📈 economics

Social Interactions Models with Latent Structures

This paper proposes a debiased estimation and inference framework for heterogeneous social interactions with latent group structures, utilizing parametric bootstrapping and tailored clustering algorithms to effectively analyze peer effects, as demonstrated through simulations and an application to students' risky behaviors.

Original authors: Zhongjian Lin, Zhentao Shi, Yapeng Zheng

Published 2026-07-21

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhongjian Lin, Zhentao Shi, Yapeng Zheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand why people make certain choices, like whether to join a club or skip a class. In the world of economics and social science, this is called studying "social interactions" or "peer effects." It's the idea that your friends influence your decisions. But here's the tricky part: people aren't all the same. A teenager in a small, tight-knit town might be influenced by their friends very differently than a teenager in a huge, chaotic city. For a long time, scientists tried to solve this by assuming everyone was basically the same, or by trying to study every single group individually. But the first way ignores real differences, and the second way is like trying to count every grain of sand on a beach—it's too messy and the numbers get shaky. This paper steps into that messy middle ground, asking: "Can we find hidden groups of people who do act similarly, without knowing who is in which group beforehand?"

The authors, Zhongjian Lin, Zhentao Shi, and Yapeng Zheng, tackle this problem using a clever mix of math and detective work. They focus on a specific type of data: students in different schools. They know that students in the same school talk to each other, but they don't know if the "rules" of that peer influence are the same for every school. Maybe in some schools, friends push each other toward risky behavior, while in others, friends actually hold each other back. The paper introduces a new method to find these hidden "clusters" of schools automatically. Think of it like a smart sorting machine that looks at a pile of mixed-up socks and figures out which ones belong to the same pair, even if you didn't label them.

The researchers built a computer algorithm that does three main things. First, it guesses how likely a student is to make a choice based on their friends' actions. Second, it uses a special mathematical tool (called "C-Lasso") to group the schools together. If two schools have students who react to their friends in the same way, the algorithm puts them in the same "cluster." Third, it calculates the exact strength of that peer influence for each cluster. To make sure their math wasn't tricking them, they ran thousands of computer simulations to test their method. They found that their approach works really well, correctly identifying the hidden groups and measuring the influence accurately, even when the data is noisy.

When they applied this to real data from the "Add Health" study, which tracks thousands of students, the results were surprising. The algorithm split the 119 schools into two distinct groups. In one group, the peer effect was weak and statistically invisible—friends didn't seem to push each other toward risky behavior. But in the other group, the effect was strong and clear: friends significantly increased the chance of risky behavior. If the researchers had ignored these hidden groups and just looked at all the schools as one big blob, they would have missed this crucial difference entirely. They would have thought the effect was just "okay" everywhere, instead of realizing it was a powerful force in some places and non-existent in others. The paper proves that by finding these hidden structures, we can get a much clearer picture of how our friends really shape our lives.

Technical Summary: Social Interactions Models with Latent Structures

Problem Statement
The paper addresses the estimation and inference of heterogeneous social interactions (peer effects) in binary choice models. Standard approaches often assume parameter homogeneity across all groups (e.g., schools), which risks severe bias if the assumption is violated. Conversely, allowing for complete heterogeneity across all groups leads to unreliable estimates due to small sample sizes within individual groups. While some literature attempts to address this by pre-specifying cluster structures (e.g., by class size), external variables for such classification are not always available. Furthermore, existing methods for social interactions often fail to account for group fixed effects, leading to incidental parameter bias in slope coefficient estimation, particularly when using Nested Pseudo Likelihood (NPL) estimators in a panel data setting.

Methodology
The authors propose a framework that simultaneously identifies latent cluster structures and estimates structural parameters, incorporating group fixed effects and slope heterogeneity. The methodology integrates three core components:

Model Specification: The study utilizes a binary choice model with social interactions where individual utility depends on observable characteristics, group fixed effects, and the average conditional choice probability (CCP) of peers within an exogenously given network. The slope coefficients are assumed to be homogeneous within latent clusters but heterogeneous across them.
Estimation Algorithm (NPL-C-Lasso):
- Step 1 (First-step NPL): The Nested Pseudo Likelihood (NPL) algorithm is applied to each group independently to obtain preliminary estimates of parameters and CCPs. This avoids the computational burden of solving for CCPs simultaneously with classification.
- Step 2 (Classification): The Classifier-Lasso (C-Lasso) algorithm is employed to group the initial estimates into latent clusters. This involves minimizing a penalized profile likelihood with a multiplicative penalty term that shrinks group-specific estimates toward cluster-specific parameters, effectively determining cluster membership without prior knowledge of the number of clusters ( $K_0$ ).
- Step 3 (Post-Classification Estimation): Once clusters are identified, the NPL algorithm is re-applied to pooled data within each cluster to obtain final estimators.
Bias Correction and Inference:
- The paper identifies that high-dimensional incidental parameters (group fixed effects) induce asymptotic bias in the slope coefficients, a phenomenon not previously characterized for NPL estimators in fixed-point game models.
- To correct this, the authors adapt a parametric bootstrap method (Higgins and Jochmans, 2024) tailored for fixed-effect models. This method constructs debiased estimators and valid confidence intervals, overcoming the limitations of split-sample jackknife methods which are infeasible here due to the joint determination of CCPs.
- Inference is also developed for Average Partial Effects (APEs), which capture both direct covariate effects and network-mediated indirect effects via a social multiplier.
Model Selection: An Information Criterion (IC) is proposed to consistently select the number of latent clusters ( $K$ ) in practice.

Key Contributions

Theoretical: The paper characterizes the asymptotic bias of slope coefficients in binary choice social interaction models with group fixed effects and a fixed-point structure. It establishes the asymptotic normality of the NPL estimator and proves that the proposed parametric bootstrap yields an asymptotically unbiased estimator with correct coverage rates.
Methodological: It introduces a multi-step algorithm that decouples the classification step (C-Lasso) from the computationally intensive NPL steps, enhancing stability and tractability. It demonstrates classification consistency (groups are assigned to true clusters with probability approaching one) and the consistency of the Information Criterion for selecting the number of clusters.
Empirical Application: The method is applied to the Add Health dataset to study risky behaviors among students. The algorithm successfully identifies two latent clusters of schools.

Results

Monte Carlo Simulations: The simulations confirm that the proposed method performs well in finite samples. The C-Lasso accurately identifies the number of clusters and classifies groups with high accuracy as group size increases. The debiased estimators significantly reduce bias and root mean squared error (RMSE) compared to pooled or original estimators, and the bootstrap confidence intervals achieve coverage rates close to the nominal 95% level.
Empirical Findings: In the application to student risky behaviors:
- The algorithm classifies 119 schools into two distinct clusters (43 and 76 schools).
- Heterogeneity: Peer effects are statistically significant and economically meaningful in the second cluster (coefficient $\approx 0.688$ ) but are not statistically different from zero in the first cluster.
- Variable Effects: The impact of intelligence on risky behavior differs by cluster (positive in the second cluster, negative in the first), a nuance obscured in homogeneous models.
- Policy Implications: Ignoring the latent structure leads to underestimating peer effects in the relevant cluster and erroneous conclusions regarding covariate effects (e.g., intelligence).

Significance and Claims
The paper claims to fill a gap between the literature on peer effects and panel data econometrics by providing the first formal results on valid statistical inference for peer effects in the presence of group fixed effects and latent structures. It argues that relaxing the classical parsimonious assumption of homogeneity is necessary to meet the challenges of big data and real-world complexities. The authors emphasize that their approach provides a practical pipeline for detecting latent structures and conducting inference on economically meaningful objects (like APEs) without requiring pre-specified cluster information. The study concludes that researchers must account for latent structures in social interaction models to avoid misleading inferences.

Technical Summary: Social Interactions Models with Latent Structures

More like this