The co-varying ties between networks and item responses via latent variables

1. Problem Statement

The paper addresses the challenge of modeling the relationship between social networks (specifically teacher advising relationships) and item response data (survey responses regarding school climate, such as satisfaction, policy influence, and student perception).

• Limitations of Existing Methods:
  • Separate Modeling: Analyzing networks and attributes separately ignores the potential dependence between them.
  • One-Directional Models: Existing approaches like Exponential Random Graph Models (ERGMs) with covariates or Social Influence models treat either the network or the attributes as fixed predictors, failing to capture the correlational nature of the relationship.
  • Summary Statistics: Reducing networks to summary statistics (e.g., node degrees) loses structural information (e.g., community structure).
  • Measurement Error: Two-step approaches (extracting latent factors first, then regressing) often ignore the uncertainty associated with the initial factor estimation.
  • Simplification of Item Responses: Treating questionnaire items as simple attributes without modeling them as indicators of latent psychological constructs oversimplifies the measurement process.

2. Methodology: The Joint Network and Item Response Model (JNIRM)

The authors propose JNIRM, a novel Covariance Joint Framework that simultaneously models network data and item response data using correlated latent variables.

A. Model Components

1. Network Component (Latent Space Model):

   • Uses a vector latent space model with a multiplicative UV effect to model directed networks.
   • For a sender $a$ and receiver $b$, the connection is modeled as $X_{a,b} = \delta + u_a^T v_b + e_{a,b}$.
   • $u_a$ and $v_b$ are $K$-dimensional latent position vectors for the sender and receiver, respectively.
   • Handles both continuous edge weights and binary networks (via a probit link where $X_{a,b}=1$ if the latent continuous measure $>0$).
   • Includes a within-dyad correlation parameter ($\rho$) to account for reciprocity.

2. Item Response Component (Latent Variable Model):

   • Models survey responses ($Y$) as a function of item slopes ($\alpha$), item intercepts ($\beta$), and person latent variables ($\theta$).
   • Equation: $Y_{p,i} = \beta_i + \alpha_i^T \theta_p + \epsilon_{p,i}$.
   • This treats item responses as indicators of latent psychological constructs (e.g., satisfaction, policy influence).

3. Joint Component (The Link):

   • The core innovation is assuming that the latent variables for the network ($u_p, v_p$) and the item responses ($\theta_p$) follow a joint multivariate normal distribution:
     $$(u_p, v_p, \theta_p)^T \sim MVN(0, \Sigma_{u\theta})$$
   • The covariance matrix $\Sigma_{u\theta}$ captures the dependence between network formation and item responses.
   • If the off-diagonal block $\Lambda_{u\theta}$ is non-zero, it indicates a dependence between the network structure and the latent traits measured by the survey.

B. Estimation and Inference

• Bayesian Inference: The model is estimated using a Gibbs sampler (MCMC).
  • Full conditional distributions are derived for latent variables ($U, V, \Theta$), covariance matrices, and hyperparameters.
  • For binary data, data augmentation is used to handle the latent continuous variables.
• Identifiability: The model addresses rotational indeterminacy by centering the latent variables and using the singular vectors of the posterior mean of the vector products ($UV^T$ and $\Theta A^T$) as estimates, rather than the latent variables themselves.
• Hypothesis Testing:
  • Multivariate Test of Independence: Tests $H_0: \Lambda_{u\theta} = 0$ using a likelihood ratio statistic (Wilks' $\lambda$).
  • Canonical Correlation Analysis (CCA): Used to interpret the nature of the dependence by finding linear combinations of network dimensions and item response dimensions that maximize correlation.

3. Key Contributions

1. Novel Framework: Introduces JNIRM, a fully specified statistical model that jointly estimates network and item response data, allowing for bidirectional inference and handling of missing data in both domains.
2. Latent-Level Integration: Unlike previous methods that link observed data, JNIRM links the latent constructs of both domains, providing a more robust understanding of the underlying mechanisms.
3. Improved Estimation: Demonstrates that joint modeling reduces parameter estimation error and improves model fit compared to separate modeling or two-step approaches, particularly by accounting for the uncertainty in latent factor estimation.
4. Handling Directionality: By using a covariance-based framework, the model avoids the need for prior knowledge regarding the direction of causality (network $\to$ attributes vs. attributes $\to$ network), treating the relationship as correlational.

4. Results

The methodology was applied to a dataset of 378 teachers across 14 schools in Nebraska, analyzing advice-seeking networks and school climate questionnaires.

• School 56 Case Study:

  • Significant Dependence: A statistically significant relationship was found between the advising network and item responses (Wilks' $\lambda = 0.084, p = 0.009$).
  • Complementarity Principle: CCA revealed that teachers tend to seek advice from those who are different from them regarding "satisfaction." Specifically, teachers with high satisfaction scores sought advice from those with lower satisfaction scores (and vice versa).
  • Model Fit: JNIRM outperformed the separate Adaptive Multilinear Embedding (AME) model in out-of-sample prediction (AUC: 0.800 vs. 0.727) and showed better recovery of network statistics (node degrees, dyadic dependence).
  • Item Covariances: JNIRM showed superior recovery of item covariances compared to separate modeling.

• Cross-School Summary:

  • Significant relationships were found in 3 out of 14 schools.
  • Across all schools, satisfaction and perceptions of students were most frequently related to advising networks, whereas perceived influence over policies showed weaker or inconsistent relationships.

• Simulation Study:

  • Simulations confirmed that JNIRM can reasonably recover model parameters (bias, variance, MSE) when sample sizes are moderate ($N=100$).
  • In sparse networks with small samples ($N=24$), slight biases in intercept and variance estimation were observed, but these were negligible for larger samples or non-sparse networks.

5. Significance

• Theoretical Advancement: JNIRM bridges the gap between Social Network Analysis (SNA) and Psychometrics (Item Response Theory), offering a unified framework for data where both relational and attribute data are latent and interdependent.
• Practical Application: The findings suggest that school climate interventions should consider the complementarity in advice-seeking; teachers may benefit from connecting with peers who have different perspectives on school satisfaction.
• Generalizability: The framework is applicable beyond education to any domain involving networks and attributes, such as neuroscience (brain connectivity vs. regional volume), economics (social media networks vs. consumer behavior), and public health.
• Methodological Rigor: By avoiding the "two-step" fallacy and explicitly modeling the covariance structure, JNIRM provides more reliable inferences for complex, high-dimensional data structures.