documents no goodness-of-fit, simulation or Bayesian command. The estimates and the log likelihood are the output. Data preparation is also different: instead of goldfish’s objects (defineNodes, defineNetwork, linkEvents, defineDependentEvents) you declare an event list with nwset ..., eventtime() and read attributes from the dataset of the actors (nwload ..., xvars). Before using nwdynam for an important result, run the same model in goldfish.

Translation table

goldfish	nwdynam
defineNodes(), defineNetwork(), linkEvents(), defineDependentEvents()	nwset sender receiver, directed eventtime(t) name(net); isolates with nwaddnodes; attributes with nwload net, xvars
estimate(f, model = "DyNAM", subModel = "choice")	nwdynam net (the default sub-model)
... subModel = "rate"	nwdynam net, submodel(rate)
... subModel = "choice_coordination"	nwdynam net, submodel(choice_coordination) on an undirected network
inertia, recip	inertia, recip
indeg, outdeg	indeg, outdeg (both sub-models)
trans, cycle, commonSender, commonReceiver, four	trans, cycle, commonsender, commonreceiver, four
nodeTrans	nodetrans
same(attr), diff(attr), sim(attr)	same(var), diff(var), sim(var)
ego(attr) (rate), alter(attr) (choice)	ego(var), alter(var)
egoAlterInt(attr)	egoalterint(var1 var2)
tertius(attr)	tertius(var)
tie(network)	tie(netname)
window = w on an effect	inertiawindow(#), recipwindow(#), indegwindow(#), outdegwindow(#)
weighted = TRUE on an effect	weightedinertia, weightedrecip, weightedindeg, weightedoutdeg
a leading 1 + in the formula (intercept, rate model)	intercept with submodel(rate)
opportunitiesList in estimationInit	opportunities(evvar actvar)
summary() of the estimate	the table printed by nwdynam; e(b), e(V), e(ll), e(N), e(nodes)

Without any effect option nwdynam fits inertia, recip and indeg for the choice model and indeg and outdeg for the rate model; naming any effect restricts the model to the effects named. In goldfish the formula always lists the effects.

Acknowledgment

nwdynam is a port of ideas, definitions and methods that others developed. If you use it, please cite the DyNAM papers and goldfish, as on the page How to cite.