Inferring Structural Characteristics of Networks With Strong and Weak Ties From Fixed-Choice Surveys

Naghmeh Momeni, Michael G. Rabbat

Mapping offline social networks is a notoriously difficult task. Unlike Facebook or Twitter, where every interaction is logged, researchers studying physical communities—like villages or schools—must rely on surveys. But there’s a catch: people have limited memories and patience. To avoid "respondent fatigue," researchers often use fixed-choice surveys, asking participants to name only a small, fixed number of their acquaintances (e.g., "list up to 5 friends").

While practical, this truncation creates a "crude" observed network that is significantly sparser and less clustered than the real thing. If you use this biased data to model how a virus spreads or how information diffuses, your predictions will be wrong. In their 2018 paper, Naghmeh Momeni and Michael G. Rabbat present a way to "de-bias" this data and infer the true structural characteristics of the underlying network.

TL;DR

This paper introduces a statistical framework to accurately estimate the size and structure of social networks from truncated survey data. By distinguishing between reliably reported "strong ties" and limited "weak ties," the authors use the method of moments to recover true network properties like average degree and clustering coefficients that are otherwise hidden by survey limitations.

What Problem Does This Paper Address?

The core challenge is degree truncation. In a typical survey, an "ego" (the respondent) lists their "alters" (neighbors). While people generally remember their close friends and family (strong ties) accurately, they often fail to list all their acquaintances (weak ties). Fixed-choice designs formalize this by capping the number of weak ties a respondent can name.

This leads to a massive underestimation of the network's connectivity. Specifically, it biases the first and second moments of the degree distribution. If we don't account for this, we underestimate the "density" of the community, which is critical for understanding social cohesion or public health risks.

Methods

The authors propose a two-layer network model (one for strong ties, one for weak ties) and employ the method of moments for inference. This involves:

  1. Generative Modeling: They define a model where "seeds" (survey respondents) are chosen via a Bernoulli process. Seeds report all strong ties but only a fixed number ($B$) of weak ties, chosen uniformly at random from their total pool of acquaintances.
  2. First Moment Estimation: They derive equations to estimate the total population size ($N$), the sampling probability ($q$), and the average strong and weak degrees ($K_s, K_w$) based on the number of observed links between seeds and non-seeds.
  3. Second Moment & Clustering Estimation: This is the most complex part. To estimate clustering, they analyze how triangles and "triads" (three-node motifs) in the real network are likely to appear in the truncated survey data. They identify 42 distinct ways a triangle can be partially observed and 31 ways an open triad can appear, using these to back-calculate the true counts of these structures.
  4. Variance Estimation: They adapt the Jackknife procedure—a resampling technique—to provide an estimate of how much their estimators might vary, ensuring the results are statistically robust.

Datasets

The researchers validated their approach using both synthetic and real-world data:

  • Synthetic Networks: They generated networks using several models, including Watts-Strogatz (small-world), Barabási-Albert (scale-free), Random Recursive Trees, and the Holme-Kim model (high-clustering scale-free).
  • Real-world Data: They used the "Villages" dataset, which contains social network surveys from 77 villages in rural India. They manually partitioned the ties into "strong" (borrowing money) and "weak" (accompanying to temple) to test their estimators against a known ground truth.

Key Insights and Ideas

  • Tie Heterogeneity is Key: Previous methods treated all social ties the same. This paper argues that because our brains process strong and weak ties differently (and surveys treat them differently), the math must reflect that.
  • The "Crude" Network is a Lie: The paper demonstrates that calculating the clustering coefficient directly from raw survey data results in a massive underestimation. Their method "stretches" the observed data back toward the true values.
  • Robustness to Skew: While the math uses average degrees as an approximation, simulations showed the method remains accurate even for networks with moderately skewed degree distributions (like the Holme-Kim model), which are typical of real-world offline communities.

Key Results

  • Accuracy: In Monte Carlo simulations, the proposed estimators for $N$, $q$, and average degrees were remarkably close to the true values, with bias decreasing as the sample size increased.
  • Superiority over SFC: Their method outperformed the "Single Fixed Choice" (SFC) method, which does not differentiate between tie types, particularly in estimating the average degree and clustering.
  • Computational Efficiency: The entire inference procedure for a village-sized network (approx. 10,000 nodes) takes about one second on a standard laptop, making it highly practical for field researchers.

Limitations and Future Work

  • Uniform Weak Tie Selection: The model assumes respondents pick which weak ties to report uniformly at random. In reality, people might be more likely to name people they saw recently or those with higher social status.
  • Strong Tie Perfection: The authors assume strong ties are reported with 100% accuracy. Future work could incorporate a "forgetting" factor for strong ties.
  • Centrality Measures: Currently, the method focuses on global statistics (averages, clustering). The authors suggest extending this to estimate individual-level metrics like betweenness centrality, which identifies "hubs" or "bridges" in a community.

Summary updated

Abstract

Knowing the structure of an offline social network facilitates a variety of analyses, including studying the rate at which infectious diseases may spread and identifying a subset of actors to immunize in order to reduce, as much as possible, the rate of spread. Offline social network topologies are typically estimated by surveying actors and asking them to list their neighbors. While identifying close friends and family (i.e., strong ties) can typically be done reliably, listing all of one's acquaintances (i.e., weak ties) is subject to error due to respondent fatigue. This issue is commonly circumvented through the use of so-called “fixed choice” surveys where respondents are asked to name a fixed, small number of their weak ties (e.g., two or ten). Of course, the resulting crude observed network will omit many ties, and using this crude network to infer properties of the network, such as its degree distribution or clustering coefficient, will lead to biased estimates. This paper develops estimators, based on the method of moments, for a number of network characteristics including those related to the first and second moments of the degree distribution as well as the network size, using fixed-choice survey data. Experiments with simulated data illustrate that the proposed estimators perform well across a variety of network topologies and measurement scenarios, and the resulting estimates are significantly more accurate than those obtained directly using the crude observed network, which are commonly used in the literature. We also describe a variation of the Jackknife procedure that can be used to obtain an estimate of the estimator variance.
BibTeX
@article{momeni2017inferring,
  author = {Naghmeh Momeni and Michael G. Rabbat},
  title = {{Inferring Structural Characteristics of Networks With Strong and Weak Ties From Fixed-Choice Surveys}},
  journal = {IEEE Trans. Signal Inf. Process. Over Networks},
  year = {2017},
  url = {https://openreview.net/forum?id=-3z4tI7fhm6},
}

Loading related content...

Figure