Networks, Crowds, and Markets: Reasoning About a Highly Connected World

1. 1Separate the two engines before you act: ask whether aligning with others gives you information about quality (informational effect) or a direct payoff from coordinating (network/direct-benefit effect). The two demand different interventions.
2. 2For informational effects, model the decision as sequential: what did each earlier actor know privately, and what could a later actor infer from the visible action alone? Trace whether at some point people begin ignoring their own private signal in favor of the accumulated public pattern — that inflection is where a cascade starts.
3. 3Count exposure, not just presence. The Milgram sky-gazing result shows the response is a function of group size: 1 gazer stops almost no one, 5 stops more but most ignore, 15 stops 45%. Track the threshold where marginal onlookers flip.
4. 4Ask what information the visible action actually conveys. A crowded restaurant or heavy YouTube/Flickr usage transmits 'others rated this highly'; if the signal is contaminated (paid promotion, herding on no real information), the inference is hollow and the cascade is fragile.
5. 5Recognize cascades are informationally thin: they rest on inferences from limited action, so they can reverse fast when new private information surfaces or a better alternative appears in a part of the network with room to take hold (MySpace displaced by Facebook).
6. 6To seed adoption, target a subpopulation where the new option can accumulate visible early adopters, then let the 'rich get richer' feedback compound — attention concentrates on a few prominent items, leaving a long tail.
7. 7Where you cannot separate conformity-pressure from information, run the contrast test: vary whether prior choices are visible but uninformative versus visible and informative, and watch whether alignment persists without any informational cause.

Take the restaurant cascade. The first diners choose restaurant B, conveying to later arrivals that they knew something favorable. Each later diner watches the actions of earlier ones and infers the earlier people's private assessments from their choices. Past some point, people abandon their own private information in favor of the inference drawn from the accumulated pattern — an information cascade. The key insight is that this is not mindless conformity: it is rational inference from limited information, each person reasoning correctly given what they see. Yet because everyone is now copying rather than contributing new private signals, the crowd's choice can be based on very little actual evidence — the early diners' signals, amplified. That is why a cascade on B can be both stable-looking and wrong, and why fresh private information can tip it.

• If the visible action carries a quality signal (usage, crowding, recommendations), then treat alignment as rational inference and ask how much independent private information the crowd is actually pooling.
• If later actors have stopped adding their own private signals and are only echoing earlier actions, then flag a cascade — the aggregate may rest on very few real signals.
• If group size crosses the threshold where marginal onlookers flip (as in Milgram, 15 gazers → 45%), then expect self-reinforcing growth rather than linear response.
• If direct-benefit/network effects dominate, then the incumbent leader is hard to displace unless a challenger offers something markedly different or starts where there is room in the network to take hold.
• If you cannot tell conformity pressure from informational inference apart, then design a manipulation that removes the informational content and see whether alignment survives.

• Treating all imitation as a 'human tendency to conform' — the tell: you can explain the behavior but cannot answer why rational agents copy, missing the informational and direct-benefit mechanisms entirely.
• Reading a crowd as strong evidence when it is a cascade built on a few early signals — the tell: the choice reverses sharply when one credible piece of new private information appears.
• Assuming a dominant leader is permanent under network effects — the tell: you dismiss challengers until one offering something markedly different (Facebook vs MySpace) captures an underserved part of the network.
• Confusing social-pressure conformity with informational inference and prescribing the wrong fix — the tell: you try to counter 'peer pressure' when people are actually pooling information, or vice versa.
• Ignoring sequence and treating decisions as simultaneous — the tell: you model aggregate behavior as an independent vote count and miss that later actions depend in complex ways on inferences from earlier ones.

• ▸Decision under study: name the aligned choice being adopted
• ▸Mechanism check — informational: does the visible action convey quality info? (y/n + what)
• ▸Mechanism check — direct benefit: does aligning raise your payoff via network effects? (y/n + how)
• ▸Sequence map: order of visible prior choices and what each plausibly signals
• ▸Cascade point: where do actors stop using private info and start echoing?
• ▸Group-size threshold: at what count do marginal adopters flip? (cite observed)
• ▸Signal integrity: is the visible action a clean or contaminated quality signal?
• ▸Reversal risk: what new private info or alternative could tip this?
• ▸Intervention: seed target subpopulation or informational correction to apply

1. 1Model each individual as a node choosing between two behaviors (A and B), where payoff comes from coordinating with network neighbors — you earn a per-neighbor benefit for matching each neighbor's choice.
2. 2Assign the coordination payoffs: let matching a neighbor on A pay one amount and matching on B pay another. The relative size of these two payoffs sets the threshold — it is not a free parameter you pick, it is derived.
3. 3Compute the threshold q as the fraction of neighbors who must already choose A before A becomes the better response for the individual. Compare expected payoff from A (proportional to fraction of neighbors on A) against payoff from B (proportional to fraction on B).
4. 4For each node, count its current neighbors: if the fraction adopting the new behavior meets or exceeds q, flip that node to adopt; otherwise leave it.
5. 5Seed a small set of initial adopters, then iterate in rounds: recompute each remaining node's neighbor fraction, flip any that now cross threshold, and repeat until no node changes (the cascade stabilizes).
6. 6Check where the cascade stalls — it will halt at the boundary of a densely connected cluster whose members have too many internal same-behavior neighbors to ever cross q.
7. 7To make a superior behavior spread, seed it in a network region where it can progress incrementally, a few nodes at a time, rather than dropping it randomly.

Take Schelling's grid as the concrete threshold instance the passages give: agents of type X and O sit on a 150x150 grid, each with up to eight neighbors (including diagonals). The threshold t equals 3 — an agent is 'unsatisfied' whenever fewer than 3 of its neighbors share its type, which is exactly a fraction-of-neighbors adoption rule applied to staying put versus moving. In each round the simulator sweeps the grid, and every unsatisfied agent relocates to a cell where at least 3 same-type neighbors satisfy the threshold. In the small hand example with t=3, one round moved agents like X1, X2, and O5 and increased same-type clustering: before the round only one agent had no opposite-type neighbor, afterward six did. Run at scale on 10,000 agents per type, the process reached full satisfaction after roughly fifty rounds, producing large interlocking homogeneous regions. The lesson: a modest per-agent threshold, applied locally and iterated, produces large-scale segregation no individual intended.

• If the fraction of a node's neighbors adopting the new behavior is at least q, then flip it to adopt; if below q, then it stays.
• If the payoff for matching on the new behavior rises relative to the old, then q falls and the behavior spreads more easily; if it falls, q rises and spread stalls.
• If a cascade halts, then look for a densely connected cluster — its internal edges keep members below threshold and block diffusion.
• If you want a better behavior to displace an entrenched inferior one, then seed it in a sparse or peripheral region where incremental progress is possible, not uniformly at random.
• If agents care more about immediate neighbors than the whole population, then use the local neighbor fraction, not the global adoption rate, to evaluate the threshold.

• Treating the threshold as an arbitrary tuning knob rather than deriving it from the two coordination payoffs — tell: you cannot explain why q has the value it does.
• Using population-level adoption instead of local neighbor fraction — tell: your model predicts smooth global spread but real cascades stall at cluster boundaries.
• Assuming small advantages always cascade to dominance ('rich get richer') — tell: you can't explain why SixDegrees.com vanished while Facebook grew.
• Ignoring that thresholds produce unintended aggregate outcomes — tell: you're surprised that modest individual thresholds (like t=3) generate extreme segregation.
• Seeding a superior behavior randomly and concluding it 'can't win' — tell: you never tested whether a peripheral, incremental seed placement would let it progress a few nodes at a time.

• ▸Behavior A / Behavior B: name the two competing practices
• ▸Payoff for matching a neighbor on A: numeric coordination benefit
• ▸Payoff for matching a neighbor on B: numeric coordination benefit
• ▸Derived threshold q: fraction of neighbors on A needed to prefer A
• ▸Node neighbor count: number of network neighbors for the node under test
• ▸Current fraction adopting A among neighbors: count adopters / total neighbors
• ▸Flip decision: adopt if fraction >= q, else stay
• ▸Initial seed set: list of starting adopters and their network location
• ▸Cascade stopping point: cluster or boundary where flips cease

1. 1Fix the two primitives before anything else: p, the probability the disease/behavior passes across a single edge to a susceptible neighbor, and k, the number of new contacts an infected/adopting node makes while contagious. These are independent inputs — do not collapse them into a single 'spread' number.
2. 2Compute the basic reproductive number R0 = pk, the expected number of new cases produced by one infected individual. This single product decides everything downstream.
3. 3Compare R0 against the threshold of 1. If R0 < 1, each case produces fewer than one successor in expectation and the outbreak trends downward to extinction; if R0 > 1, it trends upward and persists with positive probability.
4. 4Model the spread as a branching process: wave 1 is the root's k contacts, each infected with probability p; each newly infected node then meets k fresh people in the next wave, passing on independently at probability p. Track infections wave by wave.
5. 5To raise or lower spreading power, decide which lever moves more cheaply — cut p (transmission per contact, e.g. behavioral/hygienic barriers) or cut k (number of contacts, e.g. isolation). Both reduce R0 multiplicatively, so halving either halves R0.
6. 6Push R0 below 1 to force die-out: the disease reaches a wave where it infects no one, and since future waves can only catch it from higher up the tree, it is gone permanently.
7. 7When R0 > 1, remember persistence is only probabilistic, not certain: as long as p < 1 the first few cases may 'get unlucky' and infect no one. Do not treat R0 > 1 as a guarantee of an outbreak, only as a possibility with positive probability.

Take the branching process in Figure 21.1 with k = 3. In the aggressive case (b) the contagion probability p is high: the root infects 2 people in the first wave, those produce 3 in the second, then 5 in the third, and presumably more beyond — the infection spreads widely because R0 = pk exceeds 1. In the mild case (c) p is smaller: of the two people infected in the first wave, one infects no one and the other infects just one person, who in turn passes it to nobody. The disease has completely vanished after the second wave, having infected only four people total. The difference is driven entirely by p, since k = 3 is held fixed — a smaller p pushes R0 = pk below 1 and the outbreak dies out. This is the same mechanism behind the e-mail recommendations for the Japanese graphic novel that spread radially outward from four initial purchasers: high enough per-edge passing probability lets adoption cascade.

• If R0 = pk < 1, then predict extinction after a finite number of waves with probability 1 — plan for containment, not suppression.
• If R0 = pk > 1, then predict persistence with positive probability — but still expect some outbreaks to die out by chance whenever p < 1.
• If you can only intervene on one primitive, then cut whichever of p or k is currently larger relative to its floor, since R0 responds equally to a proportional cut in either.
• If the target is social/behavioral contagion (adoption) rather than biological, then treat p as involving decision-making by the affected individual, not just chance contact — the same R0 math applies but the lever for lowering p is incentive/benefit, not a pathogen barrier.
• If the spread must cross the boundary of a densely connected 'closed community', then expect blocking regardless of R0, because high internal linkage makes that cluster resistant to outside influence.

• Treating a single 'spread rate' as sufficient and never separating p from k. The tell: you can't say whether a slowdown came from fewer contacts or lower transmission, so interventions are guesswork.
• Assuming R0 > 1 guarantees an epidemic. The tell: you're surprised when a contagious disease vanishes early — but with p < 1 the first few cases can always fail to infect anyone.
• Assuming R0 < 1 means zero cases. The tell: you dismiss an outbreak that briefly grows due to random fluctuation, even though it is trending downward and will die out.
• Confusing biological and social contagion mechanisms. The tell: you apply pathogen-style barriers to an adoption process (like the graphic-novel recommendations), where p is governed by individual decision-making and direct benefits instead.
• Ignoring network structure and assuming pure tree/branching dynamics everywhere. The tell: your forecast misses that short paths in a small-world network accelerate spread, or that a closed, high-clustering community blocks it entirely.

• ▸Contagion type: mark biological (chance contact) or social (decision-based adoption).
• ▸p (transmission probability per edge): probability one infected node passes to one susceptible neighbor.
• ▸k (contacts while contagious): number of new people each node meets per wave.
• ▸R0 = p x k: compute and record the expected new cases per case.
• ▸Threshold verdict: write DIE-OUT if R0 < 1, PERSISTS-with-positive-probability if R0 > 1.
• ▸Cheapest lever: note whether reducing p or reducing k is more feasible here.
• ▸Target R0 after intervention: pick a value below 1 and back-solve the required p or k.
• ▸Structural blockers: list any densely connected closed community likely to resist the spread.
• ▸Uncertainty note: record that persistence is probabilistic — small outbreaks may still die out when p < 1.

1. 1Separate the mechanism from the metaphor. When you observe aggregate growth (a rising popularity curve like YouTube's or Flickr's), name the specific feedback source: informational effect (people adopt because they see others adopt) or direct benefit (the product becomes more valuable as more people use it). These are the two engines behind 'the rich get richer.'
2. 2Locate the early-advantage variable that compounds. For popularity, it is current adoption share; the rate of new gains is a function of existing share, so small leads convert into faster subsequent gains.
3. 3Model the reaction, not the static snapshot. Never evaluate a change (to a product, site, or ranking method) assuming the rest of the system holds still. Ask: once I ship this, how do users and content-producers re-optimize, and does that re-optimization feed the leader further?
4. 4Check for a tipping point / instability. Determine whether the system has a threshold above which one option runs away and below which it collapses. If network effects are present, expect a leader that is hard to displace once it clears the threshold.
5. 5Test the prediction against the distributional signature. Rich-get-richer processes produce power-law popularity: a few dominant items and a long tail of obscure ones. If your data shows that shape, the feedback story is consistent; if it shows an even spread, it is not.
6. 6Look for what breaks the lock-in. Identify whether a new entrant offers something markedly different, or starts in a part of the network with room to grow — the two documented ways a dominant technology gets displaced despite network effects.
7. 7Distinguish outcomes from inevitability. Do not treat early lead as destiny: some sites flourish (Facebook), others with early presence vanish (SixDegrees.com). Reserve the rich-get-richer explanation for cases where the compounding mechanism is actually operating.

Take YouTube's rise as shown in the Google Trends search-volume curve across 2006–2008. The value of a video-sharing site to you depends on where the content and audience are; regardless of whether YouTube had better features, once it became the most popular site there was by definition added value in using it. That added value pulled in more uploaders and viewers, which increased the value again — a direct-benefit feedback loop where popularity builds on itself. The same pattern appears in the companion Flickr curve and in the MySpace adoption curve for 2005–2008, driven by both informational effects (seeing others adopt) and direct benefits (using what many already use). This dominance made the leader hard to displace, yet it was not permanent: a new technology could still take hold by offering something markedly different or by starting in an underserved part of the network. The lesson is that the curve's shape reflects a mechanism, not luck, but the mechanism does not guarantee any single site wins forever.

• If a product becomes more valuable as more people adopt it, then expect network effects, a tipping point, and a leader that is difficult to displace — plan around lock-in.
• If you are changing a ranking method or product feature, then design it assuming content-creators and users will re-optimize their behavior in response, not that the environment stays static.
• If observed popularity follows a power-law (few dominant items, long tail), then a rich-get-richer feedback process is a credible explanation; if the distribution is flat, then look elsewhere.
• If an incumbent dominates via network effects, then entry is only viable by offering something markedly different or by seeding a part of the network with room to grow.
• If two similar early entrants diverge sharply (Facebook thrives, SixDegrees vanishes), then do not assume the early lead caused it — identify whether the compounding mechanism was actually engaged.

• Treating early advantage as guaranteed success. The tell: you cite 'the rich get richer' but cannot name the specific feedback mechanism, and you ignore counterexamples like SixDegrees.com vanishing.
• Evaluating a change in isolation. The tell: your projection assumes 'everything else remains static' — the sign is a forecast that has no term for how the network reacts to your action.
• Confusing feature quality with popularity dynamics. The tell: you attribute YouTube's dominance to better features, when the added value came by definition from being most popular, independent of features.
• Missing the tipping point. The tell: you model steady linear growth and are surprised by runaway dominance or sudden collapse, having not checked for stability/instability thresholds.
• Assuming lock-in is permanent. The tell: you dismiss new entrants entirely, ignoring that a markedly different offering or an underserved network region can displace an incumbent.

• ▸Item/innovation observed: what product, site, or practice is rising or dominating.
• ▸Aggregate signal: the population-level evidence (e.g., search-volume curve, adoption share).
• ▸Feedback source: informational effect, direct benefit, or both — state which.
• ▸Compounding variable: what early quantity increases the rate of further gains.
• ▸Tipping point present?: yes/no and the threshold or condition if known.
• ▸Reaction forecast: how users/content-creators re-optimize once you act — no static assumptions.
• ▸Distribution check: does observed popularity fit a power-law / long-tail shape?
• ▸Displacement path: what markedly-different offer or underserved network region could unseat the leader.
• ▸Non-inevitability note: counterexample or reason the early lead might NOT compound.

1. 1Identify which side holds private information about quality. In the used-car market the seller knows whether a car is good or bad; the buyer does not. Name the informed party explicitly before modeling anything.
2. 2Assign values to each quality type for both sides. Use the passage's numbers: buyers value a good car at 12 and a bad car at 6; workers generate $80,000 (productive) or $40,000 (unproductive), with self-employment alternatives of $55,000 and $25,000.
3. 3Write the uninformed party's expected value as a probability-weighted average over quality types. For cars, buyer willingness to pay = 12h + 6(1−h) = 6 + 6h, where h is the expected fraction of good units actually offered.
4. 4Impose the self-fulfilling expectations condition: the buyers' assumed h must equal the h that actually results once sellers decide whether to offer at the induced price. Solve for which h values are consistent.
5. 5Check each quality type's participation constraint at the induced price. A good-car seller offers only if 6 + 6g ≥ 10, i.e. g ≥ 2/3; below that threshold good cars withdraw and the pool degrades.
6. 6Trace the unraveling: when high-quality holders exit, recompute h with only lower types remaining, recompute price, and repeat until you reach a stable pool (possibly one where only the worst quality trades or no trade occurs).
7. 7Distinguish adverse selection (hidden type, present before the deal) from moral hazard (hidden action after the deal, e.g. reduced effort to stay healthy once insured). Treat them as separate information asymmetries requiring separate fixes.
8. 8Design a remedy: signaling quality, or a reputation system / mechanism that lets the informed party credibly reveal type, so the market does not collapse to the worst pool.

Take the used-car market. Buyers value a good car at 12 and a bad car at 6, and a fraction g of all cars are good. If buyers can't tell types apart, they pay 6 + 6h where h is the expected good fraction among cars actually for sale. Test the candidate equilibrium h = g: buyers then pay p* = 6 + 6g, and a good-car owner sells only if 6 + 6g ≥ 10, i.e. g ≥ 2/3 (a bad-car owner will always sell at that price). So when g ≥ 2/3, all cars trade and h = g holds. But when g < 2/3, p* falls below what a good-car owner needs, good cars are withdrawn, h drops below g, the price falls further, and the market unravels toward only bad cars being offered. The same structure recurs in the labor market ($80k/$40k productive/unproductive with $55k/$25k self-employment options) and in health insurance, where the healthiest keep opting out until the pool can collapse entirely.

• If the informed party's quality types have participation thresholds above the pooled price, then expect adverse selection: the high types exit and only low types remain.
• If g ≥ 2/3 in the car model (good types willing to sell at the pooled price), then a full-participation equilibrium exists and asymmetry does not destroy the market.
• If withdrawing high types lowers h and the recomputed price, then iterate the unraveling until you reach a stable pool — don't stop at the first round.
• If the asymmetry concerns a fixed hidden type known before the deal, treat it as adverse selection; if it concerns a hidden action taken after the deal (effort, care), treat it as moral hazard.
• If unraveling is socially costly, then introduce a credible signal or reputation mechanism so the informed side can separate itself from lower types.

• Assuming the offered pool h equals the overall population fraction g. Tell: you plug g into the price without checking whether all sellers actually choose to offer at that price.
• Treating the buyer's willingness to pay as fixed rather than as an expectation that responds to who participates. Tell: your price doesn't move when high-quality units exit.
• Confusing adverse selection with moral hazard and applying the wrong fix. Tell: you try to price-discriminate on type when the real problem is unobservable post-contract behavior.
• Stopping the analysis after one round of exit. Tell: you conclude 'good cars leave' but never recompute h and the resulting further withdrawals.
• Ignoring the alternative-income / participation constraint of the informed party. Tell: you assume everyone trades even when the productive worker's $55,000 self-employment option beats the pooled wage.

• ▸Informed party: who holds the private quality information
• ▸Quality types: list each type and its share in the population
• ▸Uninformed-side value per type: what a good vs bad unit is worth to the buyer/firm
• ▸Informed-side reservation per type: minimum price each type accepts (e.g. self-employment income, seller's own value)
• ▸Pooled price formula: expected value = sum of (share × value) using assumed offered fraction
• ▸Participation check: for each type, does pooled price meet its reservation? (yes/no)
• ▸Self-fulfilling condition: does assumed offered fraction equal the fraction that actually trades?
• ▸Unraveling trace: recompute fraction and price after high types exit; repeat until stable
• ▸Remedy: signal, reputation system, or mechanism that separates types

1. 1Test the compensation condition first: ask whether the affected party receives a mutually agreed-upon payment for the effect. If the effect is uncompensated, it is an externality; if payment equals the cost imposed, it is not. The Diet Coke case shows this — consuming a can removes it from others but you pay production cost, so the effect is fully compensated and no externality remains.
2. 2Assign a sign by asking whether the third party's welfare rises or falls. Someone joining a social networking site raises your welfare with no payment to you — positive. Your driving onto a congested road lowers others' payoff with no payment from you — negative.
3. 3Separate whether payoffs depend on the number of others using the good (population-level, as in network effects) or on the specific network connectivity (structural). The passages treat these as distinct: Chapter 17 handles count-based effects, Chapter 19 handles connectivity-based diffusion.
4. 4For network-effect goods, model the value each consumer places on the good as increasing in the number of expected users, then look for self-consistent adoption levels where expectations match outcomes — this produces multiple equilibria and tipping points rather than a single price-quantity point.
5. 5Locate the tipping point: identify the adoption threshold above which network value pulls more users in (positive feedback, rich-get-richer) and below which the good collapses. Facebook flourished by crossing it; SixDegrees.com did not.
6. 6Design a mechanism to internalize the externality: attach a price, property right, or payment so the acting agent faces the full social cost or benefit. The Coase Theorem and property-rights framing (Chapter 24) are the route from uncompensated effect back to social optimality.
7. 7For negative population externalities, check for congestion-style over-use where each additional user degrades the shared resource — the El Farol Bar problem is the canonical case where private incentives crowd a fixed-capacity good.

Take the fax machine at introduction. Its value to any single buyer depends on how many others already own compatible machines, so an early buyer confers a positive externality on every later and existing owner by expanding the reachable network — with no compensation flowing back. Model each potential buyer's willingness to pay as rising in expected adoption z. At low expected adoption, few buy, confirming low value; at high expected adoption, many buy, confirming high value — two stable outcomes bracketing an unstable tipping point. If adoption is nudged above that threshold, positive feedback drives it toward the high-adoption equilibrium; below it, the market unravels toward zero. This is why the same technology can either become ubiquitous or vanish depending only on early expectations, and why compatibility (shared file formats, shared standards) is what makes the operating-system version of this effect operate.

• If the effect on a third party carries no mutually agreed compensation, then classify it as an externality; if payment equals the imposed cost, then it is not.
• If a good's per-user value rises with the number of other users, then treat it as a positive network externality and expect multiple equilibria plus a tipping point, not a unique equilibrium.
• If adoption sits above the tipping threshold, then expect rich-get-richer amplification toward dominance; if below, then expect collapse toward zero.
• If each additional user degrades a fixed-capacity shared resource, then treat it as a negative externality (congestion) and expect private use to exceed the social optimum.
• If you want the market equilibrium to reach social optimality, then introduce a price, property right, or payment that makes the actor internalize the full external effect.

• Calling every spillover an externality. Tell: you flagged an effect that is actually paid for at cost (the Diet Coke error) — the missing check is whether compensation already offsets it.
• Assuming a unique equilibrium for network goods. Tell: your model yields one price-quantity point and cannot explain why Facebook succeeded while SixDegrees.com failed under similar starts.
• Evaluating a change as if the rest of the network stays static. Tell: your prediction ignores that the world reacts, so an apparently good product or policy tweak shifts population behavior in unintended directions.
• Confusing count-based effects with connectivity-based ones. Tell: you use population-level network-effect reasoning where outcomes actually hinge on who is linked to whom (structural diffusion, Chapter 19).
• Ignoring the sign flip on shared capacity. Tell: you treat 'more users' as always good and miss congestion, where added users impose negative externalities as in El Farol.

• ▸Action & actor: name the decision and who takes it.
• ▸Affected parties: who else's welfare changes.
• ▸Compensation check: is the effect paid for at cost? (yes = not an externality)
• ▸Sign: does third-party welfare rise (positive) or fall (negative)?
• ▸Dependence type: on number of users (population) or on connectivity (structural)?
• ▸Feedback & tipping point: is there a threshold above which use self-reinforces?
• ▸Current position: above or below the tipping point today?
• ▸Internalization mechanism: price, property right, or payment that makes the actor face the full effect.

1. 1Pick your resolution first: model the population as an amorphous aggregate (information cascades, network effects, rich-get-richer) or as a fine-grained graph where each person responds to specific neighbors. The two answer different questions about the same outcome.
2. 2At the aggregate level, decide which mechanism drives imitation: informational effects (others' choices reveal private information, as in Bayesian herding) or direct-benefit effects (payoff from aligning with others, as in network goods). This governs whether the outcome is fragile or self-reinforcing.
3. 3For the network-structural model, seed the behavior with a small set of initial adopters and let it spread radially outward, node by node, where each node adopts only when enough of its immediate neighbors have adopted.
4. 4Identify densely connected clusters — closed communities with high internal linkage. These form the boundaries that block diffusion, so map them to predict where a cascade stalls short of everyone.
5. 5Check for a tipping point: in markets with network effects, plot expected adoption against actual adoption and find where the leader becomes hard to displace versus where a new entrant can still take hold in a slack region of the network.
6. 6Judge the outcome on the full range: does the behavior die out quickly, stall at a cluster boundary, or become a complete cascade reaching everyone. State which, and name the structural reason.
7. 7To determine whether a superior-but-new behavior can win, look for a portion of the network where it can make incremental progress a few people at a time rather than requiring simultaneous mass switching.

Ryan and Gross's study of hybrid seed corn adoption among Iowa farmers is the classic informational-effects case. Farmers observed the decisions of their network neighbors, and those observations gave them indirect information that led them to try the innovation themselves. Because adoption rode on inferred information rather than direct payoff, the outcome built sequentially as each farmer's choice became evidence for the next. Coleman, Katz, and Menzel's study of tetracycline adoption among physicians follows the same informational logic. In the media-sharing analog, once YouTube became the most popular video-sharing site there was by definition added value in using it — a direct-benefit case — so the leader became difficult to displace, yet not permanently, since a markedly different entrant starting in a receptive part of the network could still take hold.

• If imitation comes from inferring others' private information, then model it as an information cascade — and expect the outcome to be fragile, since a large crowd may be acting on surprisingly little genuine information.
• If imitation comes from direct payoff to alignment (compatibility, audience size), then model it as network effects — and expect the leader to be entrenched with a tipping point protecting it.
• If the population is homogeneous and everyone is aware of everyone's prior choices, then use the aggregate resolution; if outcomes depend on who is adjacent to whom, then move to the structural graph model.
• If a superior new behavior starts inside a densely connected cluster of the incumbent, then expect it to be blocked; if it starts where it can advance incrementally a few nodes at a time, then it can displace the incumbent.
• If adoption is spreading radially from initial seeds and hits a closed community's boundary, then predict the cascade stalls there rather than reaching everyone.

• Reading crowd size as strong quality evidence: seeing a site heavily used or a restaurant packed suggests knowledge, but in a cascade the actions of many rest on little real information. The tell is that everyone cites 'everyone else' and no one cites independent private evidence.
• Treating a leader's dominance as permanent. Network-effect leadership is difficult to displace but not fixed; the tell is a new, markedly different entrant gaining a foothold in a slack region of the network.
• Confusing informational and direct-benefit mechanisms, which predicts the wrong fragility. The tell: you assumed the outcome would reverse with new information, but adopters keep the behavior because of compatibility payoffs.
• Modeling at the wrong resolution — using a homogeneous-population model for a phenomenon driven by specific neighbor structure. The tell is that your aggregate model can't explain why the behavior stalled at a cluster boundary.
• Equating social contagion with biological contagion mechanically. Social spread involves the affected person's decision making; disease spread is chance of pathogen transfer through contact. The tell is assuming exposure guarantees adoption, ignoring thresholds and choice.

• ▸Behavior/innovation: name what is spreading
• ▸Resolution: aggregate population or fine-grained network
• ▸Imitation mechanism: informational effect or direct-benefit effect
• ▸Initial adopters: who and where in the network they seed
• ▸Cluster boundaries: densely-connected closed communities likely to block spread
• ▸Tipping point present: yes/no and where the market flips
• ▸Incremental foothold: is there a network region where a new entrant advances a few at a time
• ▸Predicted outcome: dies out / stalls at boundary / complete cascade, with the structural reason

1. 1Model the market as a bipartite graph: put buyers/agents on one side, sellers/goods on the other, and draw an edge only where an agent finds a good acceptable. For sponsored search, slots are the goods and advertisers are the buyers.
2. 2Attach valuations: write down v_ij, the value agent i places on good j. In the ad case compute this as r_i·v_j — the clickthrough rate of slot i times the revenue per click of advertiser j — so higher slots (more clicks) matter more to high-value advertisers.
3. 3Find the optimal assignment: choose the perfect matching that maximizes total valuation across all agent-good pairs. This is the socially optimal allocation before any prices are set.
4. 4Construct market-clearing prices: assign a price to each good such that every agent's preferred good (the one maximizing value minus price) collectively forms a perfect matching, so no good is over- or under-demanded. Use the auction-style ascending procedure from Chapter 10.4 to build them.
5. 5Verify the market-clearing property: each agent gets a good in their preferred set at current prices, and every good is claimed exactly once. If some good is demanded by two agents and another by none, raise the price of over-demanded goods and repeat.
6. 6Choose the mechanism that induces truthful bidding: run a second-price/VCG mechanism so that bidding true values is a dominant strategy. In the single-item case this is the ascending-bid auction; for multiple slots use VCG, which charges each winner the externality it imposes on others.
7. 7Confirm optimality and division of surplus: the resulting allocation maximizes social welfare; then trace who captures the surplus — an intermediary or agent earns positive profit only if it is 'essential,' i.e., no other party can replicate its function.

Take the dormitory case: five rooms (Room1–Room5) and five students (Vikram, Wendy, Xin, Yoram, Zoe), each student listing acceptable rooms as edges in a bipartite graph. With no prices, you seek a perfect matching so every student gets exactly one acceptable room and every room is filled. Now add valuations — some students value larger or sunnier rooms more — and search instead for the assignment maximizing total valuation. To decentralize this, attach prices to rooms so that each student's most-preferred (value minus price) room forms a perfect matching: if two students both prefer Room1 at current prices, you raise Room1's price until one switches to their next-best. When every student is content with a distinct room, prices are market-clearing and the allocation is socially optimal. This mirrors sponsored search, where rooms become ad slots valued at r_i·v_j and the same clearing logic sets prices per slot.

• If a good is demanded by more than one agent while another sits unclaimed, then raise the price of the over-demanded good and re-solve until demand and supply match.
• If you want bidders to reveal true values, then use a second-price/VCG mechanism (ascending-bid for one item) so truth-telling is a dominant strategy.
• If an intermediary or trader can be fully replicated by another, then it captures zero profit — only 'essential' parties earn positive payoff.
• If the setting aggregates beliefs about a future event with a right/wrong resolution, then use a market (prices as consensus probabilities); if it aggregates subjective preferences with no correct answer, use voting instead.
• If a valuation depends on both slot and advertiser, then set benefit = r_i·v_j and match to maximize the sum, since this product is what the optimal assignment ranks on.

• Confusing the optimal assignment with any perfect matching — the tell is a filled graph that leaves total valuation lower than an alternative matching would.
• Setting prices that leave a good over-demanded and another unclaimed and calling it cleared — the tell is two agents naming the same good as their top preferred choice.
• Using a first-price or non-VCG format then assuming bids equal true values — the tell is bidders shading below their valuation to avoid overpaying.
• Assuming every intermediary in a trading network profits — the tell is a richly connected network where a trader's function is duplicated yet you still credit it with a payoff.
• Reading market prices as guaranteed-correct forecasts — the tell is treating a prediction-market price as certainty rather than an aggregate probability that can swing on events like a convention or primary result.

• ▸Agents: list every buyer/bidder on one side
• ▸Goods/slots: list every item or slot on the other side
• ▸Edges: mark each agent-good pair the agent finds acceptable
• ▸Valuation v_ij: enter each agent's value for each good (use r_i·v_j for ad slots)
• ▸Current prices: assign a trial price to each good
• ▸Preferred set per agent: circle the good(s) maximizing value minus price
• ▸Clearing check: is every agent's preferred good part of one perfect matching? yes/no
• ▸Adjustment: raise price of each over-demanded good, re-run until cleared
• ▸Mechanism + surplus: name the auction (VCG/second-price) and note who is 'essential' and earns profit

1. 1Draw the exchange network as a graph: nodes are participants, edges are the specific one-to-one deals each pair could strike. Do not add edges that don't reflect an actual possible exchange; power comes from who can trade with whom, not from raw connection count.
2. 2For each node, list its alternatives — the other partners it could turn to if a given negotiation fails. A node with many acceptable alternatives holds an outside option that raises its bargaining power over any single partner.
3. 3Identify each partner's dependence: a node is powerful over a neighbor when that neighbor has few or no alternatives, so power is relative to the partners you are connected to, not just to your own degree.
4. 4Apply the stable-outcome test: check whether any node could break its current deal and strike a better one with an unmatched or under-paid partner. If such a defection exists, the proposed division of value is not stable.
5. 5Refine stable outcomes toward balanced outcomes: within the set of stable divisions, use the Nash bargaining split between each connected pair so each side's share reflects the outside options both bring to that specific edge.
6. 6Trace ripple effects: after fixing one node's share, recompute the outside options of its neighbors, since one participant capturing more value changes what everyone connected to them can credibly demand.
7. 7Map the resulting shares back to structure to see who sits in the powerful position — typically a node whose partners are each individually dependent on it while it retains multiple substitutable partners.

Consider a buyer–seller or exchange network of the kind studied in the network-exchange experiments of Chapter 12. Place node B connected to two peripheral nodes A and C, where A and C are connected to no one else. A and C each have exactly one possible partner — B — while B can choose between them. In each round only one deal per partner can complete, so B always has an alternative and A and C have none. The stable-outcome test confirms B can defect to whichever of A or C offers more, so any split giving A or C a large share collapses. The balanced-outcome refinement, using the Nash bargaining solution on each active edge, then divides value so B captures the lion's share, matching the experimental result that centrally positioned participants extract more from partners who lack outside options.

• If a node's partners each have no other acceptable partner while the node itself has several, then that node holds strong power and captures most of the exchanged value.
• If a proposed division lets any node defect to form a better deal with someone currently unmatched or underpaid, then the division is not stable — discard it.
• If multiple divisions are stable, then select among them with the Nash bargaining split on each edge so shares track the relative outside options of the two endpoints.
• If two adjacent powerful nodes are connected to each other, then neither dominates and their split approaches an even division because each has a comparable alternative.
• If adding or removing a single edge changes a neighbor's outside options, then recompute all downstream shares because power imbalances ripple through connected nodes.

• Equating degree with power: a high-degree node whose every neighbor also has many alternatives holds little leverage. The tell is that predicted shares don't match observed outcomes because you counted connections instead of dependence.
• Ignoring outside options and splitting every deal 50/50: this misses that a partner with no alternative accepts far less. The tell is that peripheral, dependent nodes are predicted to earn as much as central ones.
• Stopping at stable outcomes without balancing: stability alone permits many divisions, so you report an arbitrary one. The tell is that your answer is unstable to the Nash bargaining critique on individual edges.
• Modeling the network as static after one deal: failing to trace ripple effects means neighbor shares stay wrong. The tell is that recomputing outside options after a single settlement flips who has power.
• Adding edges that don't reflect real possible exchanges — from geography, regulation, or actual trading access — inflates some nodes' apparent alternatives. The tell is predicted power for a node whose 'connections' cannot actually transact.

• ▸Node ID: label each participant in the exchange network
• ▸Edges (possible deals): list each partner this node can actually trade with
• ▸Alternatives count: number of acceptable partners if the primary deal fails
• ▸Partner dependence: for each neighbor, note how many alternatives THEY have
• ▸Stable-outcome check: mark yes/no whether any node can profitably defect
• ▸Nash bargaining split: record the balanced share for each active edge
• ▸Ripple recheck: note which neighbors' outside options change after this deal settles
• ▸Power ranking: order nodes from most to least value captured
• ▸Structural note: describe the position (central/peripheral) driving the imbalance

1. 1Represent the system as a graph: fix what counts as a node (person, page, firm, organism) and what counts as a link (e-mail exchange, borrower-lender relation, Web hyperlink). Follow the book's own examples — nodes are 436 HP employees, links are the pairs connected by e-mail communication.
2. 2Decide directionality. For symmetric ties (friendship, e-mail exchange) use an undirected graph; for asymmetric ties (a link from one blog to another, a Web hyperlink) use a directed graph, since the direction changes which structural positions exist.
3. 3Run breadth-first search from nodes to compute paths, distances, and connectivity. Identify the connected components — the maximal sets where every node can reach every other — and note isolated tendrils versus the giant component.
4. 4Locate the dense core versus the periphery. In the interbank loan network the annotation reveals a dense core (GSCC/GIN/GOUT/tendrils in a bow-tie); label which nodes sit in the strongly connected core and which only feed in or out.
5. 5Find bridges and structural holes: nodes or edges whose removal disconnects clusters, or that span otherwise-separated groups. In the political-blogs network the two well-separated clusters are joined by comparatively few cross-links — those are the spanning positions.
6. 6Distinguish tie strength within the structure. Overlay strong versus weak ties; weak ties are the ones most likely to bridge clusters (they carry novel information), while strong ties concentrate inside densely knit groups (triadic closure).
7. 7Read the structure against a substrate you can compare it to — here the HP e-mail network is superimposed on the official org hierarchy — so you can see where actual interaction departs from formal structure.
8. 8State what the structure constrains: who can reach, influence, or trade with whom. Use the component/path structure to bound diffusion, and the core-periphery/bridge structure to identify power and bottleneck positions.

Take Figure 1.2: the pattern of e-mail communication among 436 employees of the Hewlett-Packard Research Lab, superimposed on the official organizational hierarchy. Nodes are employees; an undirected link joins two people who exchanged e-mail. Visually you can already see the network is hard to summarize succinctly: some parts are more densely interconnected than others, with certain central figures. Comparing the interaction graph to the formal hierarchy exposes where real communication crosses reporting lines — cross-department ties that the org chart does not predict. Those cross-cutting links are candidate bridges: they connect otherwise separate clusters and let information move between divisions. Contrast this with Figure 1.4, the 2004 political-blogs network, whose links reveal two natural, well-separated clusters with few connections between them. The same reading tools — density, clustering, components, spanning links — apply to both, even though one is e-mail and the other is Web hyperlinks.

• If the tie is asymmetric (hyperlink, loan direction, citation), then model a directed graph and analyze in/out roles (as in the Web's bow-tie GIN/GOUT); if symmetric (e-mail exchange, friendship), then use an undirected graph.
• If two densely knit clusters are joined by only a few links, then treat those links as bridges/structural-hole spanners and expect them to carry the novel information and confer power.
• If a node or edge, when removed, splits a component, then it is a bottleneck; prioritize it for questions about control, robustness, or contagion pathways.
• If the observed interaction network diverges from a known formal substrate (org hierarchy), then investigate the departures — that gap is where informal influence lives.
• If the network has a dense core and a sparse periphery, then analyze core participants for systemic roles (as with interbank loans and financial-system health) rather than treating all nodes as equivalent.

• Trying to summarize the whole network in one number or one glance. The tell: you say 'it's dense' when the passages warn some parts are more or less densely interconnected — density is local, not global.
• Ignoring link direction. The tell: you treat a Web hyperlink graph as symmetric and miss the bow-tie IN/OUT asymmetry that determines who can reach whom.
• Assuming the visible network is the whole story. The tell: you reason only about observed ties, forgetting the passages' warning that effects come from implicit networks we do not directly observe.
• Confusing strong ties with important ties. The tell: you focus on the dense cluster and overlook that the cluster-spanning weak ties are the ones carrying new information.
• Reading structure as static. The tell: you fix positions once, ignoring that changes create incentives that shift behavior across the network in unintended ways.

• ▸Node definition: state exactly what one node represents.
• ▸Link definition: state exactly what a link means and its source (e-mail, loan, hyperlink).
• ▸Direction: undirected or directed — justify from tie symmetry.
• ▸Components: list connected components and the giant component.
• ▸Core vs periphery: name nodes in the dense core and those only in tendrils.
• ▸Bridges / structural holes: list links or nodes that span otherwise-separated clusters.
• ▸Tie strength: mark which ties are strong (intra-cluster) vs weak (cross-cluster).
• ▸Substrate comparison: note the formal structure you compare against and where interaction diverges.
• ▸Constraint statement: one line on who can reach/influence/trade with whom given this structure.

1. 1Classify each of a node's edges by two signals the passages treat as jointly diagnostic: interpersonal strength (close friend vs. acquaintance) and structural role (embedded in a cluster vs. reaching a different network region). Both matter; neither alone is sufficient.
2. 2For any edge A-B, test for a shared friend: check whether A and B have any common neighbor. If none exists, the A-B distance would exceed 2 if you deleted the edge — that edge is a local bridge.
3. 3Compute the span of each local bridge by deleting it and measuring the shortest alternate path between its endpoints. In Figure 3.4 the A-B edge has span 4 because removal forces the path through F, G, H.
4. 4Distinguish a true bridge from a local bridge: delete the edge and ask whether the endpoints fall into separate components (true bridge) or merely move farther apart (local bridge). Expect true bridges to be essentially absent in real social networks — assume a longer, hard-to-discover multistep path exists.
5. 5Cross-reference structural role against tie strength: local bridges tend to be weak ties (acquaintances), because triadic closure would have already generated a shared friend around any long-standing strong tie. An edge is a local bridge precisely when it forms no triangle.
6. 6Trace information flow: treat tightly-knit clusters (A, C, D, E who all know each other) as sources of redundant, similar information, and treat local-bridge edges as the only channels carrying novel information (like scarce job leads) from other network regions.
7. 7Use the clustering coefficient of a node's neighborhood as your summary metric: high clustering signals embeddedness in strong-tie clusters; low clustering signals the node holds bridging positions across regions.

Granovetter interviewed people who had recently changed employers and found many learned about their new jobs through personal contacts — but described those contacts as acquaintances rather than close friends. The puzzle: your close friends have the most motivation to help, so why do distant acquaintances supply the decisive job information? The answer joins structure and strength. Close friends sit inside your tightly-knit cluster (like A, C, D, E in Figure 3.3), so they are exposed to the same opinions and the same information you already have. An acquaintance like B typically sits at the far end of a local bridge, reaching into a different part of the network. Because information about good jobs is scarce, only the bridging edge carries opportunities you would not otherwise hear about. The weak tie wins precisely because of where it sits, not because of how warm it is.

• If A and B share no common friend, then their edge is a local bridge and is carrying non-redundant information — treat it as high-value for novel leads.
• If deleting an edge splits its endpoints into two components, then it is a true bridge; but in a real large network assume this is vanishingly rare and look for the hidden multistep alternate path.
• If an edge forms the side of any triangle, then it is not a local bridge — its information channel is redundant with the cluster around it.
• If a node's neighborhood has low clustering, then it likely occupies bridging positions; in the Bearman–Moody data, low clustering among teenage girls correlated with significantly higher likelihood of contemplating suicide, so read low clustering as both bridging power and social exposure.
• If you need novel information, then route through weak-tie local bridges; if you need reinforcement, trust, or coordinated action, then rely on the embedded strong-tie cluster.

• Equating strong ties with useful ties: assuming your closest friends are your best information source, when their embeddedness makes their information redundant with yours. Tell: everyone in the cluster keeps surfacing the same leads.
• Confusing a local bridge with a true bridge: declaring an acquaintance edge the 'only link' between two worlds. Tell: a longer path through other nodes (like F, G, H) actually exists once you view the ambient network.
• Calling an edge a local bridge without checking for common neighbors. Tell: the endpoints turn out to be at distance 2 after deletion, meaning a shared friend closes the triangle.
• Reading tie strength purely interpersonally and ignoring structure, or purely structurally and ignoring strength — the whole point is the coupling of the two perspectives. Tell: your analysis explains warmth but not information advantage, or vice versa.
• Treating span as binary. Tell: you label edges 'bridge / not bridge' and lose the graded reach that span (e.g., 4 in Figure 3.4) actually measures.

• ▸Edge (A-B): name the two endpoints under review.
• ▸Interpersonal strength: mark strong (close, frequent) or weak (casual acquaintance).
• ▸Common neighbors: list any shared friends; none = candidate local bridge.
• ▸Triangle test: does this edge form a side of any triangle? yes/no.
• ▸Bridge type: delete it — separate components (true bridge) or larger distance (local bridge)?
• ▸Span: shortest alternate-path length between endpoints after deletion.
• ▸Information role: redundant (inside cluster) or novel (spanning regions)?
• ▸Neighborhood clustering: high (embedded) or low (bridging).
• ▸Action: what to route through this edge — novel leads vs. trust/coordination.

1. 1Define the strategy set for each agent explicitly: the driver's strategies are the available routes; the bidder's strategies are the possible bids. If you cannot list what each agent can choose, you have not yet specified the game.
2. 2Write the payoff for each agent as a function of ALL strategies, not just their own. For a bidder, payoff = value of goods received minus price paid; for a driver, payoff is based on resulting travel time given everyone else's route choices.
3. 3Choose the mechanism (auction format, pricing rule, traffic routing, market prices) as your design lever, and treat it as the thing you set before agents choose.
4. 4Predict behavior by finding the equilibrium — a self-reinforcing state where no individual has an incentive to unilaterally change strategy, even knowing how others will behave.
5. 5Re-evaluate any rule change against the equilibrium, never against a static world: assume the network reacts. Ask what new incentives your change creates before assuming it delivers the gain you designed it for.
6. 6Trace second-order and population-level effects: adding a resource or tightening a rule can shift behavior across the whole network, sometimes offsetting the intended gain.
7. 7Stress-test the chosen mechanism against counterintuitive reversals (Braess's Paradox, the winner's curse, offsetting bidder behavior) before committing.

A seller wants more revenue from a single-item auction, so he introduces more aggressive pricing rules — the design lever he controls. Evaluated statically, this looks like a good idea: extract more from whoever wins. But each bidder's success — whether she wins and what she pays — depends on how everyone else bids, so aggressive rules make the bidders' strategic problem more complex. In response, bidders adjust to an optimal strategy that takes the new rules into account, and this offsets whatever gains the seller expected. The lesson mirrors Braess's Paradox in traffic: changing the rules of a shared system creates new incentives that shift behavior, and the equilibrium outcome can neutralize or reverse the designer's intent. The seller should have predicted bidder behavior at the new equilibrium rather than assuming existing bids would remain fixed.

• If your evaluation assumes 'everything else stays static,' then it is wrong — re-run it assuming the world reacts to what you do.
• If a bidder's or driver's payoff changes depending on others' choices, then treat their optimal action as a best response, not a fixed number.
• If you tighten pricing or add a resource, then check for offsetting behavior or Braess-style reversal before booking the expected gain.
• If no agent has an incentive to unilaterally deviate, then you have found the equilibrium and can use it to predict outcomes.
• If the effect only appears at the population level, then design for aggregate response, not for the isolated individual.

• Evaluating a rule change in isolation — the tell is a projection built on 'holding all else equal' that ignores how incentives shift behavior across the network.
• Assuming aggressive pricing mechanically raises revenue — the tell is surprise when bidders' adjusted strategies erase the gain.
• Adding capacity to relieve congestion and getting worse outcomes — the tell is Braess's Paradox: efficiency drops after resources are added.
• Treating a strategy set as one-sided — the tell is a payoff written only in terms of the agent's own choice, with others' choices missing.
• Ignoring implicit, unobserved networks — the tell is complex population-level responses appearing that your model, which only counted visible links, cannot explain.

• ▸Design lever: the rule/format/price you control before agents choose
• ▸Agents: list every decision-maker whose choice affects outcomes
• ▸Strategy set per agent: the concrete options each can commit to
• ▸Payoff function per agent: outcome value as a function of all strategies
• ▸Best response check: given others' choices, what does each agent optimally do
• ▸Equilibrium: the state where no agent gains by unilaterally deviating
• ▸Reaction test: how does the network shift when you change the lever
• ▸Counterintuitive-reversal flag: could Braess/winner's-curse/offsetting bidding undo the intended gain

1. 1Fix the characteristic you're testing (race, age, obesity status, articles edited) before you look at the drawn network — homophily is defined relative to a specific attribute, not a general vibe.
2. 2Compare the observed network to a null model: randomly reassign the characteristic across the same nodes and edges, then ask whether real neighbors share the attribute more than they would by chance. If they do, you have evidence of homophily on that attribute.
3. 3Distinguish immutable attributes (race, ethnicity, age) from mutable ones (residence, occupation, affluence, interests, beliefs, opinions), because mutable attributes can be caused by influence, immutable ones cannot.
4. 4Separate the two generative mechanisms: selection (people choose similar friends, so similarity precedes the link) versus social influence (people become similar to existing friends, so similarity follows the link).
5. 5Get longitudinal data if you want to attribute causation — you cannot separate selection from influence with a single snapshot, since both produce the same static clustering.
6. 6Rule out confounding homophily: check whether the attribute you observe (obesity) is just tracking another correlated attribute (shared foci, location) that actually drives the clustering.
7. 7Operationalize behavioral similarity as a ratio when nodes have action sets — e.g., (articles edited by both A and B) / (articles edited by at least one), then track how this ratio moves relative to the timing of link formation.
8. 8Treat any homophily finding as a starting point, not a conclusion: an observed similarity among neighbors demands the follow-up question of why.

Christakis and Fowler tracked obesity status and social network structure across roughly 12,000 people over 32 years. Using the numerical homophily measure, they found obese and nonobese people clustered more than in a version of the network where obesity status was assigned randomly. The hard part was then explaining the clustering: (i) selection, where people befriend others of similar obesity status; (ii) confounding homophily, where the network reflects similarity on other correlated dimensions; or (iii) influence, where changes in friends' obesity status shift your own future obesity. Their statistical analysis argued that even after accounting for (i) and (ii), there remained significant evidence for (iii). The interpretation: obesity behaves like a socially 'contagious' condition — not caught like the flu, but spread through the network by influence. The lesson is that seeing people similar to their neighbors did not, by itself, tell them why.

• If you only have a single snapshot, then report homophily as an observation and refuse to claim selection or influence — the mechanisms are indistinguishable statically.
• If the characteristic is immutable (race, age), then influence is excluded and any clustering must come from selection or confounding, not from friends changing each other.
• If two clustered nodes also share a focus or context (same school, karate club, edited articles), then attribute part of the link to focal/contextual similarity before crediting intrinsic ties.
• If observed neighbor-similarity does not exceed the randomized-attribute baseline, then do not claim homophily — the division may be an artifact of how the network is drawn.
• If you want to argue influence (type iii), then you must first statistically account for selection (i) and confounding homophily (ii), or the claim is unsupported.

• Reading homophily off a diagram: treating visible color/cluster divisions as genuine structure without the randomized null-model test. Tell — no baseline comparison is ever stated.
• Collapsing selection and influence into one claim from cross-sectional data. Tell — a causal 'friends made him obese' assertion backed only by a static clustering measure.
• Ignoring confounding homophily: crediting the observed attribute when a correlated, possibly unobserved attribute (shared foci) is the real driver. Tell — no check for other correlated dimensions.
• Attributing a single link to one mechanism: insisting triadic closure OR homophily OR focal closure caused a tie, when the passages stress most links arise from a combination. Tell — clean single-cause stories about individual edges.
• Testing homophily without fixing a characteristic first: asking 'is this network homophilous?' in the abstract. Tell — no named attribute (race, age, obesity, edit overlap) anchoring the test.

• ▸Characteristic under test: name the single attribute (e.g., race, age, obesity status, articles edited)
• ▸Mutable or immutable: mark it, since immutable rules out influence
• ▸Similarity measure: state the exact numeric definition (e.g., shared-edits / total-edits ratio)
• ▸Null-model result: observed neighbor-similarity vs. randomly-reassigned baseline
• ▸Selection evidence: does similarity precede link formation in the timeline?
• ▸Influence evidence: does a friend's later change in attribute predict your later change?
• ▸Confounders checked: list correlated attributes/foci that could explain the clustering
• ▸Data type: snapshot or longitudinal (longitudinal required to separate i/ii/iii)
• ▸Open question: the deeper 'why' this observation opens up

1. 1Separate each agent's information into two components before modeling: the private signal they receive (their own observation) and the public information they infer from watching others' actions. In the herding experiment, the private signal is a draw favoring one urn; the public information is the sequence of guesses already announced.
2. 2Assign explicit prior probabilities to each state of the world. In the Bayes' rule model, start with a prior over which is the true state (e.g., which urn is being used), typically uniform when there is no reason to favor one.
3. 3Convert each observed action into a likelihood, not a certainty. When agent B sees A guess 'majority-blue,' she treats it as evidence weighted by how likely A's guess is under each state, then multiplies it into her prior.
4. 4Apply Bayes' rule to combine private signal and observed actions into a posterior belief: posterior is proportional to prior times likelihood of the full evidence set. The agent then acts on whichever state has higher posterior probability.
5. 5Check whether public information has overwhelmed the private signal. Count the net direction of prior observed actions; if the accumulated public evidence outweighs any single private draw, the agent rationally ignores her own signal and follows the crowd.
6. 6Identify the cascade point: the first agent for whom acting on the crowd is optimal regardless of her private signal. From that agent onward, actions convey no new information, so subsequent beliefs freeze on the public inference.
7. 7When modeling markets rather than sequences, read the aggregate price as the population's synthesized posterior. Treat the market price of an event-contingent asset as the crowd's probability estimate for that event, and track how it updates as public events arrive.

In the herding experiment, an urn is chosen and each person in turn draws a private signal and guesses which urn it is, seeing all prior guesses but not the prior signals. Suppose the first two people both draw and guess 'majority-blue'; a third person who privately draws 'red' now uses Bayes' rule to combine one red private signal against two blue-implied public signals. Because the two announced guesses reveal two blue signals, her posterior favors blue even after folding in her own red draw, so she rationally announces 'blue.' From this point her guess carries no information about her private signal, so the fourth person sees three 'blue' guesses but can only infer the first two were informative. Everyone thereafter is in the same position and also guesses blue, producing a cascade that can settle on the wrong urn. The lesson from cascades is that individually rational Bayesian updating on others' actions can lock a whole population onto a choice built on very little private information.

• If accumulated public information (net observed actions) strictly outweighs your private signal, then discard your signal and follow the crowd — this is optimal, not irrational.
• If your private signal and the public inference are balanced, then act on your own signal, because your action still transmits new information to those who follow.
• If you are the agent at whom a cascade begins, then treat all later observed actions as uninformative — do not update on them, since they no longer reveal private signals.
• If you are reading a market price rather than a sequence, then interpret the price as the aggregate posterior probability and update your own belief toward it only to the extent you think it reflects information you lack.
• If observed actions come from agents subject to contradictory norms or private incentives (as with node B facing conflicting group expectations), then discount the informational content of their choices.

• Treating others' actions as independent evidence when they are actually copies of earlier actions. The tell: your posterior swings sharply with each additional agent even though later agents added no private signal — you are double-counting a cascade.
• Assuming a cascade means the crowd is correct. The tell: high consensus with very thin underlying private information; the herd converged after only one or two informative signals and can be uniformly wrong about the urn.
• Overweighting your own private signal past the cascade point. The tell: you keep acting against a strong public inference on the strength of a single draw, ignoring that the accumulated evidence already dominates.
• Reading a market price as certainty rather than as an aggregate belief. The tell: you treat a 70-cent asset as 'the market says yes' rather than as a 70% probability that reacts to events like the primary's end or a convention.
• Ignoring that reported actions can be strategically shaded by incentives or competing norms. The tell: you model B's choices as honest information transmission while she is actually navigating contradictory expectations from different groups.

• ▸State space: list the possible states of the world (e.g., the candidate urns or outcomes).
• ▸Prior probability: assign a starting probability to each state before any observation.
• ▸Private signal: record your own observation and which state it favors.
• ▸Observed actions: list, in order, the actions of others you can see.
• ▸Inferred signals: for each observed action, note what private signal it implies (blank if the actor was in a cascade).
• ▸Likelihood weights: state how strongly one signal shifts belief between states.
• ▸Posterior: multiply prior by combined likelihoods and normalize to get updated probabilities.
• ▸Cascade check: mark the first point where public inference outweighs any single private signal.
• ▸Action decision: choose the higher-posterior state and note whether your action still transmits information.