---
config:
htmlLabels: false
---
%% htmlLabels:false is load-bearing, not cosmetic. With HTML labels mermaid
%% sizes each box from getBoundingClientRect(), which reveal's slide-scaling
%% transform multiplies -- every box comes out ~40% too narrow and every label
%% is clipped. SVG labels are measured with getBBox(), which the transform
%% does not touch. (Labels still wrap at 200px either way.) The price is that
%% mermaid then never centres them, so style.scss re-anchors the text; see
%% "Research-map mindmap" there. No themeVariables.fontSize either: the SVG is
%% scaled to the slide, which grows boxes and text together anyway.
%%
%% Mindmaps have no `classDef` -- the palette for `branch` / `today` also
%% lives in style.scss. Two syntax traps: `:::class` must sit on its OWN line
%% under the node it decorates (`node:::x` is a parse error), and a bare node
%% label cannot contain parentheses, hence the quoted `["..."]` form on the
%% leaves. The layout is force-directed: identical on every load at a given
%% window size, but it can settle into a different arrangement on a screen of
%% a different shape. Check it on the room's projector, not just your laptop.
mindmap
root(("Statistical<br>computing and<br>scientific<br>software"))
NS(Network science)
:::branch
E1["Applied ERGMs"]
:::today
E2["Other graphical<br>models"]
E3["ERGM methods"]
E4["Social Influence"]
PH(Public health)
:::branch
P1["Measles"]
:::today
P2["Agent-Based Models"]
P3["Mechanistic Machine<br>Learning"]
P4["Economic Impact"]
OTH(Other domains)
:::branch
O1["Genomics"]
O2["Quantum<br>Computing"]
From Supercomputers to Health Departments
Roux Institute + Network Science Institute
Northeastern University
2026-09-14
On February 5th, 2025, public health officials sound the alarm: a measles outbreak has begun in Gaines County, Texas.
Can you build a model of measles spreading in schools in 3
monthsweeks?
Results: Shiny App
You can access the app at https://ggvy.cl/shiny/measles
Results: 1,200 letters
A personalized letter for every school in Utah
- Inputs: enrollment and up-to-date vaccination coverage
- Output: outbreaks with and without the quarantine process
The workflow is free and public on GitHub. Minnesota (MADMC, another InsightNet site) is running it; other sites are evaluating it.
The measles model
Robust C++ core. Implemented as an extension of
epiworld(Vega Yon 2022). Themeaslestool is available as an R package on GitHub and CRAN (Vega Yon 2026d).Discrete time. Each day is one time step.
Homogeneous mixing. Contact rate-based interaction.
Full disease progression. Times are geometrically distributed.
Quarantine process. Probability of detection and agents’ willingness to follow it.
Two levers: vaccination coverage and the quarantine process.
The engine: epiworld
- A fremework for fast, scalable agent-based models.
- Written in C++: 19K lines of code + 9K lines of tests.
- Multi/evolving-disease models.
- 15K downloads on CRAN (
epiworldR). - Parallel computing via OpenMP out of the box.
It is also available in R (Meyer and Vega Yon 2023), Python (Vega Yon and Banks 2026), Shiny (Meyer and Vega Yon 2025), and WebAssembly (Banks and Vega Yon 2024).
More about its <speed>
From one school to a community
- Using the school model as a baseline, we implemented a community-scale version1.
Multiple groups; heterogeneous mixing between them; group-specific vaccination; contact-tracing-based quarantine.
The R package
measles(Vega Yon 2026d) implements these models on top ofepiworldR. It is available on GitHub and CRAN and is under review at JOSS.
Utah-Arizona border outbreak
- Largest outbreak in Utah.
- 585 confirmed cases associated with that outbreak.
- A population of about 6,000 people.
- One of the areas with the lowest vaccination coverage in the state, at about 50% (see <details>).
- A complex relationship with public health.
The number of confirmed cases does not align with the data (Marye et al. 2026)
Results: Utah-Arizona outbreak
- Utah DHHS estimated outbreak size using omics.
- We used epiworldR+measles, incorporating (a) school vaccination coverage, (b) census age distributions, and (c) POLYMOD/Epistorm contact patterns.
- Performed a sensitivity analysis accounting for parameter uncertainty.
| Source | Infections1 |
|---|---|
| Confirmed cases | 585 |
| ABM estimates | ~2,226 |
| Genomic estimates | ~2,051 |
On the bright side… 
Both methods put the number closer to 2,000 than to 600, implying that roughly three in four infections went uncounted.
Where else the model goes
Healthcare-Associated Infections
Worldwide, “around 1 in 10 patients is affected by Healthcare-Associated Infections” (HAIs) (Pittet 2026).
“[I]n the [US], the CDC reported nearly 687,000 HAIs in acute care facilities, resulting in 72,000 associated deaths” (Odoom and Donkor 2025).
Contact networks are a key mechanism for the spread of HAIs (Curtis et al. 2013).
Our Data: Contact networks
ERGMs in resident–healthcare provider networks (Vega Yon, Zhang, et al. 2026) · Social Networks (in press)
Contact networks inside long-term care facilities: healthcare personnel on one side and residents on the other (bipartite).
93 bipartite networks from 47 facility units, ranging in size from 10 to 100 nodes.
Healthcare personnel: certified nurse assistants (CNAs), registered nurses (RNs), physical therapists (PTs), occupational therapists (OTs), respiratory therapists (RTs), and other healthcare providers.
Residents: dialysis, wound care, bedridden, ventilator, MDRO status.
Figure 1 from Vega Yon, Zhang, et al. (2026) drawn using netplot (Vega Yon 2026e)
The key questions
We wanted to address the following questions regarding the contact networks:
- Do different types of healthcare personnel differ in their contact patterns?
- Do residents with different care needs differ in their contact patterns?
- Can we say anything about cohorting?
- Do we observe any differences between facilities?
Exponential-family Random Graph Models (ERGMs) provide a natural framework for answering these questions.
Challenges in our data
Projectivity
An estimate for a network of size n does not necessarily apply to a network of size n' \gg n.
SOLUTION
Use size interactions for terms that are sensitive to network size (as in Krivitsky et al., 2011).
Repeated nets
The networks are not IID: some were observed more than once.
SOLUTION
Use a block bootstrap to account for the repeated measurements.
Multicollinearity
Some statistics are highly correlated, which can lead to numerical instability and convergence issues.
SOLUTION
Use an affine reparameterization of some terms.
This is one of the first applications of multilevel bipartite ERGMs.
More about the <terms> used in the model.
Results: HCP-Resident Networks
Results from 200 block-bootstrap replicates of the pooled ERGM fit:
Network size interaction
Strong evidence
sparser, not busier
A unit with twice as many residents does not have twice as many contacts
Shared partners
Strong evidence
spreads out
Staff who share a resident rarely share many other residents
Role mixing
Strong evidence
different roles pair
Two staff members caring for the same resident are usually different kinds of workers
Care need
Strong evidence
sicker, more contact
Residents who are sicker – on a ventilator, bedridden, or in need of wound care
Using these results, we can simulate realistic contact networks for epidemic modeling.
See more details <link>
R21: Conditional ERGM
How many networks do you need to detect a homophily effect? –
(Vega Yon 2023 in JASA)
More about this project <here>
Recap
Dive 1: Agent-Based Models
Fast and scalable agent-based models for measles in schools and communities.
What: Online tool for simulating school outbreaks, school-level reports distributed across the state, and a community-scale model for the Utah-Arizona border outbreak.
Takeaway: We can build models and tools that are accessible to non-academic audiences and can inform policy.
Dive 2: Network Models
Advances in exponential-family random graph models (ERGMs)
What: A new application (with methodological advances) of ERGMs to bipartite networks in healthcare, plus a glimpse into conditional ERGMs (R21).
Takeaway: ERGMs provide a flexible framework for characterizing real-world networks. Downstream applications can leverage this, e.g., epidemiological simulation.
One more thing…
Three completely open (and reproducible) book projects:
All three are also available in Spanish!
Where this goes
Statistical
inference
- Statistical inference in complex systems: theory and tools.
- ERGMs: the same, plus extensions beyond SNA.
Hybrid
models
- AI and machine learning: methods and scientific products.
- Quantum computing: explore opportunities (seed grant at Utah).
Tools and partnerships
- Translational science: more collaboration with public health.
- Innovation: explore private-sector and academic partnerships.
Collaborators
Mentees
-
Sima Najafzadehkhoei
-
Anibal Olivera
-
Olivia Banks
-
Derek Meyer (former)
-
Andrew Pulsipher (former)
DHHS
-
Randon Gruninger
-
Matthew Mietchen
-
Leisha D. Nolen
-
Abbey Marye
-
Joshua Kelley (former)
Utah
-
Bernardo Modenesi
-
Lindsay T. Keegan
-
Karim Khader
-
Andrew T. Pavia
-
Matthew Samore
-
Damon Toth
-
Jake Wagoner
-
Delaney Thornton
-
Lindsay D. Visnovsky
-
Kristina Stratford
-
Richard Nelson
CDC
-
Damon Bayer
-
Dylan Morris
-
Samuel Rosenblatt
-
Emily Pollock
-
Dina Mistry
Other institutions
-
Thomas Valente (USC)
-
Kyosuke Tanaka (Aarhus University)
-
Kayla de la Haye (UCLA)
-
Paul Marjoram (USC)
-
Sarah Piombo (Harvard)
-
Kaitlyn Johnson (LSHTM)
Thank you!
From Supercomputers to Health Departments
---
config:
htmlLabels: false
---
%% htmlLabels:false is load-bearing, not cosmetic. With HTML labels mermaid
%% sizes each box from getBoundingClientRect(), which reveal's slide-scaling
%% transform multiplies -- every box comes out ~40% too narrow and every label
%% is clipped. SVG labels are measured with getBBox(), which the transform
%% does not touch. (Labels still wrap at 200px either way.) The price is that
%% mermaid then never centres them, so style.scss re-anchors the text; see
%% "Research-map mindmap" there. No themeVariables.fontSize either: the SVG is
%% scaled to the slide, which grows boxes and text together anyway.
%%
%% Mindmaps have no `classDef` -- the palette for `branch` / `today` also
%% lives in style.scss. Two syntax traps: `:::class` must sit on its OWN line
%% under the node it decorates (`node:::x` is a parse error), and a bare node
%% label cannot contain parentheses, hence the quoted `["..."]` form on the
%% leaves. The layout is force-directed: identical on every load at a given
%% window size, but it can settle into a different arrangement on a screen of
%% a different shape. Check it on the room's projector, not just your laptop.
mindmap
root(("Statistical<br>computing and<br>scientific<br>software"))
NS(Network science)
:::branch
E1["Applied ERGMs"]
:::today
E2["Other graphical<br>models"]
E3["ERGM methods"]
E4["Social Influence"]
PH(Public health)
:::branch
P1["Measles"]
:::today
P2["Agent-Based Models"]
P3["Mechanistic Machine<br>Learning"]
P4["Economic Impact"]
OTH(Other domains)
:::branch
O1["Genomics"]
O2["Quantum<br>Computing"]
Results: How fast is epiworld?
epydemic here is Dobson’s network ABM library – not to be confused with epydemix, Epistorm’s compartmental + ABC-calibration package!
Go back to <compute>
Low vaccination coverage
We estimated overall MMR coverage of ~50% in the region, with close to 3,000 susceptible individuals in the area.
Go <back>
Calibration is the machinery under this claim, and it is its own line of work – epiworldRcalibrate (Najafzadehkhoei et al. 2026) and the ML-surrogate paper (Najafzadehkhoei et al. 2025). One sentence here; the agenda version is in “Where this goes”.
Increased reliability
Exact statistics significantly reduce error rates when estimating ERGMs.
This includes models that, in principle, are not degenerate—that is, models that generate networks that are neither very dense nor very sparse.
The other regime: exact computation
Using the exact approach makes the estimation process orders of magnitude faster.
Compared with other methods, ergmito can be up to 100 times faster at estimating pooled models.
Exponential random graph models for little networks (Social Networks, 2021) (Vega Yon, Slaughter, et al. 2021)
Model Terms
tabulergm R package (Vega Yon 2026g)Go back to the <data challenges>
ergmitos
Go back to ergms (link)
Bipartite ERGM table
Go back to the key findings (<link>)
Conditional ERGMs
R21 (NLM): “Advancing Exponential-Family Random Graph Model Estimation and Power Analysis for Biomedical and Public Health Research”
\begin{aligned} {\mathbb{P}_{}\left\{\mathbf{Y}= \mathbf{y}; \boldsymbol{\theta}\right\} } = \hphantom{\text{\hspace{3.5cm}}}\\% &= \Prcond{}{\Graph = \graph}{\sufstats{l}{\Graph} = z; \params} \times \Pr{}{\sufstats{l}{\graph} = z; \params}\\ \underbrace{\frac{% \text{exp}\left\{{\boldsymbol{\theta}_{-l}}^\mathbf{t}s_{-l}\left(\mathbf{y}\right)\right\}% }{% \eta_{s_l = z,-l}\left(\boldsymbol{\theta}{}\right) }}_{\text{Conditional ERGM}} \times % \underbrace{\frac{\eta_{s_{k}\left(\mathbf{y}\right)=x}\left(\boldsymbol{\theta}{}\right)}{\eta_{}\left(\boldsymbol{\theta}{}\right)}}_{\text{Marginal ERGM}} \end{aligned}
Go back <here>
Conditional ERGMs (cont.)
How good are conditional ERGM estimates as an initialization point?
Go back <here>