If you've worked in digital health for any length of time, you've sat through at least one product meeting where someone proposes "a community feature" as the answer to engagement.
It's almost always vague. Sometimes it's a forum. Sometimes it's a Discord. Sometimes it's a "social feed" the team will build because the team likes Strava. The pitch is that community drives retention, retention drives outcomes, and outcomes drive revenue. You get the idea.
The trouble is that "community feature" is not a behavior change mechanism. It's a UI surface. The actual mechanism, the thing that does or doesn't move the user toward the target behavior, is something underneath the UI, and most teams never name it.
A scoping review and realist synthesis published in JMIR on this exact question landed earlier this year. 81 studies. Most of them on WhatsApp and WeChat group chats used as behavior change interventions. The synthesis is the most useful thing I've read on the topic, and I want to walk through what it actually says, because the conclusions are sharper than most marketing decks I've seen on community features.
The headline counts
The review found 81 studies. WhatsApp showed up in 38 of them, WeChat in 21. So 59 of 81 studies, almost three-quarters, were on consumer messaging platforms that are not anyone's "wellness app." The reviewers also found that publication on this topic increased sharply after 2020, presumably because of pandemic-era research on remote intervention delivery.
Where did the studies cluster by topic?
- Mental health: 25 of 81 (30.9%).
- Maternal and child health: 20 of 81 (24.7%).
So roughly 55% of the literature is in two domains. That's worth knowing. If you're proposing a group-chat feature for, say, cardiac rehab, you have less direct evidence than you'd think. If you're proposing it for postpartum mental health, you have more.
Group chats most often functioned as the core intervention component (42 of 81, 51.9%). Less often as a reinforcement layer in a multi-component program (26 of 81, 32.1%). Standalone as the only thing the participant got, 13 of 81 (16.0%).
This is the first finding most product people miss. Most of the time the chat is the intervention. It is not a "community layer" sprinkled on top of a content app. It's load-bearing.
The realist synthesis bit, and why it's the actually interesting part
The reviewers ran a realist synthesis on top of the scoping review, which is a method I genuinely love because it forces the reviewer to ask "what mechanisms got activated, in what contexts, to produce what outcomes." Not "did the intervention work yes or no." Not "what was the effect size in the random-effects pooled model."
They identified 12 recurring context-mechanism-outcome configurations. These cluster into 5 domains.
- Capability and actionability.
- Confidence and motivation.
- Modeling and norms.
- Safe and supportive environment, plus access.
- Self-regulation and maintenance.
You can map these onto the COM-B model if you squint. Capability, motivation, opportunity, behavior. The synthesis is doing what COM-B was always supposed to do: name the actual mechanism, not the feature.
The five domains are activated or suppressed by:
- Facilitation quality.
- Group composition.
- Cultural alignment.
- Technological access.
- Social norms in the group.
Honestly this list is the part I'd put on the wall in a product team room. Because it tells you that the "community feature" can succeed or fail entirely on factors the engineering team doesn't control. A WhatsApp group with a great moderator and a culturally aligned cohort works. The exact same UI with a bad moderator and a mismatched cohort fails. The product is identical. The mechanism is different.
The facilitation problem nobody wants to fund
If you read between the lines of every "successful" group-chat intervention in this literature, you find a moderator. A nurse. A community health worker. A trained peer. Someone whose job is to keep the group active, redirect the conversation, surface the right content, manage the difficult member.
This is BCT 3.1 (social support, unspecified) and BCT 3.2 (social support, practical) and arguably BCT 3.3 (social support, emotional) in the Michie taxonomy. All three of them living inside one human moderator.
The product teams I've worked with do not want to fund that moderator. The moderator is variable cost. The moderator doesn't scale. The moderator has bad days and quits. The whole point of digital health, the pitch goes, is that you can serve a million users for the same cost as serving a thousand. A human moderator violates that pitch.
So what teams do instead is ship a community feature and assume "the community will moderate itself." It doesn't. The literature is unambiguous. The groups that worked were not self-organized. The groups that worked had a paid skilled facilitator with a curriculum.
This is why most "community" features in health apps are dead by month three. There's no facilitator, the early adopters churn, the lurkers never engage, and the product team writes a postmortem about "low engagement" without naming the mechanism that was missing.
What this means if you're a product team
I want to be careful here. I am not telling you to ship a WhatsApp group. The review's clearest finding is that the platform isn't the mechanism. The mechanism is the social structure inside it. Build that structure on whatever surface you like.
Concrete moves I'd actually make based on the synthesis:
Pick a domain where the literature is dense. Mental health and maternal health have ~55% of the evidence. If you're shipping a community feature in those domains, you're standing on better ground than if you ship one for, say, healthful eating in midlife adults.
Decide whether the chat is core, adjunct, or standalone. This is a real product decision and most teams skip it. If it's core, the rest of the app should not duplicate the work the group is doing. If it's adjunct, the group should reinforce what the rest of the app teaches. The synthesis maps these three modes and the outcomes look different across them.
Budget for facilitation. Skilled moderators with a curriculum. Yes, it's variable cost. Yes, that breaks the unit economics most digital health VCs want to see. The synthesis is clear that this is what makes the difference between groups that work and groups that don't. You can ship without it. Most teams will. They will mostly fail to produce behavior change.
Match group composition deliberately. Cultural alignment is named as a key activating condition. A multi-state, English-only group of strangers is the easy thing to ship. The literature suggests it underperforms a small cohort matched on language, life stage, condition severity, and culture. Cohort-matching is product work. It's not technically hard. It's organizationally hard.
Watch the norms. Healthy groups produce different norms than unhealthy ones, and norms compound. A group where the loudest voice is dismissive about adherence will tank the adherence of the quiet members. A group where the loudest voice is hopeful and consistent will pull the quiet members along. The product question is: do you measure norms? Do you have a mechanism for changing them when they go bad? In most apps, no.
The broader reframe
I think the most useful thing about this review, beyond the specific findings, is that it forces you to stop treating "community" as a feature category and start treating it as a behavior change technique with mechanisms, conditions of activation, and conditions of failure.
When a vendor pitches you a "community-driven product" for, say, diabetes self-management, you are now allowed to ask: which of the five activating domains is your product targeting, what's your facilitation model, what's your group composition strategy, and how do you measure norm drift? If they don't have answers, the product is going to do what most community products do. Generate a lot of feature surface. Move very few outcomes.
Read the review directly. It's open-access, the supplement is good, and the realist synthesis section is short enough to skim in 30 minutes.
If you've shipped a group-chat intervention and you think the review is missing something, please let me know. I'm always open to correction. The literature on this is moving fast and I won't always be right about it.