Social Contagion Claims

Frontier If peer influence on health behaviour is real at the size the network literature reports — which no observational study can establish, and section 7 says what would — then every public-health pound buys more than it is credited with, because each person treated is also an untreated person’s exposure. The claim on the table is specific and it is a number: Christakis and Fowler estimated that a person’s chances of becoming obese rose by 57% if a friend became obese. If a coefficient that size is causal, treating a fifth of a population moves the whole of it. If it is not, the same budget buys exactly the people it paid for and nobody else.

Established The part of the prize that has actually been measured is a delivery saving, not a new intervention. In a 32-village cluster-randomised trial in Honduras, choosing whom to seed by friendship nomination rather than at random raised adoption of a nutritional intervention by 12.2% (95% CI 6.9 to 17.9); in a later 176-village, 24,702-person factorial experiment by the same group, nomination targeting reduced the share of households that had to be taught to reach a given level of village-wide uptake, and the effect reached a person’s friends’ friends. Those are randomised numbers. They are the floor under this subject, and they are smaller and narrower than the observational literature implies.

Speculative The upside is distributed to whoever already runs a network census, which today is a short list. Mapping every face-to-face tie in 176 villages is a research programme, not a health-system capability; the platforms that hold equivalent data for hundreds of millions of people ran the single largest influence experiment on record — 61 million people — and do not publish that data. A targeting gain that requires a full sociocentric map accrues to the few institutions that can afford one, which is a different proposition from a finding any ministry can use.

Frontier The condition is that an observational peer coefficient predicts what happens when peers are actually assigned, and the one time that was tested head-on it failed with the sign reversed. A peer-effect estimate fitted to six years of United States Air Force Academy data predicted a 0.053 grade-point gain for the low-ability students an optimised squadron assignment was designed to help; the experiment delivered −0.061. Section 5 is the list of reasons that happened, and it is the reason this brief treats the observational magnitudes as unproven rather than merely uncertain.

Established Three distinct mechanisms produce the same statistical signature in network data: contagion, homophily and shared environment. Contagion (also called induction or social influence) is the claim that one person’s behaviour changes another’s. Homophily is the tendency of similar people to become connected in the first place — two people who both intend to lose weight become gym friends, and their subsequent joint weight loss is not transmission. Shared environment (Manski’s contextual and correlated effects) is anything that acts on connected people at once: a factory closing, a new supermarket, a heatwave, a smoking ban. Every one of the three predicts that friends resemble each other and that their behaviour moves together over time.

Established The standard design regresses a person’s outcome on the lagged outcome of a named friend, conditioning on the person’s own lag and on observed covariates. In the Framingham analyses the ego is the respondent, the alter is a person the ego named, and the coefficient of interest is on the alter’s obesity at the previous examination. The intuition offered for why this identifies contagion is that the ego’s own lagged state absorbs the stable traits that drove tie formation. Established That intuition is wrong as a matter of graph structure, and the demonstration is the anchor of this brief.

Established Shalizi and Thomas proved that the three channels are generically confounded, and that no amount of covariate adjustment fixes it without assumptions nobody can check. Their abstract is blunt: they consider processes involving homophily, contagion and the causal effect of a person’s own covariates, and “show that generically, all of these are confounded with each other. Distinguishing them from one another requires strong assumptions on the parametrization of the social process or on the adequacy of the covariates used (or both).” The mechanism is a latent trait X that both causes the outcome Y and causes the tie A. Conditioning on the tie — which every network study does, by studying friends — opens a path between the egos’ and alters’ latent traits that no observed covariate closes.

Established The problem is not small samples or noisy measurement; it is non-identification. An identification failure does not shrink with n. Ten thousand more Framingham examinations, or a hundred million more platform users, produce a tighter confidence interval around a quantity that is not the contagion effect. This is the single most important distinction in the subject and it is routinely lost in coverage, which reports the sample size as if it were evidence of causal strength.

Frontier The escape route that works is to randomise the peer rather than the treatment. If who your peers are is assigned by lottery, the homophily channel is closed by construction: the tie no longer depends on the latent trait, so the collider path Shalizi and Thomas describe does not exist. This is why the credible evidence in this field comes from military academies, university housing lotteries and village-level cluster trials rather than from observational friendship panels, and why section 7 asks for more of it rather than for better statistics.

Speculative A fourth mechanism, rarely separated from the other three, is measurement-induced correlation. Friends recall each other’s behaviour and in some designs name each other; shared reporting error produces correlated outcomes with no behavioural content at all, and the literature has no standard correction for it.

Established The founding result is precisely stated and was correctly reported. Christakis and Fowler analysed 12,067 people in the Framingham Heart Study social network from 1971 to 2003 and reported that “a person’s chances of becoming obese increased by 57% (95% confidence interval [CI], 6 to 123) if he or she had a friend who became obese in a given interval.” Among adult sibling pairs the figure was 40% (95% CI 21 to 60); between spouses, 37% (95% CI 7 to 73). Effects were not seen among immediate geographic neighbours, same-sex pairs showed greater association than opposite-sex pairs, and clusters of obese people were described as extending to three degrees of separation. This brief’s commission asked whether the 57% anchor is right: it is, verbatim, and the confidence interval — 6% to 123% — is worth as much attention as the point estimate, because a lower bound of 6% is consistent with almost nothing.

Established The same method was then applied to smoking, drinking, happiness, loneliness, divorce and sleep, and produced network effects every time. The Framingham series includes the collective dynamics of smoking (Christakis and Fowler, 2008), the dynamic spread of happiness (Fowler and Christakis, 2008), loneliness (Cacioppo, Fowler and Christakis, 2009) and alcohol consumption behaviour (Rosenquist and colleagues, 2010). Frontier That the method never returns a null is itself a diagnostic, and it is the observation that motivated the decisive critique.

Established Cohen-Cole and Fletcher ran the method on outcomes that cannot be socially transmitted and recovered contagion anyway. Using 4,300 to 5,400 adolescents from the National Longitudinal Study of Adolescent Health, they reported that “a friend’s acne problems increased an individual’s odds of acne problems (odds ratio 1.62, 95% confidence interval 0.91 to 2.89)”, that “the likelihood that an individual had headaches also increased with the presence of a friend with headaches (1.47, 0.93 to 2.33)”, and that “an individual’s height increased by 20% of his or her friend’s height (0.18, 0.15 to 0.26)”. After adjusting for environmental confounders the estimates became “uniformly smaller and insignificant”. Height is the load-bearing case: adult height is fixed years before the friendship and cannot be caught. A method that finds contagious height has a specification problem, not a discovery.

Established The theoretical result behind the reductio is a proof of non-identification, published in the same journal issue as its principal rebuttal. Shalizi and Thomas, in Sociological Methods & Research 40(2), 211–239 (2011), show that asymmetries in regression coefficients cannot identify causal effects and that “very simple models of imitation (a form of social contagion) can produce substantial correlations between an individual’s enduring traits and his or her choices, even when there is no intrinsic affinity between them”. In other words the homophily-versus-contagion inference runs in both directions: contagion can manufacture apparent homophily, and latent homophily can manufacture apparent contagion. Their own posted simulation code reproduces both.

Frontier Lyons pressed a harder claim — not that the estimates are confounded but that they do not follow from the model at all. In Statistics, Politics, and Policy (2011) he argues that the logistic specification is internally inconsistent when an ego names several alters, that the directional contrast used as evidence for induction is not statistically significant, and that the directional pattern is “actually consistent with all three considered explanations, i.e., induction, homophily, and environment”. Frontier The directionality test is the crux: Christakis and Fowler argued that a friendship named by the ego but not the alter shows a weaker effect, as induction predicts and homophily does not, and Lyons’s reply is that the two coefficients’ confidence intervals overlap.

Frontier The most careful defence of the observational estimates is a sensitivity analysis, and it splits the results rather than rescuing them. VanderWeele asked how strong unmeasured homophily or confounding would have to be to explain each finding away. His conclusion is specific: “contagion effects for obesity and smoking cessation are reasonably robust to possible latent homophily or environmental confounding; those for happiness and loneliness are somewhat less so”, while “supposed effects for height, acne, and headaches are all easily explained away by latent homophily and confounding”. Frontier This is the strongest live argument that something survives, and it is genuinely an argument rather than a proof: a sensitivity analysis states what would have to be true for the result to vanish, and cannot say whether it is true.

Frontier The methodological dispute is unusually well documented because the participants published their disagreements about each other’s corrections. VanderWeele, Ogburn and Tchetgen Tchetgen argued in 2012 that “flawed” network analyses can still yield valid tests of the null of no contagion; Shalizi replied that the paper is “a useful contribution to the on-going scientific conversation about the detection of contagion from purely observational data” which corrects “some of the more extreme statements of Lyons (2011)” but which “too, goes too far in some places, and so needs some correction itself”. A reader looking for a settled answer will not find one; a reader looking for the state of the argument has an unusually complete record.

Frontier A constructive line does exist: bound the effect instead of estimating it. Ver Steeg and Galstyan demonstrate “a general method to lower bound the strength of causal effects in observational social network studies, even in the presence of arbitrary, unobserved individual traits”, with no parametric assumptions and an algebraic proof per test, and report preliminary results on Framingham data. Their framing is the right one: “if the true goal is to test for the presence of contagion, a lower bound on the strength of contagion is all that is necessary”. Speculative Bounds have not displaced point estimates in the applied literature, and the reason is presentational rather than technical — a bound does not produce a headline percentage.

Frontier The most recent reanalysis of the original Framingham data reports no evidence of causal peer effects. Ogburn, Sofrygin, Díaz and van der Laan develop semiparametric estimation for a single network, allowing dependence both through transmission across ties and through latent similarity among tied nodes, and state: “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.” Speculative This brief reports that conclusion as the authors state it and does not report numbers from the application section, because the preprint versions differ in whether that section is present and the figures could not be confirmed twice.

Established Where peers are assigned rather than chosen, effects are real, measurable and mostly smaller than the observational estimates. Sacerdote’s Dartmouth study, using randomised freshman housing, found peer influence on grade point average and on fraternity membership but none on choice of major, with GPA effects operating at the room level. Kremer and Levy exploited a roommate lottery at a large state university (1,357 students, 1997–1998 entering classes) and found that males assigned a roommate who drank before college had roughly a quarter of a grade point lower GPA — coefficients of −0.282 (SE 0.128) for a frequent-drinking roommate and −0.263 (SE 0.101) for an occasional-drinking one — while “students’ college GPA is not influenced by roommates’ high school grades, admission test scores, or family background” and female GPAs were unaffected. That pattern — a behaviour transmits, a characteristic does not — is exactly what an honest contagion test should look like.

Established The largest randomised peer-group design is the United States Air Force Academy, where freshmen are assigned to squadrons of about thirty and physically cannot leave them. Carrell, Fullerton and West studied roughly 3,490 students across 108 squadrons in the classes of 2005–2007 and found a peer SAT-verbal coefficient of 0.416 (SE 0.103) on freshman GPA, effects concentrated in mathematics and science courses and near zero in physical education and foreign language, and persistence at declining magnitude into later years. Randomisation held: of 90 coefficients testing correlation between own and peer characteristics, six were significant at the 0.10 level.

Established The same team then used that estimate to design squadrons, and the experiment reversed the sign. Carrell, Sacerdote and West assigned about half of the entering classes of 2011 and 2012 — roughly 2,600 students across 40 squadrons — by an algorithm built to maximise the performance of the lowest-ability students. The model predicted “a statistically significant 0053 grade points” for the bottom third; the realised effect was “a negative and statistically significant treatment effect of −0061 (p = 0055)”, with a positive 0.082 (p = 0.041) for middle-ability students who were predicted to be unaffected. Established The decimal points are lost in the document’s text layer; the numbers are +0.053 predicted and −0.061 realised. The mechanism the authors identify is endogenous group formation: engineered squadrons re-sorted internally, and the low-ability students ended up with each other.

Frontier Angrist argues that a large part of the peer-effects literature is measuring its own algebra. His analysis shows that “outcome-on-outcome peer effects are vacuous, either unity or, if the average is defined as a leave-out mean, determined by a generic intraclass correlation coefficient”, and that once individual covariates are added the estimated peer effect becomes a test for the equality of two-stage least squares and ordinary least squares — a quantity that moves with measurement error, selection and weak instruments for reasons with no social content. His Dartmouth illustration is that the apparent peer effect rises from about 0.098 at room level to 0.145 at floor level to 0.232 at building level, which he attributes to many-weak-instrument behaviour rather than to stronger social forces.

Frontier Observational structural models can still find effects on health behaviours, at sizes an order of magnitude below the Framingham percentages. Card and Giuliano modelled best-friend pairs in Add Health and found that “the likelihood that one friend initiates intercourse within a year of the baseline interview increases by almost 5 percentage points (on a base rate of 11%) if the other also initiates intercourse”, with similar effects for smoking, marijuana use and truancy. Speculative Their identification rests on exclusion restrictions and a bivariate normal distribution for unobserved heterogeneity — which is precisely the “strong assumptions on the parametrization” that Shalizi and Thomas say identification requires. The estimate is not wrong; it is conditional on a model nobody can test.

Established The single largest influence experiment ever run found social transmission larger than the direct effect, on one behaviour, once. Bond and colleagues randomised political mobilisation messages to 61 million Facebook users on the 2010 United States congressional election day and reported that “the effect of social transmission on real-world voting was greater than the direct effect of the messages themselves, and nearly all the transmission occurred between ‘close friends’ who were more likely to have a face-to-face relationship”. This is a randomised result and it belongs in the positive column. Frontier It is also one behaviour, one day, one platform, with validated turnout records — a combination no health outcome has yet matched.

Established Village-scale cluster trials give the cleanest health-behaviour evidence, and the effects are modest and outcome-dependent. Kim and colleagues randomised 32 Honduran villages (5,773 participants) to three seeding rules — random 5%, highest in-degree 5%, or friends nominated by random villagers — for two interventions. Nomination targeting “increased adoption of the nutritional intervention by 12·2% compared to random targeting (95% CI, 6·9 to 17·9)”, with multivitamin redemption of 74.3% versus 66.2%. For water chlorination there was no significant difference at all (55.6% random, 55.9% in-degree, 54.5% nomination). One product spread; the other did not.

Established The largest structured network experiment to date mapped 24,702 people in 176 isolated villages and measured spillovers directly. Airoldi and Christakis (Science 384(6695), eadi5147, May 2024) randomised villages across targeting fractions of 0, 5, 10, 20, 30, 50, 75 and 100% of households and across two targeting methods, delivering a 22-month health education package and assessing 117 knowledge, attitude and practice outcomes. Their own summary: “Friendship-nomination targeting reduced the number of households needed to attain specified levels of village-wide uptake. Knowledge spread more readily than behavior, and spillovers extended to two degrees of separation. Outcomes that were intrinsically easier to adopt also manifested greater spillovers.” Frontier Knowledge moving more easily than behaviour is the finding with the most policy content in it, and it is the opposite of what the obesity-contagion framing implies.

Established Seeding algorithms also work in dense urban settings, on a different outcome. Alexander, Forastiere, Christakis and Gupta randomised 2,491 women across 50 Mumbai chawls to random targeting of 20%, friendship nomination, pair targeting, or control, for adoption of iron-fortified salt; both targeting algorithms enhanced adoption. Frontier No magnitude is asserted here, because the published effect sizes could not be confirmed twice.

Frontier Randomising the network itself, rather than the seeds, is the cleanest design available and has been done outside health. Fafchamps and Quinn linked managers of African manufacturing firms by exogenous network formation and information seeding across three countries; diffusion of business practice appeared only for “VAT registration and having a bank current account”, and the authors note concurrent VAT legislation probably contributed. Speculative Assigning ties rather than treatments is the direct experimental analogue of the identification problem, and it has been run on firms and students far more often than on health behaviours.

Established Where the field now stands: the non-identification result is not disputed, the reductio is not disputed, and what remains contested is whether any particular observational estimate survives sensitivity analysis. Ogburn’s 2018 review states the position without hedging: “the methods rely on assumptions that are likely — and sometimes guaranteed to be — false, and therefore the evidence for the conclusions is often weaker than it seems. Developing methods that do not need to rely on implausible assumptions is an incredibly challenging and important open problem in statistics.”

Frontier The open question is not whether social influence exists but whether any observational estimate of its size is recoverable. Nobody in this dispute argues that people are uninfluenced by the people around them; the randomised evidence in section 3 settles that for drinking, for voting and for adopting a nutritional supplement. The question is whether a coefficient fitted to a friendship panel means anything, and the honest answer is that it means something only under assumptions that cannot be checked from the data that produced it.

Frontier Does the obesity result survive? Two careful analyses of the same data reach opposite conclusions: VanderWeele’s sensitivity analysis calls the obesity and smoking-cessation effects “reasonably robust”, while Ogburn and colleagues’ semiparametric reanalysis finds “no evidence for causal peer effects” once network structure is accounted for. Frontier These are not the same estimand and that is part of the problem: the field has no agreed target parameter for “the contagion effect” in a single observed network, so two defensible analyses can disagree without either being wrong.

Speculative How much of the observed clustering is measurement artefact is essentially unstudied. The Framingham network was reconstructed from administrative contact sheets naming a person who would know the respondent’s whereabouts — a name generator built for follow-up, not network science. Ties are sparse, directed and truncated, and what that truncation does to a contagion coefficient has no published quantification.

Frontier Whether bounds can replace point estimates is a live technical question with an unfavourable incentive gradient. Ver Steeg and Galstyan show assumption-free lower bounds are obtainable; the applied literature has not adopted them. Speculative A journal culture that rewards “57%” over “the effect is bounded below by something small and positive” is the binding constraint here, not the mathematics.

Frontier Whether negative controls and proxy variables can identify peer effects is the most active current line. The idea is to find variables known not to be affected by the peer’s outcome but affected by the same latent confounder, and use them to net out the confounding — the same logic that turned negative-control outcomes into a standard epidemiological tool. Speculative Whether a valid negative control exists in a friendship network is a subject-matter question that statistics cannot answer; the height result in Cohen-Cole and Fletcher is in effect a negative control that failed.

Frontier Whether experimental peer effects generalise out of the institution that produced them is unresolved and the evidence is discouraging. The Air Force Academy reversal is not a failure of randomisation; it is a failure of extrapolation from a randomised estimate to a policy that changes the assignment rule. Randomisation identifies the effect of the peers people were given under the old rule. It does not identify what happens when the rule changes and people re-sort.

Frontier How far spillovers travel is contested at the second and third degree. Christakis and Fowler described obesity clusters extending to three degrees of separation; the randomised Honduras experiment detected effects to two degrees. Speculative Whether the third degree exists for any health behaviour under randomisation is currently unanswered, and the statistical power required to answer it grows steeply because the number of potentially exposed units rises faster than the effect size falls.

Frontier Whether “complex contagion” changes the identification problem or only the model is open. The complex-contagion line holds that behaviours requiring social reinforcement spread better in clustered networks with redundant ties than along long weak ties, and rests on randomised network-structure experiments in online settings. Speculative Threshold models make the diffusion prediction sharper but do not by themselves close the confounding path in observational data; a threshold fitted to a friendship panel inherits every problem a linear coefficient has.

Speculative Whether the estimator literature is converging or diversifying is unclear. New structural models continue to appear — quantile peer-effect specifications in 2025, for example — each identified under its own parametric assumptions. Handwave The optimistic reading is that the assumptions are getting weaker; nothing in the published record demonstrates that.

Established The binding constraint is network measurement, and it is expensive in a way that scales badly. A sociocentric census — every tie among everyone in a bounded population — is what the identification argument needs and what almost no dataset has. The Honduras programme mapped 24,702 people across 176 villages; the cost of that is a research budget, and the design only works because the villages are isolated enough that the boundary is real.

Established Name generators truncate, and truncation is not random. Framingham’s ties came from contact sheets naming a person who would know the respondent’s whereabouts. A design that permits one or two names systematically records the strongest tie and discards the rest, so the recorded network is a biased sample of the real one, biased toward exactly the relationships in which homophily is strongest.

Frontier Ties dissolve, and they dissolve selectively. Noel and Nyhan named this the unfriending problem: if friendships that stop being similar also stop being friendships, then a panel of surviving friendships is selected on continued similarity, and the surviving pairs look influenced whether or not anyone influenced anyone. Panel attrition in a network is not the same problem as panel attrition in a survey, and the standard corrections do not apply.

Established Outcome measurement is the second constraint, and it is why the credible results cluster on a narrow set of behaviours. Voting is validated against public records; grade point average is an administrative record; product redemption is a counted ticket. Weight, diet, activity and drinking are self-reported at long intervals, and measurement error correlated within friendship pairs reproduces contagion.

Frontier Statistical power for spillover estimands is far worse than for direct effects. The Honduras factorial design needed 176 villages to estimate a spillover surface. A trial powered for a direct effect is typically not powered for the indirect one, which is why so many network trials report the seeding comparison rather than the transmission parameter underneath it.

Established Interference breaks the statistical machinery that most trials are built on. The stable unit treatment value assumption — that one unit’s outcome does not depend on another’s treatment — is false by construction in any study of contagion. The causal-inference-under-interference literature has supplied estimands and estimators for this, but they require the exposure mapping to be specified correctly, which requires the network to be measured correctly, which returns to the first constraint.

Frontier Access to the one class of data that could settle several of these questions is controlled by firms. The 61-million-person mobilisation experiment demonstrates that platforms can randomise exposure, observe the full graph and validate outcomes against administrative records. Speculative The same capability sits behind commercial confidentiality and consent constraints, and independent replication of platform influence results is largely unavailable — a problem this corpus treats more fully in Information Integrity.

Speculative Endogenous re-sorting is a design bottleneck with no general fix. The Air Force Academy result shows that assigning people to a group does not assign them their friends inside it. Unless the design can observe or constrain within-group tie formation, a randomised peer-group experiment estimates the effect of an assignment rule, not the effect of a peer.

Established Nothing here waits on a physical technology; every input is a design, a dataset or an institution. That is unusual in this corpus and it changes what a workback plan looks like: no instrument has to be invented, no material has to be discovered, and the reason the answer is not in hand is that the studies capable of producing it have not been commissioned.

Established The first dependency is a measured network, not an inferred one. Every identification strategy in the literature — bounds, negative controls, exposure mappings, semiparametric estimation on a single network — assumes the observed adjacency matrix is the real one. Sociocentric network collection at population scale is the prerequisite, and it is a survey-methodology problem before it is a statistics problem.

Frontier The second dependency is agreement on the estimand. Contagion has been operationalised as a lagged-alter regression coefficient, as a spillover contrast between treated and untreated units in a cluster trial, as a threshold in a diffusion model and as an intervention on ties. These are different quantities with different policy meanings, and comparing them across papers is currently a category error that the field commits routinely.

Frontier The third dependency is a usable negative control for each health behaviour studied. Height worked as a negative control because nobody believes it transmits. An analogous control is needed for each outcome that does — a trait that shares the confounders but cannot be caught — and constructing one is subject-matter work that statisticians cannot do alone.

Established The fourth dependency is institutional access to a randomisation that already happens. Military academies, service academies with squadron assignment, university housing lotteries, oversubscribed school lotteries, public-housing allocation and some conscription systems already assign people to peers by chance. The randomisation exists; what is missing is a protocol that attaches pre-registered health-behaviour endpoints to it.

Frontier The fifth dependency is replication capacity, which this corpus treats as a general problem. The methodological record here is unusually good precisely because several groups reanalysed the same data and published the disagreement; that is the exception, and the general case is discussed in Forecasting Replication and Science of Science.

Speculative A sixth dependency is shared simulation infrastructure for checking whether an estimator recovers a known truth. Shalizi and Thomas posted their simulation code, which is why their claims can be checked in an afternoon. A standard benchmark of networks generated with known contagion, homophily and environment would let any new estimator be tested before it meets real data. Nobody maintains one.

Established The decisive experiment is to randomise peers rather than treatments, across the health behaviours where contagion claims carry the most policy weight, and to pre-register the endpoints before the assignment is made. Randomising who your peers are closes the homophily channel by construction: the tie no longer depends on the latent trait, so the confounding path Shalizi and Thomas describe is absent rather than adjusted for. Everything else in this brief is an argument about what to believe in the absence of that design.

Frontier The specification that would settle the question: several thousand people assigned at random to living or working groups inside an institution that already randomises, with smoking, drinking, physical activity and measured body weight as pre-registered endpoints, follow-up of at least two years, and power to detect effects an order of magnitude below the observational estimates. Body weight measured rather than reported, alcohol from administrative or biomarker sources where available, and within-group tie formation observed so that re-sorting can be modelled rather than assumed away.

Established Fragments of this design already exist and each one is informative. Kremer and Levy’s roommate lottery is the closest health-behaviour case: random assignment, a drinking exposure, and a clean outcome, with −0.282 (SE 0.128) on GPA for males assigned a frequent-drinking roommate and a reported rise in own drinking. Sacerdote’s Dartmouth lottery supplies the template. Carrell, Fullerton and West supply the scale — 3,490 students, 108 squadrons — on academic rather than health outcomes.

Established The experiment that most changed this field’s self-understanding was a failure, and it should be run again on health endpoints. The Air Force Academy squadron-optimisation trial took an observational peer coefficient, designed groups from it, and got the opposite sign for the students it was designed to help. Frontier The equivalent test for health behaviour — fit a contagion model to an observational network, use it to design an assignment, and see whether the predicted spillover appears — has never been run, and it is the sharpest available test of whether observational contagion estimates carry policy content.

Frontier The second decisive design is to randomise ties rather than groups. Fafchamps and Quinn did this for firm managers by creating connections that would not otherwise have existed; the health analogue would assign peer pairings within a cohort and track behaviour change in both directions. Speculative This is the design that most directly instruments the adjacency matrix, and no published health-behaviour version exists.

Established The village-scale factorial experiment is already running and is the best-powered spillover measurement in existence. The 176-village, 24,702-person Honduras design varies the fraction treated from 0 to 100% and the targeting rule, which is what makes spillover separable from direct effect. Its headline results — knowledge spreading more readily than behaviour, spillovers reaching two degrees — are the kind of output this brief treats as evidence.

Frontier A cheap and unambiguous test that has not been systematically run: apply every new contagion estimator to a negative-control outcome before applying it to a real one. Cohen-Cole and Fletcher did this once, informally, and the result reshaped the field. Making it a routine gate — height, or an equivalently untransmissible trait, reported alongside every contagion estimate — costs nothing and would retire a class of false positives immediately.

Frontier A benchmark of simulated networks with known ground truth would settle the estimator dispute without any new fieldwork. Generate populations under known mixtures of contagion, homophily and shared environment; require any proposed estimator to recover the mixture. Speculative The reason to expect this to be informative is that the existing simulations already discriminate: Shalizi and Thomas’s neutral-diffusion simulation produces apparent homophily from pure imitation, which most published estimators would misread.

Speculative A registered adversarial reanalysis of the Framingham data by both camps, with the analysis plans fixed in advance, would convert a decade of disagreement into a result. The data exist, the two analyses already disagree, and the disagreement is currently unresolvable because each was specified after its authors knew what the other found.

Handwave What would not settle it: more observational networks, larger platform datasets, or a better fitting model on existing data. Every one of these increases precision around a parameter that is not identified. This is the step where the field’s implicit argument works by assertion — the assumption that scale substitutes for design — and it is the assumption the non-identification proof refutes.

Established The engineering is survey instrumentation, randomisation machinery and outcome plumbing, not hardware. A working programme needs three components: a sociocentric name generator that captures enough ties to make the adjacency matrix usable, a randomisation engine embedded in an institution that already assigns people, and outcome capture that does not rely on the respondent remembering.

Established Name-generator design is the component with the highest leverage and the least attention. Framingham’s single-name administrative contact is the limiting case; the Honduras programme used photographic rosters and multiple relation-specific prompts to elicit named ties across a whole village. The difference between those two instruments is larger than the difference between most competing estimators.

Frontier Outcome capture should move from recall to record wherever possible. The results here that carry weight have hard endpoints: validated voter files, registrar grade records, redeemed product tickets, measured weight. Speculative Wearable and dispensing records could supply equivalents for activity and adherence; no network trial has used them as primary endpoints at scale.

Established The statistical engineering is exposure-mapping and interference-robust inference, and the tooling now exists. Randomisation tests for group-formation experiments were formalised by Basse, Ding, Feller and Toulis in Econometrica in 2024: they extend Fisher randomisation tests to the setting where individuals are randomly assigned to groups, note that naive exposure permutations often violate the underlying design, and show that tests built on an equivariance property are “justified by the randomization itself, require relatively few assumptions, and are exact in finite samples”. Their applications are roommate assignment and CEO meeting groups.

Frontier Semiparametric estimation on a single network is the other tooling advance. Ogburn and colleagues’ asymptotics are “the first to allow for dependence of each observation on a growing number of other units as sample size increases” and permit both transmission-driven and latent-similarity-driven dependence at once, which previous methods did not. That is a genuine engineering result and it is the reason their Framingham reanalysis is worth taking seriously.

Speculative The missing piece is a shared protocol that institutions can adopt without becoming research organisations. The cost of a randomised peer study inside an institution that already randomises is mostly the endpoint collection and the ethics work, not the randomisation. A standard package — consent language, endpoint definitions, network instrument, pre-registration template — is a modest piece of engineering that nobody has built, and it is the same pattern of missing connective infrastructure that Public Policy Foresight and Information Integrity identify in adjacent evidence domains.

Established The closest methodological neighbour is the causal-inference-under-interference literature, which shares the mathematics and has better-behaved applications. Vaccine trials face the identical problem — one person’s vaccination protects another — and solved it with cluster-randomised designs and explicitly defined indirect, total and overall effects. Infectious-disease epidemiology is where the word contagion is doing literal work, and the estimands it developed transfer.

Frontier Platform experimentation is adjacent and ahead on design, behind on access. Online field experiments can randomise exposure at a scale no field study can match; the 61-million-person mobilisation trial is the canonical example. What they cannot supply is independent verification, and the evidentiary consequences of that asymmetry are developed in Information Integrity.

Frontier Genetic epidemiology solved a structurally similar problem with sibling designs and Mendelian randomisation. Both work by finding variation that cannot have been selected on the confounder. The analogue here is an assignment people did not choose, which is what a peer lottery is.

Speculative Agent-based diffusion models are adjacent but not evidential. Reproducing observed clustering from assumed contagion rules shows sufficiency, never necessity; the homophily-only simulation usually reproduces it too.

Established The peer-effects literature in education is the methodological parent and supplies both the best designs and the sharpest warnings. Sacerdote’s roommate lotteries, the Air Force Academy squadrons and Angrist’s critique were all developed on academic outcomes and imported into health afterwards.

Speculative Exposure science is an under-exploited neighbour. Much of what looks like contagion between friends is shared environment, and the discipline that measures shared environment quantitatively is exposomics — see Exposomics. A contagion study with a real exposure measurement would be able to subtract the environmental channel instead of assuming it away.

Established The institutional record here is better than the field’s reputation suggests, because the critiques were published, cited and answered. The non-identification proof appeared in the same 2011 issue of Sociological Methods & Research as VanderWeele’s sensitivity analysis; the reductio appeared in the BMJ; the rebuttals and counter-rebuttals appeared in Statistics, Politics, and Policy.

Frontier What the institutions did not do is change the downstream claim. The 57% figure continued to circulate in public health communication and in popular science after the identification critique was published and largely conceded at the level of method. The correction mechanism worked inside statistics and did not propagate outward.

Established Data access is the central institutional constraint and it is unevenly distributed. The Framingham network data can be reanalysed, which is why Ogburn and colleagues could run their reanalysis and why the dispute is resolvable in principle. Platform graph data cannot. A field whose largest experiment cannot be replicated outside the firm that ran it has an institutional problem prior to its statistical one.

Frontier Ethics review is a real constraint on the decisive design, and a legitimate one. Randomising a roommate already happens routinely; doing it to study whether someone starts drinking needs consent architecture most review boards have not met. The obstacle is procedural rather than principled, and not trivial.

Frontier The incentive structure rewards the estimate and not the check. A contagion coefficient is publishable; a negative-control result showing the estimator finds contagious height is publishable once. Making the negative control a submission requirement is a journal policy, not a research programme — the same class of low-cost institutional fix catalogued in Scientific Advisory Institutions.

Speculative Registered reports would bite unusually hard here. The analytical choices — which lag, which covariates, which tie definition, which directionality contrast — are numerous and consequential, so the gap between a pre-specified and a post-specified network analysis is larger than in most observational epidemiology.

Established Acting on a non-identified estimate has a cost, and the Air Force Academy trial measured it. Students were assigned to squadrons designed from a peer-effect model in order to help them, and the students it was designed to help did worse. This is not an abstract risk; it is the documented outcome of taking an observational social coefficient seriously enough to act on.

Frontier Network targeting raises a consent question that individual-level targeting does not. A seeding algorithm chooses a person because of their position in a graph built from other people’s nominations. The nominators did not consent to being an input to someone else’s selection, and they are not the ones who receive the intervention. The Honduras and Mumbai trials show this is workable in practice; it is not the same consent object as a conventional trial.

Frontier Experiments that manipulate influence at platform scale sit on contested ethical ground and that contest has shaped what is publishable. The mobilisation experiment reached 61 million people; the reaction to platform emotion and influence experiments in the following years narrowed what firms would run and publish. The result is that the best-instrumented setting for studying social influence is also the least open one.

Established The identification problem is worst precisely where the claims are most politically charged, and this is a structural fact rather than a rhetorical one. Three conditions make identification hardest: the outcome is self-reported rather than administratively recorded; ties are formed on the basis of the trait in question, so homophily is maximal; and assigning peers at random is impossible or unethical. Contested claims about identity, belief and socially sensitive behaviour satisfy all three at once. Obesity, drinking and smoking satisfy at most two, and even there the observational estimates do not survive reanalysis.

Established This brief therefore restricts itself to the health and behaviour outcomes the literature actually tests with randomised designs. Body weight, smoking, alcohol, exercise, product adoption, voting, academic performance and school conflict are the domains where peers have been assigned by lottery and outcomes recorded independently. Extending a contagion framework to domains with no randomised evidence and maximal homophily would reproduce exactly the error the Cohen-Cole and Fletcher height result exposed, and with fewer checks available.

Frontier The cost of being wrong is asymmetric and the field does not price it. Overstating contagion justifies interventions aimed at people’s associations rather than their circumstances; understating it wastes the multiplier the randomised village trials do produce. Speculative The first error is the more consequential, because it licenses acting on people who were never treated and never consented.

Speculative Framing matters for blame. “Obesity is socially contagious” relocates causation from food environments and incomes into friendships, which is a claim about whom to hold responsible as well as a claim about mechanism. The environmental adjustment in Cohen-Cole and Fletcher — which made the estimates “uniformly smaller and insignificant” — is the direct evidence that the relocation is at least partly unwarranted.

Frontier The durable civilizational content of this subject is a lesson about identification, not a fact about obesity. A famous, well-executed, correctly reported study produced a number that was almost certainly not the quantity its readers took it for, and the reason was structural: the design cannot separate the mechanisms. That pattern recurs wherever observational data on connected units meets a causal question, which is most of computational social science.

Established Public health has nonetheless banked a real capability from this line of work: cheaper delivery through network-aware targeting. The 12.2% adoption gain from friendship-nomination seeding in Honduras and the household-count reduction in the 176-village experiment are randomised, replicated across two study designs and one urban replication, and do not depend on any observational contagion estimate being correct. A programme can use them without believing the 57%.

Frontier The scaling limit is that targeting requires a map. Nomination-based seeding works because nominating a friend is cheap; in-degree targeting requires knowing the graph, and in the Honduras trial it did not outperform random targeting for chlorine. Speculative A health system that wants the targeting gain at national scale needs either a nomination protocol that survives bureaucratic delivery or a graph it is not politically able to build.

Speculative If contagion effects are real but small, the implication is about compounding rather than magnitude. A 5 percentage point transmission probability is negligible for one pair and substantial across a population over decades. Handwave Getting from a per-pair probability to a population trajectory needs a diffusion model whose parameters come from the confounded data, so the compounding argument is an assertion.

Frontier The negative civilizational consequence is a governance one: contagion language licenses intervening on associations. Once a behaviour is described as spreading, the natural policy instrument becomes the network rather than the person — whom to seed, whom to separate, whom to target. The evidence supports seeding; it does not support separating, and the Air Force Academy result is direct evidence that engineered regrouping can harm the people it targets.

Speculative The most valuable long-run output may be the negative-control habit. A field that routinely reports what its estimator says about an untransmissible trait has an internal falsification test that most observational social science lacks — the same move that Pathological Science identifies as the difference between a discipline that self-corrects and one that does not.

These horizons track study designs, data access and publication norms rather than any technology; every instrument required already exists:

• 10 yr: Frontier Expect the estimator literature to continue producing identified-under-assumptions methods and the applied literature to continue publishing point estimates. Speculative The village-scale factorial experiments are the likeliest source of a decisive result in this window because they are already running and already measure spillovers directly. Frontier A randomised peer-assignment study with pre-registered health endpoints is feasible inside a normal grant cycle — three to five years from funding to first endpoints — and the binding constraint is commissioning, not capability. Speculative Negative controls becoming a routine reporting requirement is a journal-policy change that could happen at any point and would be the cheapest improvement available.
• 25 yr: Speculative If administrative and sensor endpoints replace self-report for weight, activity and adherence, the measurement-error channel that currently manufactures contagion becomes testable rather than assumed. Speculative A settled estimand for contagion in a single observed network is plausible on this horizon, and would make the existing literature comparable for the first time. Frontier Whether platform graph data becomes independently auditable is a regulatory question rather than a scientific one, and the analysis in Information Integrity applies directly.
• 50 yr: Speculative Population-scale sociocentric measurement, if it happens at all, will be a by-product of administrative data integration rather than of survey research, and will arrive with governance problems larger than the scientific ones it solves. Handwave Whether that data would actually identify contagion is unclear: a complete graph removes measurement truncation but does nothing about latent homophily, which is the binding problem.
• 100 / 250+ yr: Handwave On this horizon the interesting question is whether “social contagion” survives as a category at all, or whether it is decomposed into distinct mechanisms — imitation, coordination, information transfer, shared exposure — each with its own identification strategy, in the way that “miasma” was decomposed rather than refuted. Handwave Nothing in the current literature constrains this and it is offered as a conjecture about scientific vocabulary, not a prediction.

• Depends on Nothing on this map blocks it, and that is the finding: the binding constraints are study design, data access and publication norms, all of them decades-old capabilities. Its nearest kin here, Information Integrity, shares the structural diagnosis — a field that measures supply well and transmission badly — and supplies the platform-data-access argument this brief leans on, but produces no result this topic waits on.
• Requires (not on this map) A randomised peer-assignment trial with pre-registered health-behaviour endpoints, which is the one design that closes the homophily channel by construction and which nobody has run at the required scale. A validated negative control for each transmissible outcome — the role height played once, generalised — without which no estimator can be falsified on real data. A standing randomisation protocol that institutions already assigning people by lottery could adopt without becoming research organisations. An independent audit route to platform graph data, absent which the largest influence experiments stay unreplicable. And funding that pays for decisive replication rather than for another novel network finding.
• Enables A defensible prior for any policy that counts on spillover — network-targeted vaccination and screening campaigns, peer-delivered behaviour change, school and housing assignment rules. No typed enabling edge is asserted, because on current evidence the field cannot state the size of the peer effect for any health behaviour, and a typed edge claiming one would be the overconfidence this brief documents.
• Adjacent Causal inference under interference; the education peer-effects literature; infectious-disease epidemiology, where contagion is literal; and within this map Information Integrity, Forecasting Replication, Science of Science, Exposomics, Institutions and Growth and Pathological Science.

Established “Obesity is contagious, and the figure is 57%.” The figure is real and correctly reported, and it is an estimate of a quantity that the design cannot separate from homophily and shared environment. Its own confidence interval runs from 6% to 123%. The most recent reanalysis of the same data reports no evidence for causal peer effects once network structure is accounted for; a sensitivity analysis by a sympathetic critic calls the obesity effect “reasonably robust”. The honest summary is that two competent analyses of one dataset disagree, and that no experiment has tested the claim.

Established “Your friends’ friends’ friends make you fat — the three degrees of influence rule.” Three degrees describes the extent of observed clustering in the Framingham network, which is a description of the data and inherits every identification problem the coefficient does. The randomised evidence reaches two degrees, in villages, for an intervention deliberately designed to spread. No experiment has demonstrated a third-degree effect on a health behaviour.

Established “They controlled for everything, so it must be causal.” The Framingham analyses adjusted for a long list of covariates. Shalizi and Thomas show the confounder in question is latent by definition: if the trait that drives tie formation were observed, it could be adjusted for, and the entire dispute would not exist. Adding observed covariates does not touch the unobserved path.

Established “The directionality test rules out homophily.” This is the specific defence Christakis and Fowler offered — that a friendship named by one party and not the other shows a weaker association, as induction predicts and homophily does not. Lyons’s reply is that the contrast is not statistically significant and that the observed directional pattern is “actually consistent with all three considered explanations, i.e., induction, homophily, and environment”. Shalizi and Thomas make the general version of the point: asymmetries in regression coefficients cannot identify causal effects.

Established “The acne and height critique was a joke paper.” It was a peer-reviewed longitudinal analysis in the BMJ using a nationally representative adolescent panel, and the effect sizes are reported with confidence intervals. It is a reductio in the strict sense: the method is applied unchanged to a case where the conclusion is known to be false, and it returns the false conclusion. Adult height does not travel along friendships.

Frontier “The critique shows social influence does not exist.” It shows no such thing, and this brief takes the opposite view. Randomised roommates change drinking and grades; randomised mobilisation messages moved real votes through friends more than directly; randomised village seeding raised supplement adoption by 12.2%. What the critique establishes is that the observational estimates of the size of influence are not identified, which is a claim about measurement, not about human beings.

Frontier “Network-targeted public health works, so contagion must be strong.” Targeting gains and contagion magnitudes are different quantities. Nomination targeting beat random targeting by 12.2% for multivitamins and by nothing at all for chlorine in the same trial; the Honduras factorial experiment found knowledge spreading more readily than behaviour. A targeting gain can be produced by a modest transmission probability acting on a well-chosen seed set.

Frontier “Randomised peer studies settle it.” They identify the effect of the peers people were assigned under a given rule. The Air Force Academy experiment is the demonstration that this is not the same as identifying what happens when the rule is changed: an assignment engineered from a randomisation-based estimate produced −0.061 grade points for the students predicted to gain 0.053, because groups re-sorted internally.

Frontier “Big data solves it.” Non-identification does not shrink with sample size. A platform dataset of a hundred million users gives a narrow confidence interval around a confounded quantity. Angrist’s analysis adds a sharper version of the same warning: outcome-on-outcome peer regressions can be “vacuous, either unity or, if the average is defined as a leave-out mean, determined by a generic intraclass correlation coefficient” — arithmetic, not sociology.

Speculative “Contagion is the right frame for any correlated behaviour between connected people.” The randomised evidence fits some outcomes and not others: multivitamins spread and chlorine did not; knowledge spread and practice lagged; drinking transmitted between assigned roommates and prior academic ability did not. Transmissibility looks like a property of the behaviour rather than the network, and no theory predicts in advance which behaviours transmit.

Handwave “The same methods can be applied to socially and politically contested behaviours to show they spread.” This is the weakest claim in circulation and it fails on its own terms. The identification problem is worst exactly where such claims are made: self-reported outcomes, ties formed on the basis of the trait being studied, and no possibility of assigning peers at random. A method that finds contagious height in a general adolescent panel will find contagion in any trait on which friendships sort, and no randomised design exists to check it. This brief therefore confines itself to the health and behaviour outcomes that have actually been tested with assigned peers, and treats any extension beyond them as unsupported by the literature it cites.