nwergm and R’s ergm

nwergm follows the statnet project closely and exists because of it. The ERGM framework and R’s ergm are due to Mark Handcock, David Hunter, Carter Butts, Steven Goodreau, Pavel Krivitsky, Martina Morris and many contributors, and nwergm re-implements their terms, estimators and diagnostics in Stata. nwergm is not affiliated with or endorsed by the statnet project, and errors in nwergm are the author’s. If you can use R, ergm remains the reference and the more complete program.

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 R’s ergm.

What is the same

The model. The terms are those of ergm, with the same definitions and the same coefficient names (with . written as _). Version 1.1.0 implements 125 of the 139 terms of ergm 4.12.0 (89.9%), and 125 of the 126 that apply to binary networks; version 1.0.1 has about 110.
The estimators. Maximum pseudolikelihood and Monte Carlo maximum likelihood with the same defaults as R (the TNT proposal, bridge-sampling log-likelihood), plus the variants of the pseudolikelihood, contrastive divergence and the stochastic approximation.
Missing data. Missing dyads are left out of the likelihood and handled by the method of Handcock and Gile, as in R; missing covariate values are refused, as in R.
Fit statistics, constraints, simulation and checks. Log-likelihood, deviance, AIC and BIC as logLik(); constraints= and obs.constraints=; simulation from a fit; goodness of fit and MCMC diagnostics.

How it was compared

For each term, the observed statistic of test networks was compared with the one that ergm computes, and the maximum pseudolikelihood estimates and the Monte Carlo maximum likelihood estimates were compared with ergm 4.12.0 on the statnet example networks and on simulated networks. Because the Monte Carlo estimates are random, results agree within simulation noise, not to the last digit; nwergm version prints the version of ergm that was used.

What differs, and what is missing

Not in nwergm:

the 13 terms for valued networks (atleast, atmost, cdf, cyclicalweights, equalto, greaterthan, ininterval, nodecovar, nodeicovar, nodeocovar, smallerthan, sum, transitiveweights) and the operator B. nwergm models binary ties; the values of a valued network are not used as an outcome;
Curve() (and Parametrise()) beyond linear maps of the coefficients: map(rep) and map(matrix(name)) work, minpar(), maxpar() and cov() do not;
altkstar with an estimated decay by Monte Carlo maximum likelihood: it is available for method(mple) with starting values from init(). R cannot fit it by Monte Carlo maximum likelihood either;
in the operators, the two-sided form of attrs in S(), dyadic covariates inside S() and Project(), and curved terms inside an operator; the constraints hamming, ChangeStats(), Dyads() and I();
the packages built on ergm: dynamic models (tergm), count data (ergm.count) and multiple networks (ergm.multi).

Differences in usage: attributes must be numeric variables in the current dataset; an option name cannot have a digit followed by a letter, so R’s b1degree is bdegree1(); the operators and structural() take R-style terms inside Stata options; a fit is stored in e() like any Stata estimation command, so estimates store, lrtest, test and predict work as usual.

Translation table

R (`ergm`, `statnet`)	`nwergm`
`net <- network(...)`, `ergm.getnetwork` (the network of the formula)	the network is declared with `nwset` (or loaded with `nwuse` / `nwwebuse`) and named in the command: `nwergm net, ...`; nodal attributes are variables of the current dataset
`ergm(net ~ edges + nodematch("sex"))`	`nwergm net, edges nodematch(sex)`
`nodematch("sex", diff=TRUE)`	`nodematchdiff(sex)`
`b1degree(0:2)`, `gwb2degree(0.5, fixed=TRUE)`	`bdegree1(0 1 2)`, `bgwdegree2(0.5)`
`gwesp(0.5, fixed=TRUE)`, `gwesp(fixed=FALSE)`	`gwesp(0.5)`, `gwespfree(0.5)`
`Symmetrize(~triangle, rule="strong")`	`operators(Symmetrize(triangle, rule(strong)))` (version 1.1.0)
`ergm(..., estimate="MPLE")`, `main.method=`	`estimate(mple)`, `mainmethod(mcmle)` or `mainmethod(sa)`
`control.ergm(MCMC.burnin=, MCMC.interval=, MCMC.samplesize=, MCMLE.maxit=)`	`mcmcburnin()`, `mcmcinterval()`, `mcmcsamplesize()`, `mcmleiterations()`
`control.ergm(MCMLE.effectiveSize=)`	`mcmleeffectivesize()` (version 1.3.0; default 64 as in R, `0` for fixed-size samples)
`constraints = ~edges`, `obs.constraints = ~...`	`constraints(edges)`, `obsconstraints(...)`
`offset(edges)` with `offset.coef`	`offset(edges #)`
`coef(fit)`, `vcov(fit)`	`e(b)`, `e(V)`; `matrix list e(b)`
`summary(fit)`	the table printed by `nwergm`
`summary(net ~ terms)` (observed statistics)	the column `obs` of `estat gof`; `nwergm simulate, output(stats)` for simulated statistics
`simulate(fit, nsim = 100)`	`nwergm simulate, nsim(100)`
`gof(fit)`	`estat gof`
`mcmc.diagnostics(fit)`	`estat mcmcdiag` (with `plot` for trace plots)
`logLik(fit)`, `AIC(fit)`, `BIC(fit)`	`estat ic`, `e(ll)`, `e(aic)`, `e(bic)`; `estimates stats`
`anova(fit1, fit2)`	`lrtest`
`predict(fit)`	`predict, frame(name) pr` or `predict, matrix(name) pr`
`packageVersion("ergm")`	`nwergm version`

Where next

Terms is a map of the term families, and the tutorials (Tutorial 2 onward) use the commands above on real data. The complete syntax is in Stata’s help nwergm.