The famous “six degrees of separation” claim came from a 1969 Stanley Milgram and Jeffrey Travers study where only 64 of 296 chains actually completed. Judith Kleinfeld’s 2002 reanalysis gutted the original evidence. But Duncan Watts’s 2003 email experiment and Lars Backstrom’s 2012 Facebook study eventually vindicated the underlying network structure --- with a more honest number attached.

The phrase “six degrees of separation” is one of the few pieces of academic jargon to fully escape the academy. It is the title of a John Guare play (1990), a Kevin Bacon party game, the spine of every Malcolm Gladwell-style talk on networks, and a stock metaphor used to argue that the world is smaller than we think. The claim is so familiar it feels like folk wisdom: pick any two strangers anywhere on Earth, and a chain of acquaintance no longer than six hops connects them.

The folk version traces back to a single 1969 paper in the journal Sociometry by Jeffrey Travers and Stanley Milgram, titled “An experimental study of the small world problem.” In it, Milgram --- already famous for the obedience studies --- and his collaborator described a procedure in which “senders” in Omaha, Nebraska and Wichita, Kansas were each given a packet and asked to forward it, through personal acquaintances only, toward a target person in Boston: a stockbroker. Each forwarder added their name and passed the packet to whoever they thought was closest, socially, to the target. The packets that arrived had taken an average of about 5.2 intermediaries. The famous “six degrees” formulation, popularized in the decades after, rounded up.

This is the article in a series on famous behavioral science studies that didn’t fully survive scrutiny. Six degrees of separation is an unusual case in the series because the popular claim, despite being based on weak original evidence, has held up remarkably well under modern large-scale tests. The honest summary is not “the claim was wrong” or “the claim was right.” It is “the original study was much shakier than the famous version implies, and subsequent work using much better data has converged on a number close to --- but not exactly --- what the popular version says.” The interesting part is the distance between what Milgram actually showed and what later, harder-edged evidence eventually corroborated.

The 1969 Milgram-Travers Study

The published 1969 procedure, as described in Sociometry Vol. 32, No. 4, pp. 425—443, was specific. Travers and Milgram recruited 296 starting participants in three groups: 96 residents of Nebraska randomly selected from a mailing list, 100 “Nebraska stockholders” recruited through investor records (chosen to have an occupational similarity to the target), and 100 residents of Boston (a near-target control). Each sender received a small booklet (the “tracer”) containing the target’s name, address, occupation, place of employment, college, year of graduation, wife’s maiden name, and hometown. Senders were instructed to forward the booklet, with their own name added to a roster of intermediaries, to one person --- a first-name acquaintance --- whom they believed could move it closer to the target.

The target was the same person for all chains: a stockbroker living in Sharon, Massachusetts, working in downtown Boston. The senders were told nothing about him beyond what was in the booklet. Each intermediary, in turn, did the same thing: pick someone closer, forward the booklet, send a tracer card back to the researchers documenting the hop.

Of the 296 chains that were started, 64 completed: 24 from the Nebraska random group (out of 96 = 25%), 16 from the Nebraska stockholder group (out of 100 = 16%), and 24 from the Boston group (out of 100 = 24%). The overall completion rate was 64 / 296, or about 22 percent. The other 232 chains died somewhere along the way --- a sender who never forwarded, an intermediary who lost interest, a packet that ended up in a wastebasket. The 22 percent completion rate is not buried in the paper; it is reported in Table 1. But the result that became famous, and the only result that filtered into the popular consciousness, was the mean number of intermediaries in the completed chains: 5.2. Rounded to a slightly punchier number, this became “six degrees of separation.”

Two structural features of the study deserve to be stated plainly before any critique. First, the completed chains were short: a mean of 5.2 and a maximum of about 11 intermediaries. Second, the convergence point of completed chains was strikingly narrow: 48 percent of the chains that completed reached the target through one of three terminal contacts. One man --- “Mr. Jacobs,” a clothing merchant in Sharon --- delivered 16 of the 64 completed packets to the stockbroker himself. The network funnel was not evenly distributed; it concentrated through a few hubs.

These features are themselves intellectually interesting. They suggest that, when paths exist, they are surprisingly short, and that hub-like individuals are doing most of the social bridging. Both of these claims have, since 1969, been substantially confirmed by other methods. The trouble is not the structural finding. The trouble is the inferential leap from “completed chains had a mean of 5.2 hops” to “any two people on Earth are separated by no more than six degrees.”

Kleinfeld 2002: The Critique That Should Have Killed the Folk Version

In 2002, Judith Kleinfeld --- a psychologist at the University of Alaska Fairbanks --- published “The small world problem” in Society Vol. 39, No. 2, pp. 61—66. The paper is one of the most cited pieces of methodological criticism in modern social-science history, and the central argument is mostly bookkeeping: she just made the original 1969 results’ weaknesses unignorable by going back to Milgram’s archives at Yale.

Kleinfeld documented several things.

Completion was the exception, not the rule. The 22 percent overall completion rate is dismal as a measure of “how connected is the world.” If only one in five chains reaches the target, and the failed chains might have, on average, been much longer than the completed ones, then averaging only the completers systematically understates the typical social distance. A chain that died at step 4 doesn’t tell you the path would have been only 4 hops; it tells you the path was at least 4 hops and we don’t know how much more. Mean path length in completed chains is, statistically, a lower bound on the true network diameter --- and a biased one, because chains that found efficient routes are exactly the ones that survived.

Earlier pilots were worse. Milgram had run earlier, unpublished, pilot studies before the 1969 Sociometry paper. A 1970 Psychology Today article by Milgram described starting chains in Los Angeles aimed at targets in New York. The completion rate in that pilot was around 3 percent. A separate Milgram-Korte study aimed across racial lines in the United States (white senders, black targets) had a completion rate near 13 percent. Kleinfeld’s argument is that Milgram disproportionately publicized the results from his most favorable pilot --- the Omaha-to-Boston run --- while the wider pattern across his unpublished work was much less rosy.

Target and sender selection mattered enormously. The Sharon stockbroker was, by the standards of 1967, an unusually well-connected target: white, male, college-educated, in a profession (stock-brokering) whose work involved building large social and professional networks. One of the Omaha sender groups was itself preselected for occupational similarity (the “stockholders” group). Half the senders in the headline conditions were thus pre-tuned to be socially close to the target through a shared occupational world. The completion rate from the Nebraska stockholder group was actually lower (16%) than the random Nebraska group (25%) --- a finding which itself complicates the obvious “homophily helps” story --- but the basic point holds: this was not a random sample of human pairs, it was a sample carefully curated to maximize the chance of a path being found.

The chain-length distribution is selection-biased. Even granting the 5.2 mean as accurate for the completed chains, the 232 incomplete chains carry exactly zero weight in that mean. If you assume that incomplete chains would, if they had completed, have averaged a longer path than the completed ones (which is the natural prior --- shorter paths are easier to find), then the true mean is higher than 5.2, possibly considerably so. Kleinfeld did not propose a specific correction; she just noted that the published number is an artifact of conditioning on a small, biased subset.

Kleinfeld’s conclusion was unsparing: the empirical foundation for the “six degrees of separation” claim, as it was actually published in 1969 and as it had been popularized for thirty years, did not exist. She titled a section of her paper “An urban myth?” and answered, more or less, yes.

This was a devastating critique. In a healthy intellectual marketplace, it might have ended the popular use of the phrase. It did not. The phrase was already too useful, too pithy, and too embedded in non-academic discourse for a small piece of methodological criticism in a sociology journal to dislodge it.

What did partly dislodge it was the simultaneous arrival of much, much better data.

Watts, Dodds, and Muhamad 2003: The Email Replication

In August 2003, Peter Sheridan Dodds, Roby Muhamad, and Duncan J. Watts published “An experimental study of search in global social networks” in Science Vol. 301, No. 5634, pp. 827—829. The paper was, in spirit, a direct replication of Milgram-Travers using infrastructure that didn’t exist in 1969: email.

The procedure was a generalized small-world experiment. Volunteers registered on a website to participate. Each participant was randomly assigned one of 18 targets located in 13 different countries (Australia, Estonia, India, Italy, Norway, the United States, and others). Targets included an Ivy League professor, an Estonian archival inspector, a Norwegian veterinarian, an Indian technology consultant, and others --- a much broader spread than the single Boston stockbroker. Senders were given the target’s name, location, occupation, university (if known), and a few other identifying details, and asked to forward an email message to someone they knew on a first-name basis who they thought was socially closer to the target. Each forwarder was, in turn, asked to forward the message further, and to record their decision on the study website.

The scale was vastly larger than Milgram’s: 24,163 chains were initiated, involving over 60,000 participants across 166 countries. The total number of completed chains: 384. The completion rate: about 1.5 percent.

Two things to notice about that number. First, it is much, much lower than Milgram’s 22 percent. A skeptic could plausibly conclude from this that the world is, in fact, less connected than Milgram thought, and that high attrition is the norm rather than the exception. Second, despite the abysmal completion rate, the chains that did complete were short: median length of about 4, with most successful chains arriving in 5 to 7 hops. This is roughly consistent with Milgram’s mean of 5.2.

The Watts team did something important that Milgram could not: they modeled the attrition. Using statistical methods to estimate what the typical chain length would be if every chain had completed --- correcting for the censoring caused by attrition --- they estimated a true median chain length of about 5 to 7 in their global sample, and shorter (under 5) within homogeneous sub-populations like senders and targets in the same country. The point of this exercise is that even after honestly accounting for attrition bias (which Kleinfeld had correctly flagged as fatal to Milgram’s clean number), the underlying network really does appear to be small-world structured. The honest number is not the inflated “six,” but it is also not “twenty” or “fifty” --- it is in the same neighborhood as the Milgram-era claim, just with a wider error bar and without the false confidence.

Watts also identified why chains failed: not because the network was too large, but because people did not feel sufficiently motivated, did not have time, or did not believe their forward would help. The decay was driven by participant disengagement, not by the structural impossibility of finding a path. This matters because it suggests the network distance is real, but the social process of finding the path in a real-world chain is much noisier than the simple “people will helpfully forward” model implies.

Backstrom 2012: Four Degrees on Facebook

If the 2003 Watts study was a respectable replication, the next chapter was something Milgram never had access to: a complete, observable global social graph.

In June 2012, Lars Backstrom, Paolo Boldi, Marco Rosa, Johan Ugander, and Sebastiano Vigna --- the first author at Facebook, the others at the University of Milan and Stanford --- published “Four degrees of separation” at ACM Web Science 2012 (DOI 10.1145/2380718.2380723). They computed the actual shortest-path distance in the Facebook friend graph as it stood in May 2011: 721 million active users, 69 billion friendship edges.

This was not an experiment. It was a direct measurement. They used HyperANF, a probabilistic distance-estimation algorithm, to compute the distance distribution of the global friend graph and various sub-graphs (the US-only graph, Italy-only, Sweden-only, and so on).

The result for the global Facebook graph: average distance of 4.74, with 92 percent of all reachable user pairs separated by 5 hops or fewer. The median distance was 5. The diameter (longest shortest path between any two users in the largest connected component) was much higher, but the typical distance between two random Facebook users was less than five hops.

For sub-populations the distance was lower: within the United States alone, the mean was 4.37. Within Italy, 4.34. Within Sweden, 3.90. The within-country distances are notably shorter than the cross-country ones, which is intuitive but worth stating: most of your “small world” is within your country, language, or interest cluster, and the cross-cluster bridges add a hop or two.

A few caveats. Facebook friends are not the same as Milgram’s “first-name acquaintances.” The Facebook graph includes both close ties and weak acquaintances, and the average user in 2011 had about 190 friends --- a much denser connectivity than the average person’s real-world social network. The 4.74 figure is therefore a lower bound on the “true” social distance between any two humans, because real social ties are sparser than Facebook ties. A version of the calculation restricted to “close friends only” would yield a higher number. And of course not all humans were on Facebook in 2011 (it was 721M of ~7 billion people), so the figure speaks only to the subgraph of Facebook users.

But even with these caveats, the 2012 result is the most precise estimate we have of typical social distance in a large population, and it is in the same neighborhood as the Milgram number with a small downward correction. The popular framing, then, is roughly right --- it’s about four to six, not two and not twenty --- but the popular framing got there by happy accident from a 1969 paper that did not deserve to support such a strong claim.

The Theoretical Foundation: Watts-Strogatz 1998

In 1998, four years before Kleinfeld’s critique and five before the Dodds/Muhamad/Watts replication, Duncan J. Watts and Steven H. Strogatz published “Collective dynamics of ‘small-world’ networks” in Nature Vol. 393, No. 6684, pp. 440—442 (DOI 10.1038/30918). The paper is short --- a Nature letter --- and arguably one of the foundational documents of modern network science.

Watts and Strogatz introduced a class of network models that interpolate between regular lattices (where every node connects only to its nearest neighbors, producing long path lengths) and random graphs (where any node may connect to any other, producing short path lengths but no local clustering). They showed that even a small fraction of long-range “shortcut” edges added to an otherwise regular lattice produces a network with two properties simultaneously: high local clustering (your friends are largely each other’s friends) and short global path lengths (any node can be reached in surprisingly few hops). They called such networks “small-world networks” and demonstrated the structure in three real datasets: the actor collaboration graph, the US power grid, and the neural network of C. elegans.

The intellectual significance of Watts-Strogatz 1998 for the small-world claim is that it gives a theoretical explanation for why human social networks should have short path lengths: it doesn’t require dense connectivity, and it doesn’t require evenly distributed connectivity. A small number of “weak tie” shortcuts --- the friend-of-a-friend who happens to live overseas, the college roommate now working at a different company, the cousin who married into another country --- is sufficient. The theory predicts that human social networks should be small-world structured. The 2003 and 2012 empirical results then confirmed this prediction.

This is why “six degrees of separation,” despite the rickety original evidence, has held up. The claim is not just a popular saying. It is a prediction of a well-understood class of network model, and it has been empirically tested at scales --- 24,000 senders, 721 million Facebook users --- that were unimaginable in 1969.

What’s Honest Now

After Travers-Milgram 1969, Kleinfeld 2002, Dodds-Muhamad-Watts 2003, Watts-Strogatz 1998, and Backstrom et al. 2012, here is what can be said honestly:

The original 1969 study should not be cited as direct evidence for “six degrees of separation.” The completion rate was 22 percent, the chain-length statistic was conditional on completion and is therefore biased downward, the target was preselected for high connectivity, and at least one sender group was preselected for occupational similarity to the target. The number 6 is not in the original paper; the published mean is 5.2, and that mean is itself an artifact of selection.

The underlying claim --- that the typical social distance between two random people in a large population is short, in the rough neighborhood of 4 to 7 hops --- has been independently confirmed by methods Milgram could not access. The 2003 email experiment, after correcting for attrition, estimated 5 to 7 hops in a global sample. The 2012 Facebook measurement found an average of 4.74. Both numbers are in the same order of magnitude as the popular framing.

The theoretical foundation for why this should be true --- Watts-Strogatz small-world networks --- is well established. It does not depend on Milgram’s specific procedure. The phenomenon would exist (and be discoverable) even if the 1969 paper had never been published.

So the popular phrase is roughly correct, but it inherits its support from later work, not from its supposed source. The chain of inference --- “Milgram showed it, so it’s true” --- is wrong; the chain “small-world network theory predicts it and large-scale measurement confirms it, in the neighborhood of 5 hops” is correct. The number “six” is a folk-poetic rounding of measurements that mostly cluster around 4 to 5. The right answer to “is the world really six degrees?” is “yes, approximately, for reasons other than the experiment that gave us the phrase.”

Strategist Implications: Social Network Reach and Viral Marketing

For marketers, growth operators, and anyone trying to reason about how ideas, products, or content propagate through human networks, the small-world story carries several implications.

Viral reach math is rarely as flattering as it looks. The seductive version of the small-world idea --- “anyone in the world is six steps from anyone else, so my message could reach everyone in six hops” --- ignores everything we learned from the 1.5 percent completion rate in the Watts study. Connectivity exists in the graph, but the social process of forwarding through that graph is fragile. People don’t forward. They get distracted. They don’t see how the forward helps. They don’t know the next person well enough to bother. The 2003 result was that most chains die not because the path doesn’t exist but because some intermediary disengages. Viral propagation is bottlenecked by motivation and trust at every hop, not by network distance.

This translates directly to growth strategy: counting “potential reach” by multiplying through follower graphs (e.g., “if our 1,000 customers each have 200 friends, we reach 200,000 people”) is the marketing equivalent of citing Milgram’s 5.2 mean without mentioning the 22 percent completion rate. The graph permits reach. Motivation determines it. Empirical “K factor” (viral coefficient) measurement is the only honest way to know whether the social network actually carries the message.

Hubs do most of the bridging. In Milgram’s 1969 data, 48 percent of completed chains routed through three terminal contacts; one man, “Mr. Jacobs,” handled 16 of the 64 deliveries. On Facebook, the degree distribution is heavily skewed: a tiny fraction of users have orders-of-magnitude more friends than the median. Network reach is not democratic; it concentrates in hubs. Influencer-marketing strategies that target high-degree nodes have a real structural basis: a single hub can functionally collapse the social distance between sub-communities that would otherwise be many hops apart. The reverse is also true: censoring or losing a hub has disproportionate effects on global connectivity.

Within-cluster distance is shorter than cross-cluster distance. Backstrom 2012 found within-US distance of 4.37 and within-Sweden distance of 3.90, both substantially shorter than the global average. For practical targeting, your product is closer to your existing customers than the global “6 degrees” framing implies; it is also farther from random members of other clusters than the global average suggests. Lookalike audiences and customer-referral programs exploit the within-cluster shortness. Cold geographic or demographic expansion pays the full cross-cluster bridging cost, and that cost is real even on a small-world graph.

Weak ties carry the message across clusters. Mark Granovetter’s 1973 American Journal of Sociology paper “The strength of weak ties,” not addressed in detail here, makes the complementary point: it is the loose acquaintances, not close friends, who connect you to socially distant clusters and thus deliver novel information. The small-world result and the weak-tie result are facets of the same structural fact: the cross-cluster bridges are sparse, but they are sufficient, and they are disproportionately weak ties rather than strong ones. For content strategy, this means a post shared by a casual professional acquaintance reaches a structurally different (and often more valuable) population than a post shared by an inner-circle friend. (See granovetter-weak-ties for the longer treatment.)

Diameter is not distance. The diameter of the Facebook graph is high; the average distance is 4.74. These are different statistics. Worst-case reach (the diameter) is not the right number to plan against; typical reach (the median or mean) is. The same distinction matters in any operational decision involving network propagation: worst-case latency to reach a customer through a referral chain is much longer than the typical case, but the typical case is what drives steady-state growth.

The popular “six degrees” phrase, even as a slogan, is fine as long as it is used loosely. As a basis for any quantitative growth-modeling decision, it should be replaced with empirical measurement of your specific network and the actual completion rates of your specific propagation mechanisms. Milgram, Kleinfeld, Watts, and Backstrom collectively show that the structural claim is real but the operational claim --- “the message will, in fact, reach there in six hops” --- is much weaker than the structural one.

Sources

  • Travers, J., & Milgram, S. (1969). An experimental study of the small world problem. Sociometry, 32(4), 425—443. DOI: 10.2307/2786545
  • Kleinfeld, J. (2002). The small world problem. Society, 39(2), 61—66. DOI: 10.1007/BF02717530
  • Dodds, P. S., Muhamad, R., & Watts, D. J. (2003). An experimental study of search in global social networks. Science, 301(5634), 827—829. DOI: 10.1126/science.1081058
  • Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of ‘small-world’ networks. Nature, 393(6684), 440—442. DOI: 10.1038/30918
  • Backstrom, L., Boldi, P., Rosa, M., Ugander, J., & Vigna, S. (2012). Four degrees of separation. ACM Web Science 2012, 33—42. DOI: 10.1145/2380718.2380723
  • Granovetter, M. S. (1973). The strength of weak ties. American Journal of Sociology, 78(6), 1360—1380. DOI: 10.1086/225469

FAQ

Did Milgram actually claim “six degrees of separation”?

No. The published 1969 Sociometry paper by Travers and Milgram reports a mean of 5.2 intermediaries in completed chains. The phrase “six degrees of separation” is not in the paper. The number was rounded up and the phrase popularized over subsequent decades, most prominently by John Guare’s 1990 play of the same name.

How many chains actually completed in the original Milgram study?

64 out of 296 starting chains, or roughly 22 percent. The other 232 chains died somewhere in the middle. The famous 5.2 mean is computed only over the completed chains.

Why did Kleinfeld’s critique matter?

Because it documented, using Milgram’s own archives at Yale, that (1) the completion rate was low, (2) earlier pilots had even lower completion (in the 3 percent range), (3) the target was preselected for high connectivity, (4) some sender groups were preselected for occupational similarity to the target, and (5) the chain-length mean of completed chains is statistically biased downward as an estimate of the true typical social distance. The combination undermined the foundation of the popularized claim.

Was the small-world claim actually wrong, then?

No. Watts, Dodds, and Muhamad’s 2003 email experiment, after correcting for attrition bias, estimated true typical chain lengths of about 5—7 hops in a global sample. Backstrom et al.’s 2012 measurement of the Facebook friend graph (721 million users) found an average distance of 4.74 and a median of 5. The popular claim is roughly correct. The original 1969 evidence does not deserve credit for it.

Why did so many of the email chains in the 2003 study fail to complete?

Watts and colleagues analyzed the attrition and concluded that chains failed not because the network was structurally too large to be traversed, but because individual participants disengaged --- they didn’t have time, didn’t believe forwarding would help, or didn’t have a sufficient relationship with the next person. The structural network is small; the social process of forwarding through it is noisy.

Does this mean “six degrees” is wrong on Facebook because the number there is 4.74?

It means the popular phrase is approximately correct but rounds in the wrong direction. The actual numbers from the best modern measurements cluster around 4 to 5 within reasonably connected populations, with cross-country pairs adding a hop or so. “Six degrees” is fine as an order-of-magnitude statement and wrong as a precise one.

Is “four degrees of separation” a better catchphrase for marketers?

Empirically, yes, within reasonably dense populations like Facebook users in the same country. Operationally, no, because both numbers describe the structural shortest path, not the actual rate at which messages propagate through human networks. The operational rate depends on motivation, trust, and incentives at every hop, and is dramatically lower than the structural distance suggests. For growth-modeling purposes, the right number is your measured viral coefficient, not the structural diameter of your audience graph.

Does Watts-Strogatz network theory actually require Milgram’s empirical work?

No. The Watts-Strogatz 1998 Nature paper introduced small-world networks as a mathematical model and demonstrated the structure in three datasets unrelated to the Milgram experiment: the actor collaboration graph, the US power grid, and the neural network of C. elegans. The theory predicts that human social networks should have small-world structure regardless of whether Milgram’s 1967 packet-forwarding study had ever happened. The 2003 and 2012 empirical results confirm the prediction. Milgram’s contribution is historical and inspirational, not evidentiary.

Could anything still go wrong with the small-world claim?

Yes, several things. Modern measurements rely on observable network proxies (Facebook friends, email contacts, phone records) that are denser than real-world strong ties, so the “true” social distance using a strict definition of acquaintance is likely a couple of hops higher than the 4.74 figure. The Facebook graph is also non-random with respect to who was online in 2011 (younger, more urban, more developed-country). And the small-world property assumes a connected component; people in disconnected sub-graphs (genuinely socially isolated individuals, contacts only through dead intermediaries) are not reachable at any number of hops. The 4—5 range describes the typical case in connected populations, not a universal upper bound.

replication-crisis six-degrees-separation milgram-1969 social-networks evidence-evaluation
Share this article
LinkedIn (opens in new tab) X / Twitter (opens in new tab)
Atticus Li

Experimentation and growth leader. CXL-certified CRO practitioner, Mindworx-certified behavioral economist (1 of ~1,000 worldwide). 200+ A/B tests across energy, SaaS, fintech, e-commerce, and marketplace verticals.

About LinkedIn Newsletter