netimpute

Joint multiple imputation of network ties and node attributes in R.

Missing data in network studies usually affects both sides at once: the respondent who skipped the attribute survey is often the one whose nominations are missing too. netimpute imputes both together, in one chained-equations loop, so that each side informs the other.

Installation

Install the released version from CRAN:

install.packages("netimpute")

Or the development version from GitHub:

# install.packages("remotes")
remotes::install_github("RWKrause/netimpute")

Usage

Supply a data frame of node attributes and a list of adjacency matrices, with NA marking both missing attribute values and unknown ties:

library(netimpute)

set.seed(1)
n <- 30
friends <- matrix(rbinom(n * n, 1, 0.15), n, n)
diag(friends) <- 0
friends[sample(which(row(friends) != col(friends)), 40)] <- NA

attrs <- data.frame(age = rnorm(n, 35, 8),
                    performance = rnorm(n),
                    gender = sample(c("F", "M"), n, TRUE))
attrs$performance[sample(n, 4)] <- NA

fit <- netmice(attrs, list(friends = friends), m = 5, maxit = 20)
completed <- complete_netmice(fit, 1)
plot(fit)   # convergence diagnostics

Each attribute is imputed from the other attributes and from node-level network measures (degree, centrality, brokerage, closure, and homophily measures computed over a node's alters). Each network's ties are imputed from dyad-level predictors built from the attributes and the other networks. Both predictor sets are rebuilt at every visit.

Missing ties are redrawn tie-wise by default: one tie at a time, conditional on the ties imputed so far, with reciprocity and the shared-partner indicator refreshed after every single draw.

Further options

• models — protect a hypothesised relationship, including interactions between any internally created terms (e.g. "friends ~ age_ego:reciprocity").
• predictor_selection = "quickpred" / netquickpred() — lean, correlation-screened predictor sets per target instead of everything.
• structural, net_dependence — structural zeros and logical constraints between networks.
• net_random_intercepts — social relations model random intercepts in the tie model.
• method — any of mice's numeric-response univariate methods, globally or per target.

Learn more

vignette("netimpute") is a worked introduction: a first imputation, convergence checks, pooling with mice::pool(), the predictor naming conventions, custom models, predictor selection, structural and between-network constraints, and how undirected networks are handled.

See ?netmice for the full argument reference, including the complete lists of dyad-level term names and implemented network measures.