How the Peer Effect Does Your Member Retention for You

A 2025 study built on seven years of one gym’s attendance records puts a number on the peer effect, and on how much of your member retention your members are doing for each other. Yijun Chen at Imperial College and Nitin Mehta at Rotman worked from monthly visit and subscription records covering 1,491 adult members of an upscale class-based gym in the American mid-west, running from 2012 to 2018.

What the research found

Chen and Mehta needed to see who trained with whom, and the class register already held it. Classes ran small, typically three to eight people, so members who kept landing in the same session left a trail the data could follow. For any two members the authors worked out what share of one member’s shared-class visits the other accounted for, then counted the pair as direct peers once that share sat above the median. Peer is doing a lot of work in that sentence. It means a threshold on a spreadsheet, and not friendship in any sense a member would recognise.

The adult network this produced averaged 213 members a month, and it was badly lopsided. In March 2013, eight members out of 240 sat above fifty peers each, and those are the members the study calls high centrality, meaning the most connected tenth by that count. This is a stronger claim than social proof, which is about how people read a crowd. Here the crowd is specific people, and the data knows which ones.

Separating real influence from people who simply chose the same gym

Working out whether one member’s attendance actually moves another’s is harder than it looks, because people who train together also picked the same gym and the same six o’clock slot for reasons of their own. Chen and Mehta got at it through the cancellations. Of 1,314 on record, 86 came with a reason that had nothing to do with the gym, mostly relocation and family circumstances, and a member moving to Spain is not leaving because the Tuesday class got worse. Those 86 churns are what the causal claim rests on. It is a thin set of instruments for seven years and 1,491 people, and the size of the effect is the part worth replicating before anyone bets much on it.

The peer effect came out at 0.18, significant at p < 0.01. A gym that lifts every member by one visit a month finishes with 1.63 per member, because each member’s extra visit drags a further 0.63 out of the people they overlap with. Eighty-three per cent of members also showed strong positive carryover, so a good month leaves the next one easier and a thin one leaves it harder. That carryover is the same machinery behind what happens to attendance after a gym challenge ends.

What happens to member retention when a connected member goes

Everything past that point is modelled, not observed. The authors took the estimates, simulated a 228-member network and removed one person from it. Losing a member with two peers barely registers, and six months on the network is still averaging 11.7 visits a month against 11.8 with nobody leaving at all. The 85-peer version looks nothing like that. Six months on, average visits are down to 8.4, and projected six-month revenue for the network has gone from $189,858 to $120,256, with the share of members not subscribing in a given month climbing from 20.5% to 27.7%.

Treat that revenue figure with suspicion, because the case for some cancellations costing more than others rests on it. The peer effect itself was estimated from behaviour people actually performed. Everything downstream of it, the six-month cascade included, came out of running the model forward, and the authors say plainly that the rule governing how the network rewires after somebody leaves does not fall out of their model. They specified that rule separately, grounded in the data but not estimated from it. The full paper is open access through Imperial’s repository if you want to check the counterfactuals yourself.

Where the peer effect shows up in your member retention

Your cancellation process treats every leaver as the same event, and Chen and Mehta’s numbers say the events are not the same. A member with two training overlaps and a member sitting in your best-connected tenth are worth very different amounts of money to you, but both of them get the same call and the same line on the lapsed list.

The at-risk report has its own version of the problem, because it ranks members on their own visit frequency. A member whose closest training overlap left three weeks ago reads as perfectly healthy right up until the week she doesn’t, which is how member retention problems stay invisible until they are already priced in. It is the same blind spot behind members quitting while they are still improving: the number exists in your system and never reaches anyone who could act on it.

The member retention move, and how you’ll know it worked

Split your leavers before you decide how hard to fight for them. No network model is needed for this: when someone cancels, count how many other members regularly overlap with them in the same sessions, then put that count next to the rest of your roster. The names near the top of that ranking are the ones this research says carry a cost past their own direct debit. Everyone else costs you about what their membership was worth, and your report already tells you that.

Instinct will send you to the wrong member here. Work the group around a well-connected leaver and the members most likely to take your offer are different people from the members whose response travels furthest. Chen and Mehta found a negative correlation between how connected a member is and how much their own subscription probability lifts when you aim a promotion at them. Conventional retention advice says target whoever shows the biggest individual lift. In a network that turns out to be the short-term answer only. The paper’s own conclusion is that short-horizon campaigns should favour the less connected peers, while long-horizon ones favour the more connected. Spillovers start small and build.

Watch attendance, not churn, to see whether your member retention work is landing. Track visit frequency for the people who overlapped most with a leaver across the month or two after they go. A departure that is going to travel arrives as reduced visits long before it arrives as anyone cancelling, and that ordering is what the model produces.

What Peloton built instead

Peloton has spent two years building a member retention version of this in software. In September 2024 it launched Teams inside the app, letting members create invite-only groups of up to 100 friends and run shared goals and challenges against each other. The ceiling has since gone to 50,000 members per Team, with search added so people can find a group to join. Peloton now runs Official Teams led by its own instructors, covering everything from cross training to menopause health. Peloton’s own account of what Teams are for is that members “keep each other accountable”.

A Peloton subscription is about as easy to cancel as a subscription gets, with no contract and no conversation at a reception desk. Teams exists to give a member something to leave besides Peloton, which is a switching cost built out of other people rather than out of contract terms. But your six o’clock class already owns that asset and paid nothing to build it, which is roughly the reverse of Peloton’s position, and it is the same asset behind the instructor who calls a member’s name mid-class.

Where this member retention idea goes wrong

Their sample is narrower than it first sounds. Chen and Mehta kept the month-to-month members, who were 76% of users. They cut anyone with fewer than six visits or less than three months on the books, on the grounds that those were mostly trial members passing through. That exclusion is defensible for their question, but awkward for yours, because the members a gym owner worries about most are the members this study removed. If early leavers are your problem, the confidence a new member walks in with is a better place to start than this paper.

Much the same goes for the gym. It was an upscale site built around classes of three to eight people, where the average member came in more than seven times a month and the premium plan ran at $348.50. A 24-hour access-card gym whose members arrive and leave alone has no network of this kind for the effect to travel along.

Misusing this is worse than ignoring it. Once you can name your best-connected members, the temptation is a loyalty tier for the connectors and less attention for everybody else, and Chen and Mehta tested nothing of the sort. Visible tiering also runs into what accumulated status does to the members who don’t have it. A modelling detail argues for the same caution. Their model takes the network as given in each month and never explains how it forms, so an intervention that deliberately reshapes who trains with whom is acting on what they held fixed.

The Question

Which of last month’s cancellations came from the top tenth of your network?

People Also Ask

What is a peer effect in gym retention?

A peer effect is the influence one member’s behaviour has on another member’s, with nobody attempting to persuade anyone, and in a gym it is a live member retention mechanism rather than an abstraction. In a gym it runs through attendance, so a member becomes more likely to come in while the people they overlap with are coming in, and less likely once those people stop. Chen and Mehta estimated the effect from seven years of one gym’s records and put the coefficient at 0.18. They used cancellations with external causes, such as relocation, to separate real influence from the fact that people who train together chose the same gym in the first place. Its practical size is that lifting every member by one visit a month produces 1.63 visits per member once members start pulling on each other.

How do I find out which gym members are connected to each other?

Your class register already holds the raw material, because Chen and Mehta built their whole network out of which members attended the same small classes repeatedly over time. Their own definition is statistical, counting two members as connected when their shared-class overlap sits above the median, so anything built by hand will be rougher than theirs. Pick a member, pull the sessions they attend most, and count how many other names recur across those sessions more than occasionally. What the exercise should produce is a ranking rather than a friendship map, because the research’s claim concerns members near the top of the distribution and not any particular pair.

Do all gym cancellations cost the same?

Not according to this study, though the difference is the model’s output and not a loss anyone counted. Chen and Mehta simulated a 228-member network losing one person. A member with two connections left almost no trace six months on. The 85-connection version came with average visits across the whole network falling from 11.8 to 8.4, and projected six-month revenue dropping from $189,858 to $120,256. Scope is the qualifier that matters most here, because high centrality in the study means the top tenth of the network. It is a claim about a small group of your members, not a reason to treat every cancellation as a crisis.

Closing

Every cancellation form asks the member why they are leaving. None of them has a box for how many people used to train beside them, and on this evidence that is the more expensive number.

You’ve got the move for what happens when a member cancels. If you want the same behavioural-science read across your whole retention sequence, from onboarding to win-back, book a free 30-minute chat.

---

References: Chen & Mehta 2025 (Imperial College Business School / Rotman School of Management, Marketing Science 45(2)); Peloton, The Output, updated March 2026 (Teams and Challenges).