Code for "Exponential-Family Random Graph Models in Resident-Healthcare Provider Networks: An Application using data from Long-term Healthcare Facilities in the United States"

Analysis workflow for the paper on bipartite contact networks between health care providers (HCPs) and residents in long-term care facilities in the United States.

The publication target is the supplemental materials document:

quarto render models/05-supplemental-materials.qmd

models/05-supplemental-materials.qmd is the only document needed to reproduce the published supplement. It reads prepared network objects and saved model fits rather than refitting anything: the upstream ERGM and bootstrap steps take days of cluster time. The other Quarto documents are the intermediate steps that produced those saved objects, and are kept for provenance.

Analysis documents

File	Purpose	Key output
data/model_data.Rmd	Builds the 93 bipartite facility-unit networks from the raw interaction records. The raw CSV is not distributed with this repository.	data/network93_f2.RData
models/01-descriptive.qmd	Descriptive analysis of resident care-need attributes, degree, and network size.	figures only
models/02-individual_ergms.qmd	Exploratory single-network ERGMs, one fit per network, used to check which terms are estimable.	models/02-individual_ergms_fits/
models/ergms_filtering_nets.qmd	Early exploration of which networks to keep; documents dropping the 7 networks with a single HCP.	none (report)
models/03-pooled-ergms.qmd	Main model selection. Fits the pooled multi-network ERGMs term by term with ergm.multi::N().	models/03-pooled-ergms-final-results.rds, models/03-pooled-ergms-rds/
models/20260520-ergm-fit-diagnostics.qmd	Investigation of convergence failures and between- vs within-state heterogeneity, motivating the final specification.	none (report)
models/04-boot-ergms.qmd	Bootstrap of the baseline model using MC-MLE. Its cache is read by 04b for the SA-vs-MC-MLE comparison.	models/04-boot-ergms-cache/
models/04b-boot-ergms.qmd	Block bootstrap of the baseline model using stochastic approximation, compared against the MC-MLE run above.	models/04b-boot-ergms-results.rds
models/04c-boot-ergms.qmd	Block bootstrap of the model with state-level intercepts (facility_state).	models/04c-boot-ergms-results.rds
models/04d-boot-ergms.qmd	Block bootstrap of the final GWDSP × log(size) model. Supplies the standard errors reported in the paper.	models/04d-boot-ergms-results.rds
models/05-supplemental-materials.qmd	Final supplement. Observed statistics, GOF, MCMC diagnostics, bootstrap standard errors, and the alternative-specification comparison.	models/05-supplemental-materials.docx
models/helpers.R	Shared code: set_ergm() network preparation, boot_ergm() bootstrap driver, facility_cluster_id() blocking variable, GOF plotting, texreg wrappers, and result caching.	—

The 04* documents share one bootstrap driver, boot_ergm() in models/helpers.R. Each replicate is cached as an individual .rds under models/04*-boot-ergms-cache/, so an interrupted run resumes where it stopped.

04b, 04c, and 04d run a block bootstrap: they pass cluster_id to boot_ergm(), which switches the driver from resampling individual networks to resampling whole facilities, so every network belonging to a drawn facility enters the replicate together and within-facility correlation is respected. The blocking variable is built once by facility_cluster_id() in models/helpers.R, which blocks on state + Fac_Name and maps the 93 networks onto 24 facilities. 04 is the older MC-MLE run and still resamples individual networks; it is kept only as the comparison column in 04b.

Because the resampling scheme is part of what each cached replicate encodes, a cache written by an earlier naive run is not reusable: clear models/04*-boot-ergms-cache/ before re-running these documents, or the stale replicates will be restored instead of recomputed.

Inputs required by the supplement

These files must be present to render models/05-supplemental-materials.qmd:

• data/network93_f2.RData — the 93 prepared networks.
• models/04b-boot-ergms-results.rds, models/04c-boot-ergms-results.rds, and models/04d-boot-ergms-results.rds — block-bootstrap summaries; 04d also supplies the final fitted model.
• models/03-pooled-ergms-rds/res4_gwdsp_var_bnodematch.rds, res6_full_and_facility_mcmle.rds, and res6_full_and_logn_gwdsp.84mcmle.rds — the three specifications in the comparison table.
• models/05-supplemental-materials-gof.rds — cached GOF object; recomputed with gofN() and saved if absent.

Model objects and prepared data are .gitignored, so they travel outside git.

Environment

The analysis runs in the container defined by .devcontainer/Containerfile (R 4.5, Quarto 1.9.35, ergm.multi, data.table, texreg, ggplot2 and friends, and tabulergm from GitHub).

make container_build
make container_run
make render            # quarto render models/05-supplemental-materials.qmd

On CHPC, render.slurm runs the same target through Singularity. The 04* bootstrap documents use parallel::makeForkCluster(); lower the worker count before running them anywhere other than a compute node.

Repository layout

data/       prepared networks and the data-preparation source
models/     analysis documents, helpers.R, and saved model objects
.devcontainer/  container definition
Makefile    render, container, and Singularity targets
render.slurm    CHPC entry point

Superseded analyses and side projects (abm/, chong/, fig/, LTCF_final_report/, summary-stats.R) are no longer tracked. They may still exist in a local working copy, and remain in the git history.

Note on terminology

Saved model objects and coefficient names use the legacy labels n_patients and n_hcp. Throughout the manuscript these mean residents and health care providers, respectively.