TL;DR: An app like Headspace does provide strong improvements on wellbeing, but remains ≥2$\sigma$ below cash transfers due to high costs. Reducing those costs may bring mental health apps in line with cash transfers or group therapy. The linked Observable notebook allows you to interact with the estimates yourself.

Software provides enormous benefits to cost-effectiveness; there is zero marginal cost of supplying a software service to a new user (only acquisition costs). There are very few health interventions that have this property, as most health issues require physical treatment, usually from paid experts. As software has made content delivery orders of magnitude more cost-effective, it might also have this effect on non-physical health treatments.

I am interested in understanding if delivering fully self-guided, software-based mental health interventions might have a similar cost-effectiveness to GiveWell’s estimates for wellbeing improvements (selecting these because they’re a higher bar to clear than most other estimates of wellbeing improvement, and generally agreeable to the community).

I’m not interested in this being the most comprehensive CEA; this is purely exploratory, to diagnose whether this is an issue worth pursuing. I will usually try to underestimate benefits and overestimate costs throughout, and I’m only interested in getting within an order of magnitude or two of the most cost-effective interventions. The only way to really test these apps would be to run experiments in collaboration with them, which is much, much harder.

I’ll define ‘self-guided mental health app’ broadly, as anything vaguely intending (explicitly) to improve mental health, and self-guided as anything with zero 1:1 contact with other people to guide you (i.e. an app that provides a library of guided meditations is self-guided).

To estimate the cost-effectiveness, let’s break it down into:

• What is the short-term impact of a short treatment course on mental health symptoms?
• How does that impact translate to other measures?
• How many users will continue to use an app in the long term?
• What is the customer acquisition cost of such an app?
• What is the development cost of such an app?

Prior Art

I’m extremely grateful to the long history of the mental health topic on this forum. I found 2 specific CEAs/discussions that focused on mental health app development:

• Stan Pinsent & CEARCH’s analysis of self-administered psychotherapy: Finds a 12% chance of being as effective as a GiveWell top charity. I find a lower chance that existing apps are as effective, due to lower effect sizes for self-guided interventions (I think). We both agree that effectiveness is only achieved on the long tail and that our assumptions are shaky.
• George Bridgwater & Charity Entrepreneurship’s 2-hour CEAs for promising mental health interventions: Specifically, they analyse guided self-help workbooks, gratitude journalling apps, apps for CBT, and apps for mundane, sub-optimal happiness. For a CBT app, they estimate lower retention, but lower cost-per-install, and achieve a much higher cost-effectiveness (possibly on the order of cash transfers or group therapy).

I wanted to dive deeper specifically on self-guided mental health. I hope also by including distributional estimates (which Pinsent does) that we can gain some clarity on where the uncertainty is, and what levels we can pull to increase effectiveness.

Short-term impacts of treatment

(We will focus specifically on depression as a health outcome. It is the largest mental health category by incidence and health burden, and the most studied. By narrowing in this way we can only serve to underestimate total impact; although, as below, self-guided app-based treatments for depression seemingly generalise to anxiety and vice-versa.)

Linardon et al. (2019) appears to be a very well-conducted, well-cited meta-analysis of RCTs of smartphone apps that aim to treat mental health conditions (a broad definition, but one which is sub-divided later in the analysis). Further research uncovered other meta-analyses, but this was more robust in all relevant aspects (comprehensiveness, bias/quality checks, etc.), which I hope excuses the singular focus on the study (ex. it specifically aimed to improve Firth et al. (2017) by including ~50 more RCTs).

They analysed 54 RCTs of apps that were measured against depressive symptoms and found a pooled standardized mean difference of 0.21–0.36. When adjusting for potential publication bias using trim-and-fill, the SMD increased to 0.32–0.49 (but only had data for 37 RCTs).

• For anxiety, it was 0.20–0.40 ($k$ = 39) and adjusted to 0.28–0.49 ($k$ = 31)
• For stress, it was 0.21–0.48 ($k$ = 27) and adjusted to 0.30–0.57 ($k$ = 22)
• For quality of life, it was 0.29–0.42 ($k$ = 43) and adjusted to 0.32–0.46 ($k$ = 37)

There was no statistically significant difference between whether apps targeted depression or anxiety directly; they had comparable uplift effects on both measures regardless. This indicates we may be able to generalise between conditions, which seems to be supported by the similarity in effect sizes above.

Apps which included professional guidance had a larger depressive symptom SMD (0.34–0.62; $k$ = 15) than those which didn’t (0.15–0.31; $k$ = 37; difference $Q$ = 0.002).

• For anxiety, it was 0.36–0.70 with ($k$ = 12), and 0.12–0.30 without ($k$ = 27; difference $Q$ = 0.001)
• For stress, it was 0.35–0.79 with ($k$ = 10), and 0.12–0.36 without ($k$ = 17; difference $Q$ = 0.010)
• For quality of life, it was 0.39–0.64 with ($k$ = 13), and 0.22–0.35 without ($k$ = 30; difference $Q$ = 0.001)

We don’t have adjusted SMD for the self-guided condition. However, since trim-and-fill only improved the SMD in the overall condition, we can be conservative and just use the direct values for self-guided interventions.

const selfGuidedDepressionSMD = dist`${normalInterval(0.15, 0.31, { p: 0.95 })}`

GiveWell (2023) estimate the cost-effectiveness of direct cash transfers, group psychotherapy, and bednet distribution, providing us with a useful benchmark.

They apply the following discounts:

• 0.7–0.9 for study social desirability bias
• 0.75–0.9 for publication bias
• 0.7–0.85 for lower effectiveness at scale / outside trial contexts
• A lower ‘effect multiplier’, which doesn’t directly apply to our analysis as it accounts for decay of effects over time (which we will do later with our own dropout analysis).

I believe it is reasonable that the first three multipliers have an effect here, so we can include them before making our comparison.

const selfGuidedDepressionSMDDiscounted = dist`${selfGuidedDepressionSMD} * (0.7 to 0.9) * (0.75 to 0.9) * (0.7 to 0.85)`

Translating impacts to life-satisfaction-point-years (WELLBYs)

Unfortunately, there isn’t much literature on the relationship between a standard-deviation improvement in depression symptoms and quality- or disability-adjust life years (QALYs/DALYs), which are an extremely well-studied measure of disease burden. The GBD’s calculations for these numbers are confusing at best.

Instead, WELLBYs are a somewhat standard measure that complement (QALYs/DALYs) in cost-effectiveness analyses. A treatment that improves a condition by one WELLBY increases the individual’s life satisfaction by a single point for the length of a year.

The best expert guess on the standard deviation of life satisfaction as a whole is about 2.108, which comes from this GiveWell analysis combining extensive satisfaction data from high-income countries with other surveys from low-income countries.

const expertLifeSatisfactionStdev = dist`(1.9 to 2.7) * 0.6 + 1.86 * 0.4`

Finally, how should an standard-deviation decrease in depressive symptoms correspond to a standard-deviation increase in life satsifaction? It’s not clear that this is a 1:1 relationship. Plant & McGuire (2023) discuss possible discounts. We will make a conservative estimate of 0.7–1, as GiveWell do.

We can put these together to find an average value of 0.21 WELLBYs caused by using a self-guided mental health app consistently (at post-test).

const selfGuidedWellby = dist`${selfGuidedDepressionSMDDiscounted} * ${expertLifeSatisfactionStdev} * (0.7 to 1)`

Our model doesn’t assume any ‘spillover’ effects. GiveWell strongly adjust them downward because they are very hard to estimate and there isn’t a lot of evidence for them. To save myself effort, I’ve omitted them entirely in the hopes that it will steelperson my argument a bit better; but even if we believe they’re significant I really doubt they will change the final order of magnitude of cost-effectiveness.

Self-guided app adherence

Under our model, the above WELLBYs are only achieved at post-test in studies that account for dropout rates. In reality, a huge proportion of users will pick up an app, use it a couple of times, and never open it again. The remainder tend to become long-term users.

• Baumel et al. (2019) systemtically review Android mental health apps using SimilarWeb data, which have a 3.3% 30-day retention rate (IQR 6.2%)
• Apptopia reports a 6.29% 30-day retention rate on health and fitness apps, with Headspace and Calm at 7.65% and 8.34%. They also noted top iOS apps tend to have a ~30% higher retention rate than top Android apps.
• Forbes report Insight Timer has a 16% 30-day retention rate (via Apptopia)

(As an aside, a lot of this data is obtained through user deception and privacy violations. I did not pay for the pro versions of these products so that I could access more up-to-date information, at the risk of lower-quality data. There may also be a sampling bias inherent to this collection as some of the identified data collectors are wellness-focused apps such as screen time trackers.)

Baumel et al. (2019) also found full decay curves for apps across categories, and the daily percentage open rate for installations (calculated by dividing an estimate for total installs by daily active opens). For all mental health apps, this was 4.0% (IQR 4.7%), suggesting that this is approximately the limiting retention. This implication concurs with the decay curve plot from the paper, and our intuitions about habit-building (some number of users will reach a sort of ‘escape velocity’ and stick with the app for the long term).

Comparing Baumel et al. (2019)’s data against Apptopia’s, it seems wise to assume mental health apps have a lower retention rate than health & fitness overall. We should also increase the estimate slightly to account for iOS apps making a good share of total users.

A ~4.0% mean retention isn’t an awful guess, and we can then assume that a 99th-percentile retention would be Insight Timer’s 16% (evidenced by its newsworthiness). We should assume an exponential distribution; most apps should achieve no retention.

const mentalHealthAppRetention = dist`${exponential(1 / 0.04)}`

In our model, we do not assume that continuing to use such an app has diminishing returns over time. Rather, continuing to use the app ‘tops up’ the effect, and neglecting to use it doesn’t have any meaningful decay rate. This is in line with other mental health interventions such as therapy or exercising, where continued use is required for the effect to continue.

We can now compute the number of WELLBYs per new user of the app.

const perUserWellby = dist`${selfGuidedWellby} * ${mentalHealthAppRetention}`

User acquisition costs

We are specifically interested in CPI (Cost Per Install). It’s hard to get good data on CPI, since most ad networks don’t publish it (and it obviously depends on a huge number of factors).

• Revealbot (a Facebook Ads optimiser) estimate U.S. CPI in February 2024 at ~$1, with a 12-month average of $3.43
• Business of Apps (a blog) estimate CPI as ranging between $0.50–$6 over a long time period
• Skimming various other sources through search engines generally concurs with this

Notably, CPI is dramatically lower in poorer countries (less crowded markets with less potential ROI). It’s not clear what the CPI is for the health category. We can also note that Facebook and other ad networks provide extremely good optimisation tools that can even take a desired Return On Ad Spend (ROAS) and deliver as much breadth as possible.

All of this is to say that I am not confident we can produce a narrower estimate without empirically verifying this with data from a real app in our exact category and market.

const marketingCostPerDownload = dist`(0.5 to 6) * ${acquisitionCostAdjustment}`

Development costs

(Note: We assume zero service costs, as apps run on device and use free OEM services for things such as syncing. Most content apps will host their files on a CDN, but I’m not convinced the hosting costs would be a significant proportion of revenue; anyhow, I add a cost multiplier later).

After wrangling with a few different accounting methods, I think the most reasonable way to account for development costs is to establish a rough amortised cost per user. This is a frozen-in-time estimate (i.e. it divides the total cost of providing an app in a given year by the total number of users it has in that given year), so it doesn’t account for past costs. Apps are likely to get more cost-effective over time.

We will assume the major development cost is labour. Here are the rough employee & user counts for the top mental health apps:

• Headspace: 1,200 @ 70,000,000 downloads
• Calm: 400 @ 135,000,000 downloads
• Breeze: 11–50 @ ~2,000,000 downloads
• BetterHelp: 5,000 @ 80,000,000 users
• Paired Couples Counselling: 250 @ 8,000,000 downloads
• Insight Timer: 350 @ ~10,000,000 downloads
• Balance: 57 @ 8,000,000 downloads
• Shmoody: 2–10 @ ~1,000,000 downloads
• Daylio: 2–10 @ ~20,000,000 downloads
• Bearable: 7 @ 600,000

Let’s see how that relationship works.

Obviously, a company which optimises for cost-effectiveness should be able to get these numbers down. But with our observed values, we can take that the number of downloads per employee is around 25,000–2,500,000.

Employee salaries probably average out around 25,000–250,000 USD. I am very unconfident in this estimate—some of these companies operate in low-income areas (Daylio, Breeze) and others employee cheap content generation labour (Headspace, Calm, BetterHelp).

const developmentCostPerDownload = dist`(1 / (25000 to 2500000)) * (25000 to 250000)`

Benchmarking against other interventions

I will also include a 1–1.5x multiplier on costs, as there is a strong possibility of unaccounted costs.

const totalCostPerDownload = dist`(${marketingCostPerDownload} + ${developmentCostPerDownload}) * (1.1 to 1.5)`

We can finally arrive at our total cost-effectiveness. Let’s plot the rough density below, on a log scale, and compare it to GiveWell’s estimates for the cost-effectiveness of GiveDirectly, StrongMinds, and Against Malaria Foundation.

We can compute some statistics on this to make the relationship a bit clearer. The mean of this distribution is 3.47 WELLBYs per US$1,000, and its standard deviation is 5.50, which means:

• GiveDirectly is 1.5$\sigma$ better
• StrongMinds is 3.1$\sigma$ better
• AMF is 12.7$\sigma$ better

I would be happy to conclude that the average self-guided mental health app is unlikely to achieve similar cost-effectiveness to a top GiveWell charity. You should probably not start donating to Headspace tomorrow.

How would we improve it?

This analysis considers the current cost-effectiveness of self-guided mental health apps. I think it is quite likely that significant improvements can be made to surpass at least GiveDirectly. And even if we can’t, this section can be useful for determining the sensitivity of this cost-effectiveness analysis in terms of its parameters.

Reduce hiring

Meditation apps, for example, require an evergreen content library and therefore have significant ongoing costs and headcount needs. This could be significantly reduced to just a small development team by developing an app that doesn’t structurally require ongoing content development. It may even be possible to reduce the size of the marketing team with advancements in marketing automation over the next decade.

I personally believe most modern tech companies are very bloated, so I think this could reduce development costs by an order of magnitude, or perhaps even more.

Reducing hiring by an order of magnitude adds about 40% to the cost-effectiveness, but this isn’t nothing—it could bring us roughly up to the same cost-effectiveness as cash transfers. Given the wide range of observed employee-download ratios, this would be a very promising direction.

Reduce cost per employee

This is trickier, since the counterfactual employee likely would’ve worked at another effective organisation or donated any extra salary. However, it might be possible to lower employee salaries a little bit? In practice, halving salaries (as low as we could reasonably go) only adds 14% more cost-effectiveness. I would not pursue it as a direction.

Market to low-CPI regions

CPI is usually half an order of magnitude lower in low- and middle-income countries. This is an excellent opportunity, as those countries also tend to have lower baseline life satisfaction & worse existing mental health support.

Lower-income countries tend to have CPIs around US$0.05–1.50, and adjusting to roughly this value nearly triples our cost-effectiveness (although there is a lot of uncertainty about those CPIs in practice). I think this would be an extremely promising direction—apps could accept donations for customer acquisition and subscription costs in LMICs, while taking revenue from higher-income countries, without significant changes to their structures or business models. Advocacy in this area toward existing providers might be effective!

Improve dropoff rates

It might be possible to lower dropoff rates. We saw Insight Timer report a 16% retention rate, which would double our impact. They claim it’s due to their lack of growth hacking techniques, which might be true if users are put off by an app that feels too slick.

At the very least, it seems reasonable that meditation apps have high dropoffs because they require significant upfront commitment, and a delayed emotional payoff (and are just not effective for some segment of users). All of these elements could be improved upon to increase the dropoff rate. 

Improving the dropoff rate from 4% to 6% increases the cost-effectiveness by about 40%—but we’d have to be careful about DAU-hacking in practice.

Improve effect size

It may also be possible to achieve a larger-than-average effect size by structural means. The easiest-to-imagine mechanism here would be to simulate a therapist using a chatbot; this could even have a larger effect due to the increased availability & lower cost. Or, it could be true that GiveWell’s discounts are overly conservative. But in general, I can’t speak for mechanisms as clearly as I can for the above. Improving the effect size by 50% increases the cost-effectiveness by 60%.

Conclusion

I mostly did this work for myself, but I hope it’s useful for other EAs looking into this area. I am very new to the forum. If you’re giving feedback, I am less interested in specifics on accuracy, and much more in large, directional things I’ve overlooked that could have caused me to overestimate cost-effectiveness. Thank you!