Simpson's Contagion Dynamics

• Simpson's contagion is defined as a process that appears superlinear at the population level despite linear or sublinear dynamics within each subgroup.
• The model employs group-based transmission kernels to demonstrate how hidden heterogeneity and shifting subgroup compositions create threshold-like aggregate behavior.
• It challenges conventional contagion inference by showing that observed nonlinearity may stem from aggregation bias rather than genuine social reinforcement.

Searching arXiv for recent and foundational papers directly relevant to Simpson's contagion, Simpson's paradox, and contagion inference in networks. Simpson’s contagion denotes a contagion process that looks superlinear when observed over an entire population, but is mechanistically linear or even sublinear in all of its subgroups. The concept was introduced to explain why threshold-like, nonlinear, and “complex contagion” signatures can appear ubiquitously in aggregate incidence data even when no subgroup exhibits intrinsically superlinear transmission (Hébert-Dufresne et al., 1 May 2026). In this formulation, the phenomenon is a contagion-specific manifestation of Simpson’s paradox: the population-level kernel is not a direct average of subgroup kernels, but a composition-weighted average over a heterogeneous mixture whose composition changes with incidence. As a result, empirical patterns often interpreted as social reinforcement, peer pressure, or nonlinear dose dependence may instead reflect aggregation bias over hidden heterogeneity (Hébert-Dufresne et al., 1 May 2026).

1. Definition and conceptual scope

The defining statement is explicit: a Simpson’s contagion is “a contagion process that looks superlinear when observed over an entire population, but is mechanistically linear or even sublinear in all of its subgroups” (Hébert-Dufresne et al., 1 May 2026). The key contrast is between mechanistic contagion dynamics within subgroups and observed population-level dynamics. Within a homogeneous subgroup with fixed parameters, the transmission kernel is straightforward and interpretable; after aggregation over groups with unobserved or hidden covariates, the effective kernel can display increasing returns or threshold behavior that no subgroup possesses.

This usage places Simpson’s contagion within the broader logic of Simpson’s paradox. In the classical setting, the paradox arises because marginal and conditioned associations can point in opposite directions. A general resolution proposed for the minimal binary case is that the relevant association should be defined after conditioning on the common cause $C$, and that in the binary-common-cause setting this selects the same direction as the usual “condition on $B$” option (Hovhannisyan et al., 2024). Simpson’s contagion imports that caution into contagion modeling: the aggregate pattern need not reveal the correct mechanism unless the relevant heterogeneity is modeled or conditioned upon.

A plausible implication is that “complex contagion” should be treated as an observational description unless subgroup structure and hidden covariates have been explicitly addressed. The 2026 analysis makes this point directly by stressing the pitfall of model selection that ignores correlations and heterogeneity in populations (Hébert-Dufresne et al., 1 May 2026).

2. Mechanistic model and coarse-grained population kernel

The formal starting point is the contagion kernel $\beta(i)$, the total rate incident on a susceptible individual as a function of the number $i$ of infectious contacts (Hébert-Dufresne et al., 1 May 2026). In a simple contagion,

$\beta(i)=\lambda i,$

so the increment

$\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$

is constant. Superlinear kernels have increasing $\Delta\beta(i)$; sublinear kernels have decreasing $\Delta\beta(i)$; threshold-like kernels are nearly flat at low $i$ and then jump upward (Hébert-Dufresne et al., 1 May 2026).

The subgroup model is group-structured. Each group has a static type $(n,\lambda,\nu)$, where $B$0 is group size, $B$1 the local transmission rate, and $B$2 the scaling exponent. If a group contains $B$3 infectious individuals, each susceptible becomes contagious at rate

$B$4

This recovers linear contagion for $B$5, superlinear contagion for $B$6, and sublinear contagion for $B$7 (Hébert-Dufresne et al., 1 May 2026).

The global dynamics are represented by group-based approximate master equations, where $B$8 is the fraction of groups of type $B$9 with $\beta(i)$0 infectious agents: $\beta(i)$1 with recovery rate set to $\beta(i)$2 by rescaling time. The cross-group coupling is

$\beta(i)$3

This is a susceptible-weighted average of transmission across groups, multiplied by the mean membership $\beta(i)$4 (Hébert-Dufresne et al., 1 May 2026).

That susceptible weighting is the core of the paradox. The observed population kernel is obtained by coarse-graining over hidden heterogeneity: $\beta(i)$5 Near equilibrium, detailed balance yields

$\beta(i)$6

and hence

$\beta(i)$7

The weighting is therefore by the equilibrium abundance of subgroup types at incidence $\beta(i)$8, so groups with larger local transmission are disproportionately represented at high incidence (Hébert-Dufresne et al., 1 May 2026).

3. Aggregate nonlinearity as a Simpson’s-paradox effect

The central mechanism is compositional change. As incidence $\beta(i)$9 increases, the sample of groups observed at that incidence is not a random sample of all groups. It is enriched for groups with stronger local transmission, which pushes the aggregate kernel upward. This changing mixture can make the pooled curve accelerate even if no subgroup’s kernel accelerates (Hébert-Dufresne et al., 1 May 2026).

The paper distinguishes two notions of increment: $i$0 which compares population-average kernels at adjacent incidence values, and

$i$1

which compares the actual increase in a fixed subgroup as it moves from $i$2 to $i$3 (Hébert-Dufresne et al., 1 May 2026). In heterogeneous populations these quantities differ. The apparent nonlinearity is captured by $i$4, whereas the within-subgroup mechanism is better reflected by $i$5.

This is the precise sense in which the phenomenon is a Simpson’s paradox. Subgroup-level trends may be linear or sublinear, but the pooled trend is nonlinear because the composition of the pooled sample changes with $i$6 (Hébert-Dufresne et al., 1 May 2026). The broader statistical parallel is reinforced by the common-cause treatment of Simpson’s paradox, where the “correct” association depends on conditioning over the variable that explains the dependence structure rather than marginalizing it away (Hovhannisyan et al., 2024).

A common misconception is to treat the population-level threshold or superlinearity as direct evidence of reinforcement inside individuals or groups. The 2026 analysis argues instead that such signatures may simply be Simpson’s contagions arising from hidden group structure, local correlations, or parameter heterogeneity (Hébert-Dufresne et al., 1 May 2026).

4. Canonical case studies

The 2026 paper develops three mathematical case studies showing how Simpson’s contagions arise under distinct forms of heterogeneity (Hébert-Dufresne et al., 1 May 2026).

Case	Subgroup mechanism	Aggregate appearance
Heterogeneous $i$7, fixed $i$8	All groups linear	Threshold-like coarse-grained kernel
Fixed $i$9, heterogeneous $\beta(i)=\lambda i,$0	All groups sublinear	Threshold-like, asymptotically linear kernel
Correlated $\beta(i)=\lambda i,$1 and $\beta(i)=\lambda i,$2	Mixed hidden structure	Threshold-like or more complicated patterns

In the first case, all groups are mechanistically linear, with $\beta(i)=\lambda i,$3 on $\beta(i)=\lambda i,$4. The effective kernel is

$\beta(i)=\lambda i,$5

with

$\beta(i)=\lambda i,$6

Despite linear subgroup dynamics, the coarse-grained kernel becomes threshold-like because low incidence is associated mostly with low-$\beta(i)=\lambda i,$7 groups, while high incidence is increasingly dominated by high-$\beta(i)=\lambda i,$8 groups (Hébert-Dufresne et al., 1 May 2026).

In the second case, $\beta(i)=\lambda i,$9 with $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$0, so every group is sublinear. The derived effective kernel is

$\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$1

and for large $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$2,

$\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$3

Thus the aggregate dynamics become threshold-like and eventually asymptotically linear even though no group is superlinear (Hébert-Dufresne et al., 1 May 2026).

In the third case, $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$4 and $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$5 are correlated through a bivariate normal copula with uniform marginals $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$6 and $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$7. Positive Pearson correlation $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$8 yields threshold-like behavior approaching roughly $\Delta\beta(i)=\beta(i)-\beta(i-1)=\lambda$9, while negative correlation can create alternating regimes of decreasing and increasing returns before settling to a lower-rate linear contagion as $\Delta\beta(i)$0 (Hébert-Dufresne et al., 1 May 2026). This shows that the paradox is not restricted to marginal heterogeneity; dependence structure between hidden parameters can reshape the effective kernel.

These case studies support a general interpretation: threshold-like or complex contagions can be effective models of aggregate data without being mechanistic models of any subgroup (Hébert-Dufresne et al., 1 May 2026).

5. Relation to contagion inference in networks

Simpson’s contagion is closely related to longstanding identification problems in network science, especially the difficulty of distinguishing genuine influence from compositional or structural artifacts. One such problem is contagion versus homophily. In a fixed network with two waves of a binary nodal outcome, contagion is generically unidentified from observational data; the observed connected-dyad association can be distorted because connected dyads are a selected subpopulation and latent traits may jointly influence tie formation and outcomes (Clark, 16 Jun 2026).

The 2026 sensitivity-analysis framework formalizes this as selection bias. It defines controlled direct effects on the risk-ratio scale while holding the tie present and shows that the gap between the observed connected-dyad risk ratio and the causal target is governed by how strongly a latent homophily variable shifts the composition of connected dyads (Clark, 16 Jun 2026). This is not identical to Simpson’s contagion, but it is structurally analogous: an observed aggregate association may weaken, disappear, or reverse once the relevant hidden structure is accounted for.

An additional network-specific caution comes from causal inference for infectious disease outcomes. In the vaccination setting, one individual’s treatment can affect another’s outcome through distinct causal pathways called contagion and infectiousness (Ogburn et al., 2014). That work decomposes the indirect effect of an alter’s vaccination on an ego’s outcome into a contagion effect, operating through susceptibility or acquisition by the alter, and an infectiousness effect, operating through transmissibility given infection (Ogburn et al., 2014). It also shows that naive independence-based generalized linear models in social network data are invalid and expected to be anticonservative because outcomes are non-independent; valid GLM-based inference requires conditioning on contact-history summaries so that residuals become uncorrelated across sampled groups (Ogburn et al., 2014).

The connection is methodological. Simpson’s contagion concerns spurious nonlinearity produced by aggregation over heterogeneous subgroups, whereas the vaccination paper concerns causal decomposition under interference and the homophily framework concerns latent selection into ties. Yet all three warn against reading mechanistic claims directly from pooled network-level associations [(Ogburn et al., 2014); (Clark, 16 Jun 2026)]. This suggests that Simpson’s contagion belongs to a wider family of aggregation and selection phenomena in contagion inference.

6. Methodological implications and neighboring contagion models

The principal methodological implication is explicit: empirical claims of nonlinear, threshold, or complex contagion should be treated cautiously unless heterogeneity and group structure are explicitly modeled or controlled for (Hébert-Dufresne et al., 1 May 2026). The authors suggest that true superlinearity can be better distinguished by experiments, by controlling all known covariates, or by comparing kernel estimates such as $\Delta\beta(i)$1 and $\Delta\beta(i)$2 (Hébert-Dufresne et al., 1 May 2026). They also argue that inference methods that reconstruct contagion kernels from incidence data should be extended from single global-parameter models to mixtures of kernels (Hébert-Dufresne et al., 1 May 2026).

This caution matters because contagion models with explicit nonlinear structure are widely used in social systems. A distinct example is the repeated-activation model for demonstrations, protests, riots, and shifts in public opinion, in which activation is recurrent rather than absorbing (Piedrahita et al., 2017). There, each node progresses toward threshold according to

$\Delta\beta(i)$3

receives social pulses

$\Delta\beta(i)$4

activates at threshold $\Delta\beta(i)$5, then resets and can activate again (Piedrahita et al., 2017). The outcome of interest is synchronization or large-scale coordination, defined as at least $\Delta\beta(i)$6 of nodes activating simultaneously (Piedrahita et al., 2017). That model is mechanistically nonlinear in time and interaction structure, but Simpson’s contagion warns that threshold-like aggregate patterns alone do not establish such a mechanism.

A plausible implication is that observed threshold behavior should be interpreted as insufficient evidence for reinforcement unless the analysis separates within-subgroup dynamics from mixture effects. Simpson’s contagion does not deny that genuinely complex contagion exists; rather, it shows that pooled incidence curves cannot by themselves adjudicate between intrinsic nonlinearity and heterogeneous linear or sublinear mixtures (Hébert-Dufresne et al., 1 May 2026).

7. Significance, limitations, and interpretive boundaries

The significance of Simpson’s contagion lies in its unifying explanation for the ubiquity of threshold and complex contagions as effective models in aggregate data (Hébert-Dufresne et al., 1 May 2026). By showing that superlinear-looking kernels can emerge from populations composed entirely of linear or sublinear subgroups, the framework relocates the source of apparent complexity from local mechanism to population composition. The result is especially relevant for empirical settings with hidden group structure, unmeasured covariates, temporal burstiness, or correlated heterogeneity (Hébert-Dufresne et al., 1 May 2026).

Its interpretive boundary is equally clear. The result concerns what aggregate data can look like under coarse-graining; it does not imply that every observed threshold or complex contagion is illusory. Nor does it replace causal identification strategies for interference, vaccination effects, or latent homophily. Instead, it supplies a specific warning: when high-incidence observations are disproportionately drawn from subgroups with stronger transmission, the pooled curve can display threshold-like or superlinear behavior even if every subgroup is mechanistically simple (Hébert-Dufresne et al., 1 May 2026).

Within the larger Simpson’s-paradox tradition, this amounts to a domain-specific lesson about conditioning and aggregation. The common-cause formulation of Simpson’s paradox argues that one should define the relevant association by conditioning on the variable that screens off the dependence structure (Hovhannisyan et al., 2024). Simpson’s contagion applies the same logic to epidemic and social-transmission settings: if heterogeneity and subgroup structure are hidden, population-level kernels can be misleading summaries of local transmission laws.

For contagion research, the durable consequence is methodological restraint. Aggregate threshold signatures, by themselves, do not identify reinforcement, peer pressure, or complex exposure response. They may instead be the observational footprint of a Simpson’s contagion (Hébert-Dufresne et al., 1 May 2026).