Pipeline: netify to ergm (statnet)

ERGMs (Exponential Random Graph Models) are the workhorse of inferential network analysis in the statnet ecosystem. netify provides a clean bridge: build your network with netify(), attach attributes with add_node_vars() / add_dyad_vars(), then convert to the network format that ergm expects with to_statnet().

This vignette covers:

The cross-sectional pipeline (single network → single ergm fit)
The longitudinal pipeline (per-time ergm fits)
The multilayer pipeline (per-layer ergm fits — to_statnet() now iterates layers automatically)
Round-tripping ergm-simulated networks back into netify for descriptive checks

library(netify)
library(ggplot2)
library(network)
#> 
#> 'network' 1.20.0 (2026-02-06), part of the Statnet Project
#> * 'news(package="network")' for changes since last version
#> * 'citation("network")' for citation information
#> * 'https://statnet.org' for help, support, and other information
library(ergm)
#> 
#> 'ergm' 4.12.0 (2026-02-17), part of the Statnet Project
#> * 'news(package="ergm")' for changes since last version
#> * 'citation("ergm")' for citation information
#> * 'https://statnet.org' for help, support, and other information
#> 'ergm' 4 is a major update that introduces some backwards-incompatible
#> changes. Please type 'news(package="ergm")' for a list of major
#> changes.

data(icews)

1. Cross-sectional pipeline

The simplest case: one snapshot, one model.

# build a single-year netify object with nodal and dyadic attributes
icews_2010 <- icews[icews$year == 2010, ]

verb_coop <- netify(
    icews_2010,
    actor1 = "i", actor2 = "j",
    symmetric = FALSE,
    weight   = "verbCoop",
    nodal_vars = c("i_polity2", "i_log_gdp"),
    dyad_vars  = "matlCoop"
)
#> ℹ `missing_to_zero` is set to "TRUE" (the default).
#> ! Missing dyads will be filled with zeros. For latent space or other
#>   statistical network models, structural zeros and missing data have different
#>   meanings. Set `missing_to_zero = FALSE` to preserve NAs if this distinction
#>   matters for your analysis.
#> This message is displayed once per session.

# convert to statnet 'network' object
sn_2010 <- to_statnet(verb_coop)
#> ! Nodal columns with "NA" detected: "i_polity2" and "i_log_gdp". Ergm terms
#>   like `nodecov()`/`nodematch()` will refuse to fit.
#> ℹ Use `drop_na_actors(net, cols = c('i_polity2', 'i_log_gdp'))` (or impute)
#>   before refitting.
#> ℹ Dyad covariates attached as per-edge attributes under "matlCoop_e" and as
#>   network-level matrices under their original names ("matlCoop").
#> ℹ For `ergm::edgecov()` use the matrix name (e.g. `edgecov('matlCoop')`); the
#>   "_e" per-edge attribute is for descriptive use such as edge styling.
#> This message is displayed once per session.
sn_2010
#>  Network attributes:
#>   vertices = 152 
#>   directed = TRUE 
#>   hyper = FALSE 
#>   loops = FALSE 
#>   multiple = FALSE 
#>   bipartite = FALSE 
#>   verbCoop: 152x152 matrix
#>   matlCoop: 152x152 matrix
#>   total edges= 9976 
#>     missing edges= 0 
#>     non-missing edges= 9976 
#> 
#>  Vertex attribute names: 
#>     i_log_gdp i_polity2 vertex.names 
#> 
#>  Edge attribute names not shown

to_statnet() carries nodal attributes through as vertex attributes and dyadic attributes through as edge attributes.

Before you fit: three sanity checks

ERGMs fail in cryptic ways when the underlying network is malformed. Three checks catch most “invalid output from statistic” headaches before you ever call ergm():

# 1. NAs in nodal covariates referenced by nodecov / nodematch
nd <- attr(verb_coop, "nodal_data")
na_cols <- names(nd)[vapply(nd, function(c) any(is.na(c)), logical(1))]
na_cols <- setdiff(na_cols, c("actor", "time", "layer"))
na_cols
#> [1] "i_polity2" "i_log_gdp"

# 2. isolates -- ergm fits but some terms (e.g. gwdegree) degenerate
m <- get_raw(verb_coop)
bin <- (m != 0) & !is.na(m)
deg_total <- rowSums(bin) + colSums(bin)
isolates <- rownames(m)[deg_total == 0]
length(isolates)
#> [1] 0

# 3. symmetry: the netify flag must match how your model treats ties
attr(verb_coop, "symmetric")
#> [1] FALSE

If na_cols is non-empty, drop the affected actors with drop_na_actors(verb_coop, cols = na_cols) or impute before converting. Now fit an ergm:

# Note: this chunk is not evaluated by default to keep vignette build fast.
# Replace eval = FALSE with eval = TRUE to actually run.
set.seed(6886)
m <- ergm(
    sn_2010 ~ edges +
        nodecov("i_polity2") +
        nodecov("i_log_gdp")
)
summary(m)

One footgun worth flagging: nodecov("i_polity2") will refuse to fit if any vertex has an NA polity score. Either subset to actors with complete covariates or impute before passing to to_statnet(). netify emits a one-shot inform when it detects NAs in a nodal attribute and stashes the offending column names on the resulting network object so you can introspect:

attr(sn_2010, "netify_na_cols")   # character vector of NA-bearing nodal vars

If non-empty, drop or impute those columns before fitting ergm() formulas that reference them with nodecov() / nodematch(). The drop_na_actors() helper does this in one call and works for cross-sectional, longitudinal, and bipartite netlets:

clean <- drop_na_actors(verb_coop, cols = attr(sn_2010, "netify_na_cols"))
sn_2010_clean <- to_statnet(clean)

Dyadic edge covariates: the `_e` suffix

Any dyadic covariate you passed to netify() (e.g. dyad_vars = "matlCoop") is attached to the resulting network object in two places:

as a network-level attribute under its original name (the full n x n matrix), accessible via network::get.network.attribute(sn, "matlCoop"), and
as a per-edge attribute under <var>_e (here, matlCoop_e), populated only on edges that actually exist.

The trailing _e disambiguates the per-edge edgelist from the network-level matrix. For ergm::edgecov(), pass the original (matrix) name — edgecov() resolves its argument as a network-level matrix attribute, so the _e per-edge alias will not work there:

m <- ergm(sn_2010 ~ edges + edgecov("matlCoop"))

The _e per-edge attribute is exposed for descriptive uses (for example, coloring edges by covariate in network::plot.network). to_statnet() emits a one-shot inform listing both forms the first time it attaches dyadic covariates.

Goodness-of-fit, mcmc.diagnostics, and other postestimation tools live in ergm itself.

2. Longitudinal pipeline (per-time fits)

For a longitudinal netify, to_statnet() returns a named list — one network object per time period:

verb_longit <- netify(
    icews[icews$year %in% 2010:2012, ],
    actor1 = "i", actor2 = "j", time = "year",
    symmetric = FALSE,
    weight = "verbCoop",
    nodal_vars = "i_polity2"
)

sn_list <- to_statnet(verb_longit)
#> ! Nodal columns with "NA" detected: "i_polity2". Ergm terms like
#>   `nodecov()`/`nodematch()` will refuse to fit.
#> ℹ Use `drop_na_actors(net, cols = c('i_polity2'))` (or impute) before
#>   refitting.
#> This message is displayed once per session.
length(sn_list)
#> [1] 3
names(sn_list)
#> [1] "2010" "2011" "2012"
class(sn_list[[1]])
#> [1] "network"

Fit a separate ergm per period:

set.seed(6886)
fits <- lapply(sn_list, function(n) {
    ergm(n ~ edges + nodecov("i_polity2"))
})

For coevolution models (where ties change as a function of past ties), look at tergm from the statnet suite — netify provides the data, tergm does the modeling. A typical tergm 4.x call against the same per-period list looks like this:

# Longitudinal ERGM via tergm 4.x
library(tergm)
set.seed(6886)

nets <- to_statnet(verb_longit)   # named list of network objects
fit <- tergm(
    nets ~ Form(~ edges + nodecov("i_polity2")) +
           Persist(~ edges),
    estimate = "CMLE",
    times = seq_along(nets)
)
summary(fit)

In the tergm 4.x split formulation, Form() models the formation of new edges between periods, and Persist() models the persistence of existing edges from one period to the next; both formulas accept the same ergm terms (edges, nodecov, nodematch, mutual, gwesp, etc.). (Note: i_polity2 has missing values for some country-years — to_statnet() flags this and stashes the offending column names on attr(net, "netify_na_cols") so you can drop or impute before fitting.)

3. Multilayer pipeline (per-layer fits)

Multilayer ergm modeling is its own research literature. The simple pragmatic approach is to fit one ergm per layer. to_statnet() handles this automatically: pass a multilayer netify, get back a named list keyed by layer.

verb <- netify(icews_2010, actor1 = "i", actor2 = "j",
               symmetric = FALSE, weight = "verbCoop")
matl <- netify(icews_2010, actor1 = "i", actor2 = "j",
               symmetric = FALSE, weight = "matlCoop")
multi <- layer_netify(list(verbal = verb, material = matl))

sn_multi <- to_statnet(multi)
length(sn_multi)
#> [1] 2
names(sn_multi)
#> [1] "verbal"   "material"

Each element is a network object you can plug straight into ergm().

4. Round-tripping simulated networks back to netify

A useful descriptive check after fitting an ergm is to simulate from the fit, compute descriptives on the simulated networks, and compare to the observed network. simulate.ergm() returns network objects, which means you can pipe them straight back into a netify for any of netify’s descriptive tools:

set.seed(6886)
sims <- simulate(m, nsim = 100)  # list of network objects

# convert each simulated network back into a netify for comparison
sim_nets <- lapply(sims, function(s) to_netify(s))

# compare observed vs simulated at the structural level
all_nets <- c(list(observed = verb_coop), sim_nets)
struct_comp <- compare_networks(all_nets, what = "structure")

This gives you observed-vs-simulated comparisons for density, reciprocity, transitivity, mean degree, etc. — useful for assessing whether your ergm captured the descriptive features you care about.

tl;dr

# build → attach attrs → export → model
net <- netify(df, actor1 = "i", actor2 = "j", symmetric = FALSE, weight = "x")
net <- add_node_vars(net, attrs, actor = "id")
sn  <- to_statnet(net)               # single network, longit list, or multilayer list
m   <- ergm(sn ~ edges + ...)        # modeling happens in ergm/statnet

For modeling beyond ergm:

Latent factor / additive-multiplicative: pipeline_lame_dbn vignette
Social relations model: netify-dev/srm via to_amen(net)
Social influence regression (Minhas & Hoff): netify-dev/sir

Have fun!

References

Butts, C.Butts, C. T. (2008). network: A Package for Managing Relational Data in R. Journal of Statistical Software, 24(2), 1–36. https://doi.org/10.18637/jss.v024.i02
Cranmer, S. J., Desmarais, B. A., & Morgan, J. W. (2021). Inferential Network Analysis. Cambridge University Press. https://doi.org/10.1017/9781316662915
Hunter, D. R., Handcock, M. S., Butts, C. T., Goodreau, S. M., & Morris, M. (2008). ergm: A Package to Fit, Simulate and Diagnose Exponential-Family Models for Networks. Journal of Statistical Software, 24(3), 1–29. https://doi.org/10.18637/jss.v024.i03
Krivitsky, P. N., & Handcock, M. S. (2014). A Separable Model for Dynamic Networks. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 76(1), 29–46. https://doi.org/10.1111/rssb.12014
Snijders, T. A. B., Pattison, P. E., Robins, G. L., & Handcock, M. S. (2006). New Specifications for Exponential Random Graph Models. Sociological Methodology, 36(1), 99–153. https://doi.org/10.1111/j.1467-9531.2006.00176.x

Cassy Dorff and Shahryar Minhas

2026-05-29

1. Cross-sectional pipeline

Before you fit: three sanity checks

Dyadic edge covariates: the `_e` suffix

2. Longitudinal pipeline (per-time fits)

3. Multilayer pipeline (per-layer fits)

4. Round-tripping simulated networks back to netify

tl;dr

References

Pipeline: netify to ergm (statnet)

Cassy Dorff and Shahryar Minhas

2026-05-29

1. Cross-sectional pipeline

Before you fit: three sanity checks

Dyadic edge covariates: the _e suffix

2. Longitudinal pipeline (per-time fits)

3. Multilayer pipeline (per-layer fits)

4. Round-tripping simulated networks back to netify

tl;dr

References

Dyadic edge covariates: the `_e` suffix