R Package iglm: Regression under Interference in Connected Populations
summary
The gist
Regression under interference in connected populations is studied through a comprehensive R package called iglm which implements a framework for analyzing relationships among predictors and outcomes…
In short
The iglm R package introduces a comprehensive regression framework to study relationships among predictors and outcomes in connected populations where data points are dependent. It allows researchers to model spillover effects by capturing local attribute dependencies and connection structures, offering scalable solutions with provable theoretical guarantees superior to existing methods.
Key concepts
- Regression under Interference
- This refers to statistical modeling when the attributes being measured (like X and Y) are not independent due to their connections in a system. The framework specifically addresses how the dependence between connected units influences the relationship between predictors and outcomes.
- Pseudo-likelihood Functions
- These functions are used as a basis for model estimation. They are sums of exponential-family loglikelihood functions that help researchers find the best parameters ($ heta$) for the regression models by maximizing them, even when dealing with complex joint probability distributions.
- Scalability and Provable Guarantees
- IGLM achieves scalability by solving optimization problems using methods like Minorization-Maximization instead of slow simulation techniques. It provides provable theoretical guarantees for consistency and convergence rates, meaning the model's performance can be mathematically verified based on a single observation.
- User-Defined Model Terms
- This feature allows users to extend the package by adding custom statistical terms to the regression model. Users define these terms by decomposing functions ($g_i$ and $h_{i,j}$) and writing C++ code to compute necessary change statistics, enabling highly tailored modeling.
Terminology used across episodes
This episode discusses
- R Package iglm: Regression under Interference in Connected Populations · Paper Radio
- Exponential-family Random Network Models
The paper
R Package iglm: Regression under Interference in Connected Populations · Read on arXiv
School of Computer Science and Statistics Trinity College Dublin · Department of Statistics The Pennsylvania State University
We introduce R package iglm, which implements a unifying regression framework for studying relationships among binary, count-, and real-valued predictors and outcomes in connected populations with interference due to spillover, contagion, and other complex phenomena. On the computational side, the regression framework provides scalable methods that can be applied to small and large data sets, by solving a convex optimization program based on pseudo-likelihoods via accelerated Minorization-Maximization algorithms. On the statistical side, the regression framework comes with provable theoretical guarantees for statistical estimators, including consistency results and rates of convergence for maximum pseudo-likelihood estimators. To increase the versatility of iglm, users can add custom-built model terms. We showcase iglm using data on hate speech on the social media platform X and communications among students.
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "R Package iglm".
Jane: Here's the gist of the paper. Regression under interference in connected populations is studied through a comprehensive R package called iglm which implements a framework for analyzing relationships among predictors and outcomes under dependence in connected systems The implemented regression framework facilitates the study of spillover and other phenomena in connected populations and has important advantages over existing packages, among them scalability and provable theoretical guarantees.
Tom: Pruning redundant reasoning tokens The paper introduces the R package iglm, which provides a comprehensive regression framework for dependent attributes (Xi, Yi) and connections Zi,j with fixed and random designs.
Jane: This framework is built on a probability measure dominated by a σ-finite measure νX,Y,Z The joint probability density function of (X, Y, Z) is given by a form involving unit-dependent functions gi(xi, yi) and pairwise functions hi,j(xi, xj, yi, yj, z).
Tom: This framework can capture relationships among attributes (Xi, Yi) of units i via gi(·) and dependence among attributes (Xi, Yi) and (Xj, Yj) of connected pairs of units i and j via hi,j(·).
Title and authors: Tom: So this paper is about this new R package called iglm which studies regression under interference in connected populations. Jane, what's the big picture here?
Jane: Well, it sounds like they’ve built a whole framework for looking at how things influence each other when they are connected. It’s all about spillover effects and how those connections change the outcomes for different groups.
Lu: Exactly. The authors introduce this comprehensive regression framework that handles dependent attributes and connections with both fixed and random designs, which is a big step forward in modeling these kinds of systems > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg2>
Meng: But what does this mean for us in terms of actual use? I mean, I'm not sure how scalable this thing actually is when we look at real traffic and massive datasets.
Tom: That's a fair question, Meng. The paper says the framework solves a convex optimization program based on pseudo-likelihoods using Minorization Maximization and Quasi Newton algorithms > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg1>
Jane: Right, so they’re trying to make it work for both small and large data sets, which is important because we always run into the issue of needing time-consuming Markov chain Monte Carlo methods when things get big.
Lu: And they claim scalability as one of the main advantages, which they describe in Section one point three > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3> <ref:2604.22791#pg1,R Package iglm: Regression under Interference in Connected Populations>
Meng: Scalability is nice, but what about the guarantees? When you're dealing with these complex dependencies, I worry about getting estimates that are totally unreliable.
Tom: They claim provable theoretical guarantees for the regression models and methods they implement > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg1>
Jane: That’s a strong claim, Tom. What does that actually mean for someone trying to use this package? Does it mean we don't have to guess how reliable our estimates are?
Lu: It means they have consistency results and rates of convergence based on just a single observation of dependent attributes and connections > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3>
Meng: Consistency is one thing, but what about when the number of weights gets really big, like when we try to capture unobserved heterogeneity through unit-dependent weights?
Tom: That’s where they point out a problem with standard asymptotics; those can be misleading when you have dependent attributes and connections with an increasing number of weights > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3>
Jane: So they are specifically addressing the issue that standard methods break down when the number of weights goes up, which is a key hurdle for modeling these connected systems.
Title and authors: Lu: They’re showing that conclusions based on those standard asymptotics can be misleading in non-Bayesian approaches, and this framework helps quantify uncertainty about estimators using the exact covariance matrix rather than relying on standard assumptions > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3>
Meng: Quantifying uncertainty with the exact covariance matrix sounds better than just assuming something about how the error looks asymptotically. But how do you actually get that simulation to run without taking forever?
Tom: They use Markov chain Monte Carlo simulations, specifically Gibbs sampling for attributes and Metropolis Hastings algorithms with TNT proposals for simulating connections > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3>
Jane: So they’re using these simulations to approximate the uncertainty about the maximum pseudo-likelihood estimators, which is a bit different from just relying on standard asymptotic theory.
Lu: And for flexibility, they allow users to add custom-built model terms by decomposing functions like gi and hi in a specific way > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg1>
Meng: Custom terms sound powerful, but that implies a lot of manual work on the user's part. How easy is it for someone to actually build one of these custom terms?
Tom: The recipe involves decomposing those functions into change statistics, defining those statistics, writing a C++ function to compute them, and then registering it as a term in the package > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg1>
Jane: So it’s not just plug-and-play; you have to be able to write some custom code for the change statistics before you can use that flexibility. That adds a layer of complexity.
Lu: And they specifically show how this framework can capture relationships among attributes of units through gi and dependence between attributes of connected pairs through hi,j > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg2>
Meng: So, for someone deploying this in production, the main concern is the computational cost of that optimization problem they solve with Minorization Maximization and Quasi Newton methods.
Tom: They say it allows them to handle up to ten thousand units and large scale simulations without needing those slow Markov chain Monte Carlo methods > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg1>
Jane: That’s a big win for practical applications, Tom. If you can run these models on ten thousand units efficiently, that opens up a lot more real-world possibilities than what was possible before.
Lu: The potential here is that this framework helps construct models with local dependence among attributes and connections in overlapping neighborhoods > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3>
Meng: So, the practical implication is that we can build richer models for connected populations by looking at local structures, not just treating everything as independent.
Tom: And they even show how you can interpret the conditional distributions of outcomes and connections using GLMs > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg2>
Title and authors: Jane: That’s interesting because GLMs are pretty standard tools in data science, so if you can connect the framework to something familiar like a logistic regression model for connections, it makes sense for people who aren't deep into the interference theory.
Lu: The results suggest that models should include spillover terms to better predict outcomes and there’s a tendency to connect with others who are similar in terms of gender, race, and state > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg2>
Meng: So what does that mean for the user? It means if you're studying connections, you should explicitly try to model those spillover terms because the data seems to support that.
Tom: The conditional distribution of the connection Zi,j is shown as a Bernoulli distribution with a mean defined by an exponential family parameter > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg2>
Jane: So, to wrap up this section, this paper is introducing R Package iglm which gives us a scalable and theoretically sound way to model relationships where things influence each other within connected groups.
Lu: It’s a comprehensive regression framework for dependent attributes and connections with both fixed and random designs > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg2>
Meng: And the main takeaway for deployment is that you get scalable optimization methods alongside provable guarantees, which is what we need to ship models that actually work reliably at scale.
Tom: It’s about making sure our models don't just look good on paper, but that they actually have solid mathematical backing when we push them onto real data > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg1>
Jane: That seems to be the core message, Tom. It’s not just about having a fancy framework, it’s about having a framework that's robust enough for real-world connected populations.
Lu: And the future work is clearly pointing toward expanding this into more complex structures where these local dependencies become even more intricate > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3>
Meng: I just hope they keep pushing on the scalability front, because if it stays slow, it's just a cool theoretical exercise and not something we can actually deploy.
Tom: That’s the engineering reality, Meng. The trade-off between complexity and speed is always there when you build these types of frameworks > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg1>
Jane: And that’s what we'll be looking at next time, because we need to see if this framework can handle even more complicated structures than just overlapping neighborhoods.
Lu: Right, let's keep an eye on how they extend the structure-based transfer learning ideas mentioned in other papers to this iglm framework > R Package iglm: Regression under Interference in Connected Populations <ref:2604.22791#pg3>
The paper's summary: Tom: So, we just talked about how this new R package, iglm, handles relationships when things influence each other in connected groups. Now let's talk about what they actually propose in their summary section and what that means for us right now on air.
Jane: Okay so the summary is saying they’ve built a whole framework for dependent attributes and connections with both fixed and random designs using a specific probability measure called nu X, Y, Z. It sounds like the core idea is that we can capture how things are related by looking at local dependencies within neighborhoods of connected units.
Lu: Exactly. They describe the joint probability density function involving functions gi for the attributes themselves and hi for the connections between pairs of units i and j. This means they’re not just looking at isolated effects; they’re modeling how the connection between unit i and unit j affects their respective attributes.
Meng: So, when I think about shipping this to production, what's the practical implication of that probability measure thing? Does it mean we can actually build a model that accurately predicts spillover effects without resorting to super slow simulations?
Tom: They claim scalability is one of the main things. They solve a convex optimization problem using methods like Minorization Maximization and Quasi Newton algorithms instead of needing those time-consuming Markov chain Monte Carlo methods. That’s a big deal for large datasets, Meng.
Jane: And they back that up with provable theoretical guarantees for the models they implement, which means we know there are actual mathematical rules about how accurate these regression models will be, even when things get complicated.
Lu: That consistency result based on just a single observation of dependent attributes and connections is what I find interesting. It gives us confidence in the structure of the model without needing massive amounts of data to prove it works.
Meng: So for me, that means if we use this iglm package on, say, ten thousand units—which is a lot—we can get results relatively quickly instead of waiting days for a simulation to converge. That makes deployment much more feasible.
Tom: Right, so the numbers are there and they're trying to make sure the math holds up when we scale things up.
Jane: And they’re showing how you can interpret these models by looking at conditional distributions, using standard tools like GLMs for connections, which makes it easier for people who aren't deep into this specific interference theory to get the results.
Lu: They also show that their joint model is better than just using separate GLMs for each attribute and connection on its own because it explicitly includes the spillover terms that matter.
Meng: So, if I’m looking at my own models, I should be thinking about adding those spillover terms because the paper suggests they are important for accurate prediction. That’s a concrete thing I can act on.
Tom: And they even give us some hints on what kind of connections we might see—like a tendency to connect with others who share traits in gender or race, which gives us something to test against.
Jane: So the big picture is that iglm moves us beyond simple independent predictions and lets us model the complex social or biological networks where influence flows between connected parts.
Lu: It’s about building models that reflect the reality of how systems are actually structured, not just a simplified version of them. That structure is what matters here.
The paper's improvements: Tom: Alright, so we’re moving on to how the authors suggest improving this framework they built for regression under interference in connected populations. Basically, they're talking about taking things further than just having a basic model setup.
Jane: So what are these specific improvements? I mean, it sounds like the package itself is pretty good already, but are there new ways to make the math stronger or the application wider?
Lu: They focus on adding structure to the models by leveraging neighborhoods. They suggest using this local dependence among attributes and connections within overlapping neighborhoods. This means instead of just looking at one unit in isolation, you can model how it’s related to its neighbors in a specific spatial arrangement.
Meng: So they're suggesting we use geographical or network proximity to build our models, right? That seems like it could be useful for things like traffic flow or even disease spread where local connections matter most.
Tom: Exactly. And this neighborhood structure helps them construct models that account for the local dependencies in the attributes and connections simultaneously, which is a key part of capturing spillover effects more accurately.
Jane: Then there’s the way they handle custom terms. They show how users can add their own specific model terms by decomposing those functions into change statistics, which lets them incorporate unique knowledge directly into the regression framework.
Lu: That decomposition process is quite detailed; you have to specify those change statistics, like delta x or delta y, and then write a C++ function to compute them for the package to register it as a term. It’s flexible but requires some specific coding work from the user side.
Meng: That sounds powerful, but I gotta ask about the cost of that customization. If you want a custom term, how much extra development time are we looking at compared to just using a standard formula?
Tom: They do acknowledge that adding custom terms involves writing code and decomposing those functions according to their specific equations. It’s not just dragging and dropping a new variable into a formula object.
Jane: And they emphasize that the framework lets you compare your joint probability model against simpler models, like standard GLMs for individual attributes, to see exactly how much predictive value those connection terms add on top of just looking at each thing separately.
Lu: That comparison is crucial because it directly shows the added value of modeling the connections together versus treating them as independent effects. It’s a way to prove why the interference model is necessary.
Meng: So, if I'm trying to decide whether to build a complex interference model or stick with simpler GLMs, this paper gives me a way to quantify that added predictive gain in terms of how well the joint prediction performs compared to the separate ones. That’s a solid metric for my deployment decisions.
Tom: It gives you that comparison using predictions, which is tangible data rather than just theoretical math.
Jane: So we see improvements in structure, flexibility with custom terms, and a clearer way to compare the complexity of the joint model versus simpler models.
Lu: And they’re pointing toward future work that involves expanding these ideas into even more complex structures where these local dependencies become even more intricate across larger areas.
Conclusion: Tom: So to wrap up, we’ve seen that the R package iglm provides a scalable way to model regression under interference in connected populations using optimization instead of slow simulations, with provable theoretical guarantees built in.
Jane: And they show us how to use neighborhood structures and custom terms so that we can build really rich models that capture those spillover effects we talked about earlier.
Lu: Yeah, the framework is powerful because it explicitly separates the attribute relationships from the connection relationships, allowing for a much clearer understanding of what drives the prediction.
Meng: For deployment, this means we have a tool that can handle big data efficiently without getting bogged down in computationally expensive methods like Markov chain Monte Carlo simulations.
Lalam: I’m confident that because they've addressed both scalability and theoretical consistency, the iglm package is a reliable way to move past simpler models when dealing with complex social or biological networks.
Tom: It really makes sense when you consider what this means for real-world applications, whether it’s modeling how information spreads on a network or understanding human behavior in connected groups.
Jane: So we’ve seen that iglm offers a concrete method to build models that respect the local dependencies within connected systems.
Tom: It’s a solid piece of engineering for anyone who wants to move beyond basic regression when dealing with interconnected data structures.
Meng: It’s good stuff, but the real test is seeing how robust it is when you throw messy, real-world traffic or biological data at it and see where those assumptions break down.
Lu: That’s exactly where the next steps are going to be, testing those boundary conditions on even more intricate structures that go beyond simple overlapping neighborhoods.
Lalam: I'm looking forward to seeing how these structure-based ideas transfer into even more complicated systems using this framework as a foundation.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates