I used AI to assist in writing this post, and it’s likely that >30% is AI-generated text.
Continuing our April 2026 post on the cascade effect
Our two previous posts established the problem and why it compounds: approximately 18 million women in Mexico are living through some stage of the climacteric (between ages 39 and 60+), and each woman we reach may represent 5–6 people benefited through spillover effects in her caregiving network. This post answers a different question: what kind of digital intervention can actually reach them and retain them?
The short answer: fully automated digital interventions (apps, chatbots, unguided self-help) are cheap per user but fail on the metric that determines whether cost-effectiveness calculations are real. The 30-day retention rate for mental health apps in real-world deployment is 3.3%. Human-guided online interventions are completed at nearly double that rate, with effect sizes two to three times larger.
In Mexico, the infrastructure to deliver human-guided programs exists not through home broadband, but through smartphones and mobile data: 97.1% of internet users in Mexico connect via smartphone, and WhatsApp reaches 95.3% of them. That is the channel. Intervention design has to match it.
In our first post we described Lunava as a CBT-based intervention for the menopausal transition. In the second, we expanded the framework to the full climacteric: a 15-to-25-year arc beginning around age 40 and extending well into post-menopause. Here we sharpen that frame one step further.
Drawing on data from the INEGI 2020 Census and the CEPAL 2024 Demographic Observatory, approximately 16–18 million women in Mexico are currently between ages 40 and 65, living through some stage of this transition. The average age of natural menopause in Mexico is 47.9 years (compared to 51 globally), meaning perimenopause frequently begins in the late thirties. The INEGI 2020 Census data covers ages 40–65; the clinical floor of the intervention starts at 39, given the earlier onset of perimenopause in Mexico.
The psychological burden across this range is well-documented. Clinical studies conducted in Mexico report anxiety symptoms in 91.6% of women in the climacteric, depressive symptoms in 85.2%, and difficulty making decisions in 85.5%. A comparative analysis found that Hispanic women experience a 43% greater deterioration in quality of life relative to other groups studied, a figure that likely reflects, in part, the accumulated caregiving burden we documented in our previous post.
The treatment gap is stark: only 4% of people with mental health conditions in Mexico receive psychological care. The public health system has 6.6 psychologists per 100,000 inhabitants. There are no culturally adapted, scalable, and psychologically rigorous programs for this population at any stage of the climacteric anywhere in Latin America.
The instinct when designing for scale in low-and-middle-income countries is to minimize cost per user by removing human labor from the delivery chain. Fully automated apps and chatbots have a marginal cost close to zero per additional user. The economic logic looks solid.
The clinical and behavioral evidence contradicts it.
The most important number in this discussion is not the effect size reported in a controlled trial. It is the proportion of users who actually complete the intervention under real-world deployment conditions.
An analysis of 93 mental health applications found a median 30-day retention rate of 3.3%. More than 80% of users dropped off within the first ten days, before any therapeutic dose could accumulate. This is not an isolated finding: it aligns with systematic reviews showing dropout rates of 26–47% even under controlled trial conditions, where participants are actively recruited, monitored, and incentivized.
Mental health apps have worse retention than any other app category. General health apps have a 30-day retention benchmark of 6.29%. Mental health apps fall below that floor.
A cost-effectiveness calculation built on this retention rate is not optimistic. It is wrong. If an intervention reaches 10,000 women but only 330 complete it, the scalability that makes automation attractive evaporates when adjusted for who actually finishes.
A pre-registered and replicated meta-analysis (published 2023, PMC) synthesized evidence across different levels of human contact in online interventions for depression. The results follow a clear dose-response gradient:
Level of human contact | Effect size (Hedges' g) | Completion rate |
|---|---|---|
No human contact (fully automated) | 0.21 | ~52% |
Minimal contact before the intervention | 0.44 | — |
Support during the intervention | 0.56 | — |
Contact before, during, and after | 0.76 | ~64% |
The difference in effect size between no contact and full contact is not marginal. It is the difference between a weak signal and a clinically meaningful one. A separate meta-analysis covering 56 randomized controlled trials and 13,335 participants confirmed that the degree of human contact was the strongest predictor of both clinical outcome and completion rate.
The 12-percentage-point difference in completion rates (64% vs. 52%) understates the real gap, because trial completion rates are systematically higher than real-world deployment rates. Under real conditions, the comparison is closer to 60% for guided interventions versus 3–4% for fully automated ones.
Generative AI chatbots for mental health have attracted significant investment and media attention. Some recent RCTs show promising results against waitlist controls. This deserves an honest examination.
The critical finding is what happens when AI is compared to active human-guided interventions, rather than no-treatment controls. Every published study comparing AI chatbots to human therapists for moderate-to-severe anxiety or depression has found that human therapists produce better outcomes.
There are additional concerns specific to the population and context in which Lunava operates. Language models are trained predominantly on English-language data from Western, educated, industrialized contexts; their capacity for culturally pertinent responses in Mexican Spanish, for a population of middle-aged women navigating a life stage with significant social and cultural specificity, has not been validated.
For populations in low-and-middle-income countries who lack access to alternatives, the risk is not just inefficacy.
If fully automated digital tools are insufficient and in-person care is geographically and economically inaccessible, the viable path is online delivery with human guidance. The question is whether the infrastructure for that exists in Mexico especially outside major urban centers and for women who may not have home broadband.
The answer, from Mexico's own data, is more encouraging than is commonly assumed with an important caveat about where access actually lives.
According to the INEGI ENDUTIH 2024 (published May 2025):
But the home broadband figure understates real connectivity, because it measures access through fixed connections. The infrastructure most relevant to a mobile-delivered intervention is different:
Mexico's internet infrastructure is, in practice, a smartphone infrastructure. This has a direct consequence for intervention design: a program that requires a computer, stable broadband, or a formal institutional platform will structurally exclude a significant portion of the target population. A program that runs on smartphone, through apps women already use, will not.
The most important data point for our delivery model:
Women in the climacteric age range in Mexico, particularly between 40 and 55, are not intensive users of new platforms. But WhatsApp, Facebook, and Instagram have penetrated deeply into this demographic group, even in lower-income contexts and rural areas. These are not platforms that need to be learned; they are platforms women already use daily.
This matters for intervention design in a way that is frequently overlooked in mental health programming in middle-income countries. The question is not only whether the target population has internet access, but whether it has frictionless access to the intervention channel. Downloading a new app, creating an account on an unfamiliar platform, navigating a new interface, each step introduces dropout risk before the first therapeutic session begins. A group intervention delivered via WhatsApp, Zoom, or Google Meet eliminates most of that friction.
For women with the caregiving profile we described in our previous post 39.7 hours per week of unpaid domestic and caregiving work (ENUT 2024, INEGI), frequently responsible for children, elderly parents, and in some cases grandchildren attending in-person psychological care requires not only transportation costs but an irreplaceable time block in an already saturated week.
Online delivery eliminates:
Qualitative research consistently documents time and transportation as the primary barriers to mental health care access for women in low-and-middle-income contexts, particularly those with caregiving responsibilities. Removing these barriers does not guarantee participation, but it does strip away the first layer of structural exclusion.
Online delivery does not eliminate the human facilitator that is the central distinction of our model. What it eliminates is everything else: the physical space, the commute, the in-person materials. The marginal cost of adding participants to an online group cohort is close to zero up to group capacity. That is where the model's efficiency lives not in eliminating human contact, but in eliminating everything except human contact.
The literature on group CBT delivered online is thinner than on individual internet-based CBT, but the available evidence is consistent with findings on individual contact.
A 2022 descriptive study comparing adherence in online versus in-person mindfulness group therapies found that online group formats maintained comparable retention rates to in-person groups, with shared group accountability functioning as an engagement mechanism independent of modality. This aligns with what we have observed in our early group work: women are less likely to drop out of a session when they know other participants in the same group are expecting them.
For the climacteric population specifically, group dynamics serve an additional therapeutic function: normalization. Qualitative data from our early prototype and from the broader literature on peer support in menopause suggest that the moment of recognition hearing another woman describe exactly the same 3 a.m. insomnia, the same unexpected emotional volatility, the same professional anxiety carries therapeutic weight that individual therapy cannot fully replicate. This is not simply a delivery feature; it is part of the mechanism of change, and it is entirely absent in automated interventions.
We are close to launching our first pilot. We do not have outcome data yet, and we are not reporting any here. What we are reporting is the evidence base for our design decisions, so that the EA community can evaluate the reasoning before the evidence is available.
Our design decisions and the evidence behind each:
Coverage of the full climacteric arc (39–65+): Evidence from the SWAN study and the Lancet 2024 Series documents that psychological burden persists and accumulates throughout the entire transition, not only during perimenopause. A program limited to women in active perimenopause omits the majority of the affected population.
The questions we are still working through:
The three open questions from our previous post remain open:
How much of the 5–6x cascade multiplier can we credibly claim before we have second-order outcome data?
Should CBT content be adapted and how for established post-menopause versus perimenopause?
And what does real retention look like across 8 sessions with a population of women carrying the caregiving load we described?
On that last question: we expect our own retention data to be more informative than anything we can cite. The 3.3% figure for automated apps is a floor. The 64% from guided RCTs is a ceiling under controlled conditions.
We are sharing this analysis before we have results because the design reasoning is something the EA community can evaluate now, and early feedback helps us improve what we measure and how.
We especially welcome:
We are beginning to generate our own data. Until we have it, we are trying to be precise about what we know, what we are assuming, and where those assumptions might be wrong.
Karina Benitez and Fabiola Balmori | [email protected] / [email protected] | lunava.ngo
