Causal inference for social network data

Elizabeth L. Ogburn; Oleg Sofrygin; Ivan Diaz; Mark J. van der Laan

Source-linked AI summary

Causal inference for social network data

Elizabeth L. Ogburn, Oleg Sofrygin, Ivan Diaz, Mark J. van der Laan

arXiv:1705.08527v6May 23, 2017stat.MEmath.ST

TL;DR

Causal inference for social-network data lacks methods that appropriately handle dependence among connected individuals. The paper develops semiparametric methods for a single network and finds no evidence of causal obesity peer effects after accounting for network structure.

Problem
Existing causal-inference methodology has not kept pace with social-network data, leading researchers toward inappropriate statistical approaches.
Method
The paper proposes semiparametric causal and statistical inference methods for a single interconnected network, allowing dependence informed by network ties and growing numbers of ties.
Results
The reanalysis of obesity peer effects in the Framingham Heart Study finds no evidence for causal peer effects after accounting for network structure.
Takeaways & Limitations
The Framingham analysis illustrates the dangers of estimating peer effects without adequately accounting for dependence and network structure.
Takeaways & Limitations
Dependence among network observations can produce spurious associations, constraining causal interpretation of social-network data.

Abstract

from arXiv · show

We describe semiparametric estimation and inference for causal effects using observational data from a single social network. Our asymptotic results are the first to allow for dependence of each observation on a growing number of other units as sample size increases. In addition, while previous methods have implicitly permitted only one of two possible sources of dependence among social network observations, we allow for both dependence due to transmission of information across network ties and for dependence due to latent similarities among nodes sharing ties. We propose new causal effects that are specifically of interest in social network settings, such as interventions on network ties and network structure. We use our methods to reanalyze an influential and controversial study that estimated causal peer effects of obesity using social network data from the Framingham Heart Study; after accounting for network structure we find no evidence for causal peer effects.

1. INTRODUCTION

The paper develops causal inference methods for observational data from a single social network, addressing dependence across connected individuals and limitations of existing approaches. It allows both transmission-based and latent dependence, introduces network-specific causal estimands, and establishes conditions for inference on network interventions.

1. INTRODUCTION: Existing social-network analyses often use methods that do not handle dependence across individuals, motivating specialized causal-inference procedures.The paper notes criticism of standard GLM and GEE-based approaches for causal peer effects.
1. INTRODUCTION: Prior interference methods generally require multiple independent networks or randomized treatment, whereas this paper targets observational data from a single network.The authors seek inference when exposures cannot be randomized and all observations come from one social network.
1. INTRODUCTION: The paper proves a central limit theorem showing that estimators remain consistent and asymptotically normal under more realistic dependence structures and growing dependence regimes.The framework extends earlier work that assumed each unit depended on only a small, effectively fixed number of others and that dependence arose solely through direct transmission.
1. INTRODUCTION: The methods accommodate both observable transmission along network ties and latent, unstructured dependence among socially connected nodes.The authors identify latent dependence as a common feature of social networks and a limitation of methods designed only for direct transmission.
1. INTRODUCTION: The paper introduces causal estimands for network settings, including interventions on network ties and network structure, and provides conditions for causal inference about such interventions.Together, these contributions form a framework for observational causal inference using a single social network.
1. INTRODUCTION: The paper focuses primarily on theory and briefly discusses estimation, while implementation and computation are addressed in companion papers and an R package.The authors distinguish the theoretical focus of this paper from implementation details covered elsewhere.

2. BACKGROUND AND SETTING

The paper frames causal inference in a single social network, where ties may transmit effects and also reflect dependence among connected units. It defines network-wide causal estimands and interventions while conditioning on the observed network structure.

Motivating example: The motivating setting is observational data from one interconnected social network, illustrated through the Framingham Heart Study and its reconstructed social ties.The FHS example concerns peer effects involving obesity, smoking, and happiness.
Causal questions: Causal questions include interference between tied subjects and peer effects in which one subject’s outcome affects alters’ future outcomes.These questions require distinguishing effects transmitted across network ties from other relationships among units.
Network structure: Accurate network measurement is essential because missing ties can create unmeasured confounding, omit exposure pathways, and cause exposure misclassification.The framework therefore assumes that the complete network is observed.
Causal estimands: The framework averages potential outcomes across network nodes while conditioning estimands and estimators on the observed adjacency matrix and network size.Potential outcomes may depend on the full exposure vector because other nodes’ exposures can affect an individual’s outcome.
Interventions and inference: The paper considers static, dynamic, and stochastic interventions, including interventions on network structure, and establishes identification despite network dependence.Its inferential framework targets a single network and proves asymptotic normality under general dependence conditions.

3. METHODS

The paper develops a semiparametric framework for causal effects in social networks, using structural equation models that accommodate network-mediated dependence and latent similarities while supporting identification, estimation, and inference under stated assumptions.

The framework generalizes causal-effect estimation to broader dependence among network observations and more realistic asymptotic regimes.
3.1 Structural equation model: The structural equation model represents covariates, exposures, and outcomes as functions of neighboring variables and exogenous errors, with unknown functions that may depend on node degree.
3.2 Definition and nonparametric identification of causal effects: Causal effects are identified under conditional ignorability given covariates, while the paper develops semiparametric estimation and inference with doubly robust TMLE properties.
3.1 Structural equation model: The assumptions permit dependence from network ties and limited latent-variable dependence, but exclude latent homophily that jointly confounds exposure and outcome.
3.1 Structural equation model: Summary functions W and V reduce neighboring covariates and exposures to quantities used in the exposure and outcome models, while remaining deterministic functions of observed network data.
3.4 Asymptotic normality: The asymptotic theory establishes normal limiting behavior and variance estimation when second-order terms are controlled under the stated assumptions.

4. EXTENSIONS

The paper extends causal inference to social-network settings by defining peer effects and interventions on network structure, including dynamic and stochastic interventions. These effects require assumptions about transmission, positivity, and how network structure affects outcomes.

4.1 Peer effects: The framework defines causal effects for social contagion or peer effects and interventions that alter network structure itself.Network-structure interventions can add, remove, or relocate ties.
4.1 Dynamic and stochastic interventions: Dynamic interventions assign exposures as deterministic functions of covariates, whereas stochastic interventions replace the exposure mechanism with a user-specified random distribution.The same asymptotic results apply to static, dynamic, and stochastic interventions.
4.2 Peer effects: Peer-effect identification requires the time between measurements to permit transmission only between nodes and their immediate alters.Broader contagion creates dependence and possible confounding beyond what the methods account for.
4.3 Interventions on network structure: Network interventions may replace the adjacency matrix with a fixed matrix or a random draw from matrices sharing specified features, such as a bounded maximum degree.Local features such as degree or clustering may be identifiable when global structure is not.
4.3 Interventions on network structure: Network-structure effects rely on a strong assumption that network structure affects outcomes through the modeled exposure mechanism; positivity also requires counterfactual exposure support in the observed data.Interventions on global features may be unidentified even when local-feature interventions are identifiable.

5. SIMULATIONS

Simulations evaluate the proposed TMLE and variance estimators under direct-transmission and latent-variable dependence in preferential-attachment networks. Ignoring dependence produces anticonservative inference, whereas dependent and bootstrap procedures improve coverage, with bootstrap inference especially robust in smaller samples.

Intervention effects: Targeting highly connected and physically active individuals can yield larger population-level intervention effects than assigning the same exposure to an untargeted group.The simulations also demonstrate feasibility of interventions on network structure and combinations of exposure and network interventions.
Direct transmission: Ignoring network dependence produced confidence-interval coverage as low as 50%, even for large samples, under direct transmission.The naive i.i.d. variance estimator was anticonservative.
Direct transmission: Dependent influence-curve variance estimates approached nominal 95% coverage for sufficiently large samples, while parametric bootstrap intervals were most robust in smaller samples.Bootstrap intervals also attained nominal 95% coverage in large samples across nearly all scenarios.
Simulation limitations: Small-sample performance can deteriorate because of limited asymptotic normality and near-positivity violations.This limitation is reported for dependent variance estimates in smaller samples.
Latent-variable dependence: Under latent-variable dependence, variance estimates assuming conditionally i.i.d. outcomes were anticonservative.That assumption is valid for direct transmission alone but fails when latent-variable dependence is present.

6. DATA ANALYSIS

The Framingham Heart Study reanalysis applies network-aware methods to obesity peer effects while accounting for dependence across the full social network. Unlike naive analyses, it finds that the reported strong contagion effects may be spurious rather than true associations or causal effects.

Methods: The reanalysis used all ten social-connection types simultaneously for 3,766 participants and modeled each ego as a potentially dependent observation.This contrasts with pairwise analyses that treat network ties as independent observations.
Caveats: The authors caution that the Framingham estimates are not true causal effects because of unobserved confounding and exposure measured concurrently with the outcome.The comparison remains informative for contrasting network-aware and naive methods.
Findings: Accounting for interdependence in the Framingham data undermines findings of strong contagion effects for obesity.The analysis is consistent with the significant earlier results being driven by dependence or model misspecification.

7. CONCLUSION

The paper develops methods for causal and statistical inference from a single interconnected social network, allowing dependence from network ties and growing numbers of connections. Its Framingham reanalysis illustrates that naive independent-unit methods can produce misleading peer-effect findings.

Contributions: The proposed methods support causal and statistical inference from a single social network while incorporating dependence informed by network ties.They address both causal and statistical dependence among interconnected subjects.
Contributions: Unlike existing methods, the framework does not require randomized exogenous treatment and remains valid when network ties grow slowly with sample size.The asymptotic results accommodate dependence on a growing number of other units.
Application: The Framingham obesity analysis demonstrates the danger of applying naive independent-unit methods to peer effects.Accounting for network dependence undermined the study’s reported strong contagion effects.

8. REGULARITY CONDITIONS

The theorem’s regularity conditions constrain nuisance-function classes, empirical-process complexity, boundedness, and the TMLE update so asymptotic normality can be established.

The assumptions require nuisance estimators and their TMLE updates to remain within specified function classes for m and h̃.The procedure models h̃ rather than g directly.
The regularity conditions are stated as requirements for Theorem 1’s asymptotic normality result.
Uniform consistency is imposed for asymptotic equicontinuity, while bounded entropy and a universal envelope control the empirical-process class.Uniform consistency is not needed for convergence-rate proofs.
The universal bound is typically obtained by choosing a function class that satisfies the stated entropy condition.

Consistency and rates for estimators of nuisance parameters: Assume that ∥ˆm −m∥

The nuisance estimators must satisfy convergence-rate and second-order remainder conditions, with rates that can be weakened under more general technical assumptions.

The second-order term must be asymptotically negligible at the theorem’s normalization rate.
The parametric TMLE update has negligible effect on the initial estimator’s convergence rate, so the updated estimator converges at nearly the same rate.
The network assumptions impose limited connectivity and dependence through a bound on maximum degree.
Both nuisance models can satisfy the required condition when they converge to truth at rate C_n^-1/4.The text notes that more general conditions are available.
The argument allows a pre-specified parametric model, while the general asymptotic-normality conditions follow prior TMLE theory apart from convergence rates.

9. OVERVIEW OF THE PROOF OF THEOREM 1

The proof decomposes the TMLE error into first- and second-order terms, controls the remainder using empirical-process arguments, and establishes normality through a dependency-neighborhood CLT.

Network structure changes the proof by requiring bounds on Orlicz norms for empirical processes associated with influence-function components.
The proof requires second-order terms to be smaller than 1/√C_n, while the first-order sum supplies the asymptotic distribution.
The second-order control extends empirical-process theory while following the format of earlier TMLE proofs.
The key combinatorial step bounds overlapping friend groups, extending the argument from fixed to growing numbers of network connections.
The first-order terms converge to a normal distribution using a CLT for dependent data with growing, irregular dependency neighborhoods.The paper proves this CLT in Lemmas 1 and 2.

10. CENTRAL LIMIT THEOREM FOR FIRST ORDER TERMS

The first-order influence-function components are analyzed with dependency neighborhoods formed by friends and friends of friends, and Stein-based CLTs yield normal limits under network-growth conditions.

The proof assumes independence for nodes lacking a shared tie or mutual contact, conditional on the relevant covariates.
Each unit’s dependency neighborhood contains itself, its friends, and its friends’ friends, reflecting the network-based dependence structure.
Stein’s method bounds Wasserstein distance to a standard normal distribution, and the bound vanishes under the stated maximum-degree condition.
The argument permits unequal degrees because removing ties preserves the relevant upper bound, so convergence holds whenever K_i ≤ K_max,n.
Conditional normal convergence is transferred to marginal convergence by dominated convergence, with separate limits for the influence-function components.
Z′_nX, Z_nY, and Z_nC are shown to have asymptotically normal distributions, yielding a normal limit for their orthogonal sum.

11. VARIANCE ESTIMATION

The paper develops variance estimators for TMLE under network dependence and evaluates confidence-interval performance across dependence structures, network models, and sample sizes.

Ignoring network dependence produced anticonservative variance estimates and confidence-interval coverage as low as 50%.
Dependent-influence-curve intervals approached nominal 95% coverage at sufficiently large sample sizes but could degrade with small samples.The reported small-sample problems were linked to lack of asymptotic normality and near-positivity violations.
Parametric-bootstrap intervals had the most robust small-sample coverage and reached nominal 95% coverage in large samples across nearly all scenarios.The apparent robustness extended to sample sizes as low as n = 500.
Re-scaled TMLE distributions converged toward their theoretical normal limits in preferential-attachment and small-world networks, with similar approximate normality under latent dependence.