Studies that link social networks to mental health share a common design. They record who texts and calls whom, compute where each person sits in that communication network, and predict questionnaire scores for stress, happiness, or anxiety. These predictions work surprisingly well. The catch is that the questionnaire is not a thermometer. People shade their answers toward what looks socially acceptable. The shading plausibly depends on who is watching, and in a social network, "who is watching" is a property of network structure. If people in tightly interlinked friend groups answer the same items differently from people with scattered, non-overlapping contacts, even when their underlying states match, then part of what "network predicts wellness" models learn is a reporting style, not wellbeing.
This note proposes to test that directly. We measure each person's local path geometry: the shape of the network around them as seen through shortest paths, reachability, and loops, beyond simply how many contacts they have. We then ask whether the questionnaire measures the same thing across geometrically different positions. The psychometric tool for this question is a measurement invariance test, which checks whether a scale's items relate to the underlying trait the same way in different groups. The result is either a correction that shrinks published network–wellness effects by a measurable amount, or evidence that those effects survive it.
Core Claim
Two people can have the same number of contacts but sit in very different surroundings. One has a compact position: their contacts know each other, their shortcuts close short loops, and most of the network is reachable in a few hops. The other has an extended position: a tree-like fan of contacts that do not overlap. Our hypothesis is that compact positions put a person in front of a single overlapping audience, which should raise social-desirability pressure on evaluative items such as happiness and self-rated health.
The claim is specific and falsifiable. After matching on degree and standard centralities, questionnaire items should show differential item functioning between compact and extended positions. Differential item functioning means that people with the same latent level give systematically different answers to a given item. Its size should track how socially desirable independent raters judge each item to be. Correcting for it should reduce the network–wellness associations reported on the same cohort. If geometry-based groupings reveal no more item-level bias than degree-based groupings, the claim fails.
Why the Two Literatures Need Each Other
The network–health work on the NetHealth cohort motivates the gap. Social network structure is predictive of health and wellness builds an ensemble classifier (NetCARE) on degree, triangles, clustering, betweenness, and closeness from phone and SMS logs. It reports 38–65% relative F1 gains over Fitbit-only models for self-reported stress, happiness, positive attitude, and health, but treats those survey answers as ground truth. Network analysis of the NetHealth data tracks weekly centrality changes and finds that students whose network position co-evolves with physical activity are "more likely to be introverted and anxious." Both traits are self-reported, so the finding is exactly the kind that differential reporting could produce. On StudentLife, Monitoring Behavioral Changes Using Spatiotemporal Graphs goes further in the wrong direction for our purposes: its graph edges are built from feature similarity that includes self-reports, so any reporting bias gets written into the graph itself. We differ from all three by making the self-report the object of study rather than the target.
The self-report side supplies the method and the stakes. Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders shows the multi-group invariance analysis we reuse, applied there to gender. We apply the same test with network position as the grouping variable. Tag2Risk responds to questionnaire bias by abandoning the questionnaire for passive music-tag traces, which reached 66.4% at-risk accuracy. However, it still validates those traces against the K10 distress scale and treats users as isolated points, so it does not remove the bias so much as inherit it. The stakes are visible in Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns. That micro-randomized trial finds that notifications lower mood when previous-week self-reported mood is high (est −0.052) and help only below a mood of 7. If mood reports carry position-dependent bias, such thresholds misfire for whole subgroups.
The geometry side supplies the measurements. Globally defined signatures can each be computed from a single node's point of view. The first is the shape of the shortest-path distance distribution fitted with a generalized Gamma (Parameterizing the Distance Distribution of Undirected Networks). The second is the reachability exponent $h$ in $C(r)\propto r^h$ (Power-Hop). The third is local diffusion dimension from a heat pulse spreading on the graph (Relative, local and global dimension in complex networks). The fourth is the length of a node's shortcuts, measured along the spanning tree of highest-betweenness edges (Scale-free trees). The compact-versus-extended distinction comes from Classification of scale-free networks, which shows that degree alone cannot separate these architectures.
Research Program
We first compute a per-person geometry profile on the NetHealth communication network for each semester. The profile has four parts:
- the fitted distance-distribution parameters $(a,d,p)$ from that person's shortest-path distances;
- their local reachability exponent $h$;
- their local diffusion dimension at two time scales;
- the median skeleton length of the shortcuts they participate in.
We cluster these profiles within degree bins, so that the resulting "compact" and "extended" labels cannot be reproduced by degree.
NetHealth provides both a whole network, including outside contacts, and a participant-only network. The participant network is a truncated sample, and path metrics are known to be distorted by exactly this kind of truncation. We therefore check the stability of each feature across the two versions and drop any feature whose rank correlation between them is poor. If the Mexico call-record graph (Harnessing Mobile Phone Social Network Topology to Infer Users Demographic Attributes, over 70 million users) is accessible, we also sample cohort-sized subgraphs from it and measure how far local features drift from their full-graph values. If it is not, we run the same drift study on synthetic graphs with matched degree sequences.
Second, we fit item-level measurement models to the multi-item NetHealth scales (Big Five, depression, anxiety, and the wellness items). For person $i$ and item $j$ we model the response as
$$y_{ij} = \tau_j + \lambda_j \eta_i + \delta_j g_i + \varepsilon_{ij},$$
where $\eta_i$ is the latent trait, $\tau_j$ and $\lambda_j$ are the item's baseline and loading, $g_i$ is the geometry class, and $\delta_j$ is the item's differential functioning. Nonzero $\delta_j$ means that equally anxious people in compact versus extended positions answer item $j$ differently. We also allow loadings to vary by class.
Third, we collect desirability ratings for every item from independent raters who are not part of the cohort. The mechanism predicts that $|\delta_j|$ correlates with rated desirability. Plain group differences in the latent trait predict no such item-level pattern.
Fourth, we refit NetCARE-style predictors on both raw sum scores and invariance-corrected latent scores. We then report how much of each published association survives.
Baselines
The naive reference point is the status quo. We use raw sum scores as targets and assume invariance, which amounts to ignoring bias entirely. Every correction is measured as a change relative to this point.
The strongest existing predictive method is NetCARE from Social network structure is predictive of health and wellness. It is an ensemble of support vector machines, nearest-neighbor classifiers, and random forests trained on local centralities plus Fitbit features. It reports relative F1 improvements over sensor-only models of 38% (stress), 65% (happiness), 55% (positive attitude), and 54% (self-assessed health). No absolute F1 values are reported on Lacuna, so we reproduce it ourselves on the same 325-participant Fall 2016 sample, with the same edge filter (at least three interactions in five months) and the same tuning budget used for our variants. It is the right competitor because it is the published claim our correction would shrink.
The critical ablation isolates geometry. We rerun the entire invariance analysis with groups defined by degree, clustering, and betweenness only, at identical group sizes. We also run a null in which geometry labels come from degree-preserving rewired networks. The proposal beats these baselines under three joint conditions: geometry grouping yields significantly more non-invariant items than local-centrality grouping; the rewired null yields none; and the correction moves NetCARE's target-level associations by more than their bootstrap uncertainty.
A second competitor is psychometric and uses no network. We compute standard response-style indices (acquiescence, extreme responding) from each person's full item set and include them as covariates. If these indices absorb all the geometry-linked differential functioning, the network adds nothing to bias detection.
Why not just…?
Why not just add the geometry features to NetCARE and report higher F1? Doing so would predict the biased targets more accurately and would say nothing about whether the targets are biased. That design is the circularity we identified in Monitoring Behavioral Changes Using Spatiotemporal Graphs. We will report this number for completeness, but a higher F1 on raw scores would be consistent with our hypothesis rather than evidence against it.
Why not just control for response-style indices? This is the real threat. Acquiescence and extremity indices capture bias that is uniform across items. Our mechanism predicts bias that is item-specific and scales with desirability, which a single per-person offset cannot represent. The response-style baseline settles the question. If adding those indices removes the geometry-linked $\delta_j$ and erases the desirability correlation, we will conclude that the network carries no bias information beyond generic response style, and we will report that.
Why not just replace questionnaires with passive traces, as Tag2Risk does? Passive traces such as music tags (Tag2Risk) or mobility still need a validation target. For internal states like anxiety, that target is a questionnaire, so the bias moves one step back rather than disappearing. Correcting the questionnaire also improves every passive proxy validated against it.
Why not just group by degree or clustering? Degree cannot separate compact from extended architectures (Classification of scale-free networks), and clustering captures only the loops through a node's immediate contacts. The degree-matched ablation tests directly whether the richer geometry matters. If it does not, the simpler grouping wins.
Decisive Checks
The first check is the within-person test. Over half of NetHealth students change network position across semesters; the co-evolution study reports significant centrality change for 74% of compliant users (Network analysis of the NetHealth data). We identify students whose geometry class flips between semesters. The desirability hypothesis predicts that their answers to high-desirability items shift in the direction of $\delta_j$, while their Fitbit activity and sleep, and their answers to low-desirability items, do not. Because this compares each person with themselves, stable traits cannot explain it. It is also the most discriminating test we have.
The second check is the desirability correlation across items. It should be positive and absent in the rewired null.
The third check is the attenuation audit. We report NetCARE-style effects and the co-evolution study's "introverted and anxious" finding on both raw and corrected scores. A drop in these effects is the practical payoff. A null result, meaning invariance holds across geometry classes, would also be informative, because it would license the existing literature's use of raw scores.
With about 160 people per class, invariance tests have adequate power only for moderate item-level effects on short scales. We will therefore pre-register the item set and the effect size we can detect, and treat smaller effects as unresolved rather than absent.
References
- Social network structure is predictive of health and wellness
- Network analysis of the NetHealth data: exploring co-evolution of individuals' social network positions and physical activities
- Monitoring Behavioral Changes Using Spatiotemporal Graphs: A Case Study on the StudentLife Dataset
- Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders
- Tag2Risk: Harnessing Social Music Tags for Characterizing Depression Risk
- Assessing Real-Time Moderation for Developing Adaptive Mobile Health Interventions for Medical Interns: Micro-Randomized Trial
- Parameterizing the Distance Distribution of Undirected Networks
- Power-Hop: A Pervasive Observation for Real Complex Networks
- Relative, local and global dimension in complex networks
- Scale-free trees: The skeletons of complex networks
- Classification of scale-free networks
- Harnessing Mobile Phone Social Network Topology to Infer Users Demographic Attributes