What we actually know, and what we only repeat.

This is the single most important number in the field and it is still routinely softened in decks to "engagement is a challenge." Three in a hundred users are still there after a month. The reason it matters is the source: this is panel data, not self-report and not the developer's own analytics, which is exactly where optimistic numbers usually come from. If you are building in this space and your retention projections do not start from something like this baseline, you are planning against a fiction.

Read this next to Baumel and the gap does the arguing for you. Trial attrition is bad; real-world attrition is far worse. That gap is the most under-discussed number in digital mental health, because efficacy is established in the first condition and the product ships into the second. An intervention that works in people who stay is not the same claim as an intervention that works.

Twenty years old, from outside health entirely, and still the most useful theoretical frame available for this problem. It is the foundation my own instrument is built on. The practical consequence is the part teams miss: removing friction and adding value are different interventions with different returns, and doing one well does not substitute for the other. A product can be delightful and still bleed users, because delight does not cancel friction. It sits beside it.

The most cited thing on this list by a wide margin, and deservedly, but it is frequently used as a citation rather than as a tool. NASSS is a framework for organizational and system complexity. It is excellent at explaining why a telehealth program dies in a health system, and it was never designed to tell a product team which screen is losing them users on Tuesday. Both questions are real. Reaching for NASSS to answer the second one is a category error I see often in grant applications.

The finding I did not expect and now consider the most practically important is the graduation effect. It means a meaningful share of what the field counts as failure is success wearing the same clothes, and no engagement dashboard currently distinguishes them. The honest limitation: the sample is young adults, cross-sectional, and the model operates at the composite level. Anyone applying it to a clinical or older population should say so out loud.

What struck me writing it was how narrow the studied populations are relative to the populations these tools are marketed to. The methodological range is wide; the demographic range is not. That mismatch should make anyone cautious about generalizing engagement findings, mine included.

Included here because the failure mode is the same one at a different scale. People fall out of crisis systems at the handoffs between services, exactly as they fall out of apps at the transitions between screens. In both cases the loss is invisible to whoever owns only one segment of the path.

This is where my method started. Large-scale user-generated text is an underused evidence source in this field, partly because it is messy and partly because it says things campaigns would rather not hear. It is also, unlike a survey, unprompted. People are telling you what they think without being asked a question that shapes the answer.