Aggregate Relational Data (ARD)

• Aggregate Relational Data (ARD) is a method that captures network information by asking respondents for counts of acquaintances with specific traits.
• ARD techniques enable the estimation of hidden-population sizes, personal network degrees, and latent structures through both Bayesian and penalized regression approaches.
• Modern ARD methods incorporate diagnostic tools and efficient computational frameworks, though they face challenges in accurately recovering individual network edges.

Aggregated Relational Data (ARD) are summary counts obtained when sampled respondents are asked questions of the form “How many $X$’s do you know?”, rather than being asked to enumerate every dyadic tie in a network. In the standard social-network formulation, the observed data are an integer matrix $Y=[y_{ik}]$, where $y_{ik}=\sum_{j\in G_k}A_{ij}$ counts the number of alters of respondent $i$ belonging to subpopulation $G_k$; equivalent notation writes $y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$, with $g^*_{ij}$ the unobserved link indicator and $w_{k,j}$ the trait indicator. ARD are used to estimate hidden-population sizes, personal network sizes, degree distributions, mixing patterns, and latent network structure, and they replace the $O(n^2)$ dyadic questioning required by a classic network census with a small set of aggregate questions (Ward et al., 26 Jun 2025, Laga et al., 23 Jan 2026, Alidaee et al., 2020).

1. Survey representation and measurement logic

In the canonical survey setting, ARD arise when respondents are asked “How many $X$’s do you know?”, producing counts $Y=[y_{ik}]$0 for respondent $Y=[y_{ik}]$1 and group $Y=[y_{ik}]$2. Typical examples include questions about acquaintances with specified demographic or behavioral traits, such as having been arrested, owning a farm, or having graduated high school. The most well-known application is the Network Scale-Up Method (NSUM), which uses ARD to estimate respondent degrees and group prevalences simultaneously (Laga et al., 23 Jan 2026).

Formally, if $Y=[y_{ik}]$3 respondents are surveyed over $Y=[y_{ik}]$4 groups, then $Y=[y_{ik}]$5. When the full adjacency matrix $Y=[y_{ik}]$6 is unobserved, ARD provide indirect access to network structure through counts aggregated over known groups. In one notation, $Y=[y_{ik}]$7. In another, with $Y=[y_{ik}]$8 total population members and $Y=[y_{ik}]$9 interviewed agents, $y_{ik}=\sum_{j\in G_k}A_{ij}$0, and in matrix form $y_{ik}=\sum_{j\in G_k}A_{ij}$1, where $y_{ik}=\sum_{j\in G_k}A_{ij}$2 records trait membership and $y_{ik}=\sum_{j\in G_k}A_{ij}$3 collects the unobserved links from the population to sampled respondents (Alidaee et al., 2020).

This measurement design is motivated by feasibility. A classic network census asks every ordered pair $y_{ik}=\sum_{j\in G_k}A_{ij}$4 whether they are linked, which requires $y_{ik}=\sum_{j\in G_k}A_{ij}$5 questions. ARD replace this with a small number of aggregate questions per sampled ego. The resulting compression is central to both the utility and the difficulty of ARD: collection is cheaper and often more practical, but inference becomes an inverse problem because the data are coarse summaries of an unobserved graph (Alidaee et al., 2020, Breza et al., 2017).

2. Inferential targets and identification

The inferential target varies across ARD applications. In hidden-population estimation, the target is typically the vector of prevalences

$y_{ik}=\sum_{j\in G_k}A_{ij}$6

for unknown subpopulation sizes $y_{ik}=\sum_{j\in G_k}A_{ij}$7 in a population of size $y_{ik}=\sum_{j\in G_k}A_{ij}$8, together with degree-distribution parameters such as $y_{ik}=\sum_{j\in G_k}A_{ij}$9 and $i$0 under a prior like $i$1 (Seymour et al., 2 Jun 2026). In broader Bayesian ARD models, latent parameters often include personal network sizes $i$2, baseline tie-formation rates $i$3, and possibly overdispersion, mixing, or latent-position parameters (Ward et al., 26 Jun 2025).

A more general network-recovery formulation treats the unobserved links as generated by

$i$4

where $i$5 are i.i.d. errors with unknown cdf $i$6, $i$7 are latent node-specific effects, and $i$8 is an unknown link-utility function. The primary object of interest is then the link-probability matrix

$i$9

so that estimating $G_k$0 is equivalent to estimating the distribution of links (Alidaee et al., 2020).

Identification results in the ARD literature exploit cross-group link probabilities. One line of work shows that, for the beta-model with node-specific unobserved effects, the stochastic block model with unobserved community structure, and latent geometric space models with unobserved latent locations, cross-group link probabilities identify the model parameters, meaning that ARD are sufficient for parameter estimation under those model classes (Breza et al., 2019). Another line of work uses a latent-distance formation model with node positions $G_k$1, gregariousness parameters $G_k$2, and homophily parameter $G_k$3, under which

$G_k$4

Up to sphere-wide rotations and reflections, three trait centers must be fixed to remove rotational indeterminacy; under support conditions such as non-collinearity of centers, parameters are uniquely identified (Breza et al., 2017).

These identification results delimit what ARD can and cannot recover. They support consistent estimation of model parameters and many derived network statistics, but they do not imply exact recovery of realized edges in a single observed graph.

3. Estimation frameworks for network recovery

Bayesian parametric estimation was an early dominant approach. In the latent-distance specification, ARD counts are modeled with a Poisson approximation,

$G_k$5

and the posterior is explored by Metropolis-within-Gibbs updates over $G_k$6, $G_k$7, trait centers and concentrations, and $G_k$8. Posterior draws of the latent parameters induce link probabilities $G_k$9, from which full graphs can be simulated and arbitrary node-level or graph-level statistics estimated by Monte Carlo averaging (Breza et al., 2017).

A distinct route dispenses with strong parametric assumptions and instead exploits low effective rank. Writing the ARD moment condition in expectation as

$y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$0

network recovery can be posed as nuclear-norm penalized regression:

$y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$1

Here $y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$2 is the Frobenius norm, $y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$3 is the nuclear norm, $y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$4 is a tuning parameter, and $y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$5 can enforce nonnegativity, symmetry, and a zero diagonal. The nuclear-norm penalty is motivated by the empirical observation that many economic network models, including latent-space models, blockmodels, random-dot-product graphs, and degree-heterogeneity models, have low-effective-rank $y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$6 (Alidaee et al., 2020).

The corresponding optimization is implemented with an accelerated proximal gradient method rather than direct semidefinite programming. Each iteration applies a gradient step,

$y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$7

followed by singular-value thresholding. Theory gives a finite-sample bound in terms of the effective rank

$y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$8

under which

$y_{k,i}=\sum_{j=1}^{n_2} g^*_{ij}w_{k,j}$9

so that the average per-entry mean squared error is $g^*_{ij}$0; when $g^*_{ij}$1 has true rank $g^*_{ij}$2, this becomes $g^*_{ij}$3 (Alidaee et al., 2020).

This suggests that ARD-based recovery is especially effective when the latent link-probability matrix is low-rank or low-dimensional, and less informative when the inferential target is the exact realized adjacency matrix.

4. Bayesian modeling, diagnostics, and intractable likelihoods

The contemporary Bayesian ARD literature includes a unified Stan-based implementation of existing models. A common baseline is the Poisson model $g^*_{ij}$4, with overdispersed extensions using the NegativeBinomial family, and priors such as $g^*_{ij}$5 and $g^*_{ij}$6 for $g^*_{ij}$7. A central technical issue is non-identifiability: the transformation $g^*_{ij}$8, $g^*_{ij}$9 leaves $w_{k,j}$0 unchanged. Within-iteration rescaling addresses this by anchoring to reference subpopulations with known prevalence, eliminating the ad hoc post-processing scaling step and improving convergence diagnostics; in synthetic experiments, $w_{k,j}$1 dropped from approximately $w_{k,j}$2–$w_{k,j}$3 in the unscaled models to approximately $w_{k,j}$4 in the scaled models (Ward et al., 26 Jun 2025).

Model criticism has become a separate focus. A point-estimate-based diagnostic workflow proceeds in three stages: covariate-structure diagnostics, correlation diagnostics, and distributional diagnostics. Baseline Poisson or negative-binomial models are fit by MLE or MAP; randomized-quantile residuals are then used in local-covariate plots and global-covariate plots to decide whether covariates should enter with group-specific or common coefficients. Residual dependence across groups is assessed from the largest eigenvalue of the sample covariance matrix of residuals, centered and scaled to form a Tracy–Widom test statistic. Distributional adequacy is checked with rootograms and per-group dispersion indices $w_{k,j}$5. The prescribed ordering matters: misordering the diagnostics can produce spurious detection of correlation or overdispersion when omitted covariates create extra residual noise (Laga et al., 23 Jan 2026).

A further development addresses ARD models whose likelihood cannot be written down or efficiently evaluated. In that setting, a simulation-based neural estimation framework trains a permutation-invariant neural Bayes estimator that outputs a posterior median and a $w_{k,j}$6 credible interval for each marginal parameter by minimizing a multi-quantile pinball loss with a cumulative-gap construction that rules out quantile crossing by design. The framework was demonstrated on three structurally distinct intractable extensions of NSUM-style inference: a stochastic block model, a latent-space model, and a recall-subset model. On held-out simulated data, credible-interval empirical coverage was within $w_{k,j}$7 of nominal on all three simulators; training required approximately $w_{k,j}$8 minutes on a single CPU core per model, while inference on a single real survey took less than $w_{k,j}$9 ms (Seymour et al., 2 Jun 2026).

These developments reflect a shift in ARD methodology from single-model estimation toward a broader workflow of identification, computation, diagnostic checking, and sensitivity analysis.

5. Empirical performance and substantive applications

Simulation and field evidence indicate that ARD can recover many network summaries reliably, though not all of them. In the latent-distance Bayesian framework, simulations with $O(n^2)$0, $O(n^2)$1, and $O(n^2)$2 trait centers showed that degree and eigenvector centrality typically have mean squared error tending to zero as $O(n^2)$3 grows, and graph-level statistics such as maximal eigenvalue, proximity, and clustering also had low MSE. At the same time, single-edge MSE remained large, and betweenness and component count had larger errors (Breza et al., 2017).

The same paper applied the method to $O(n^2)$4 village networks in Karnataka, using a $O(n^2)$5 random subsample with ARD on $O(n^2)$6 traits. Posterior-mean versus true village-level statistics yielded $O(n^2)$7 for max-eigenvalue, $O(n^2)$8 for proximity, $O(n^2)$9 for global clustering, and $X$0 for eigenvector-cut. For node-level quantities, degree achieved $X$1 on ARD nodes and $X$2 on all nodes via kNN imputations, while eigenvector centrality achieved $X$3 on ARD nodes. In a top-decile centrality classification exercise, the true-positive rate was approximately $X$4 for ARD nodes versus $X$5 for all nodes. The same study reported that, in rural India, ARD surveys were approximately $X$6 cheaper than full network surveys (Breza et al., 2017).

The penalized-regression approach exhibits a different performance profile. Under latent-space, random-dot-product, and stochastic-block simulation designs with $X$7 and $X$8, the reported average MSE $X$9 fell from approximately $Y=[y_{ik}]$00–$Y=[y_{ik}]$01 at $Y=[y_{ik}]$02 to approximately $Y=[y_{ik}]$03–$Y=[y_{ik}]$04 at $Y=[y_{ik}]$05. Computation time was less than $Y=[y_{ik}]$06 seconds for $Y=[y_{ik}]$07 on a standard laptop, and comparison to parametric latent-space MCMC showed orders-of-magnitude speedup, with similar or better accuracy when the rank assumption held (Alidaee et al., 2020).

Diagnostic work has also been evaluated on large real datasets. In Ukraine 2008–2009 ARD with $Y=[y_{ik}]$08 respondents, $Y=[y_{ik}]$09 groups, eight respondent-level covariates, and one respondent–group “respect” covariate, the diagnostic workflow identified local effects for Gender, Age, Employment, Internet Access, and Secondary Education, and global effects for Ukrainian ethnicity, Vocational Education, and Academic Education. After fitting a negative-binomial regression with the selected covariates, the Tracy–Widom test yielded $Y=[y_{ik}]$10 with $Y=[y_{ik}]$11, strongly rejecting independence and indicating residual group correlation (Laga et al., 23 Jan 2026).

Likelihood-free ARD models have been applied to the 2011 Rwanda household survey, which interviewed approximately $Y=[y_{ik}]$12 respondents and asked ARD questions for $Y=[y_{ik}]$13 known populations and $Y=[y_{ik}]$14 hidden populations. Across stochastic-block, latent-space, and recall-subset NSUM models, posterior medians for hidden-population sizes differed by factors of $Y=[y_{ik}]$15–$Y=[y_{ik}]$16. The paper emphasizes that no single model is “truth,” and treats these discrepancies as quantification of model uncertainty beyond sampling error (Seymour et al., 2 Jun 2026).

ARD has also been used outside the standard hidden-population setting. In a three-wave King County study of people experiencing homelessness, RDS-II population estimates of mean acquaintance degree were $Y=[y_{ik}]$17 in 2023 and $Y=[y_{ik}]$18 in 2024, while close-friend estimates were $Y=[y_{ik}]$19 in 2023 and $Y=[y_{ik}]$20 in 2024, and kinship estimates were $Y=[y_{ik}]$21 in 2023 and $Y=[y_{ik}]$22 in 2024. The reported interpretation was declining visibility, stable but low strong-tie support, and modestly growing family clustering within the unsheltered population (Almquist et al., 2024).

6. Limits, misconceptions, and terminological scope

A recurrent misconception is that ARD recover the realized graph itself. The literature is more circumspect. Parametric and nonparametric methods can recover model parameters, link-probability matrices, and many node-level or graph-level statistics, but they do not generally recover individual edges accurately. The latent-distance study explicitly reports that single-edge MSE stays large, and that high-order quantities such as betweenness and component count are harder to estimate (Breza et al., 2017). This suggests that ARD are best understood as instruments for inference on distributions of ties and network summaries, not as substitutes for a complete edge list.

Another central issue is modeling assumptions. Much Bayesian ARD inference assumes that, conditional on degree, counts for different subpopulations are independent. Recent work argues that homophily, latent-space clustering, imperfect recall, non-random mixing, visibility biases, unobserved heterogeneity, and residual dependence can violate that assumption. The practical response in the recent literature is not a single replacement model but a layered workflow: covariate diagnostics, correlation testing, distributional checks, posterior predictive checks, and cross-validation where feasible (Laga et al., 23 Jan 2026, Ward et al., 26 Jun 2025, Seymour et al., 2 Jun 2026).

The acronym “ARD” also has a separate meaning in database-oriented machine learning over relational data. In that literature, an ARD query is a SQL-style aggregate over a feature-extraction join, and a batch of ARD queries is computed together to form the sufficient statistics for objectives such as ridge regression, PCA, support vector machines, decision trees, and k-means. Optimized ARD processing based on factorization, FAQ/InsideOut, aggregate fusion, tries, and incremental maintenance can avoid materializing large joins; on a retailer dataset, a structure-aware system was reported to reduce runtime from $Y=[y_{ik}]$23 seconds in a Postgres+TensorFlow pipeline to $Y=[y_{ik}]$24 seconds, a $Y=[y_{ik}]$25 speedup, while moving $Y=[y_{ik}]$26 KB instead of $Y=[y_{ik}]$27 GB (Olteanu, 2020).

The coexistence of these usages makes terminological precision important. In the social-network and survey-statistics literature, ARD denote aggregated counts over personal networks; in the database literature, ARD denote aggregate queries over relational joins. The two traditions share an emphasis on aggregation as a surrogate for more expensive full-data access, but they address different inferential objects and computational problems.