Summary

Considering the expected benefits of greater access to mental health drugs (i.e. improved health from reduced depression and correspondingly greater economic output), the expected costs (i.e. worse health from adverse drug side-effects and correspondingly lower economic output), as well as the tractability of direct delivery (i.e. a charity working to distribute mental health drugs), I find that the marginal expected value of distributing mental health drugs to be 28 DALYs per USD 100,000, which is around 4% as cost-effective as giving to a GiveWell top charity.

Key Points

  • Importance: This is a low-to-moderately important cause, with 1.28 * 106 DALYs at stake from lack of access to mental health drugs (note: not depression in general) from now to the indefinite future. Around 94% of the net burden is health related, while 6% is economic in nature.
  • Neglectedness: Whatever efforts that governments/nonprofits/business are making to solve the problem (e.g. channelling more resources to treating depression) is clearly insufficient, given the growing disease burden lost to depression per capita.
  • Tractability: A moderately tractable solution in the form of direct delivery of SSRIs to LMICs patients by collaborating with local primary care physicians is available.

Discussion

  • The low cost-effectiveness is strongly driven by the lack of stickiness of the solution, from individual people relapsing back into depression after treatment and from population turnover (but mainly the former).
  • However, this report does not imply that mental health as an overall cause area is unimportant – there could well be other cost-effective causes/intervention pairs within mental health, particularly on the psychological therapy side (c.f. the promising work of Strong Minds or Vida Plena).
  • This analysis may understate the effectiveness of SSRIs given the possibility of continued patient improvement after the observed trial periods (usually 6 to 12 weeks) of studies examining SSRI effectiveness.
  • The calculation of the proportion of the global disease burden of depression attributable to lack of access to treatment is highly uncertain, relying as it does a complicated model that already makes a number of simplifying and potentially biasing assumptions.
  • Note that diagnosis accuracy (by doctors, of potentially depressed patients) in LMICs is potentially worse than the rate used in the analysis, given how little time LMIC doctors have with their patients (less than a minute, in Bangladesh!), making the estimate here upwards-biased.
  • Impact scales with cost for direct delivery charities, so our ultimate tractability estimates aren't sensitive to costing here compared to other interventions and cause areas.


Expected Benefit: Improved Health from Better Access to Mental Health Drugs

The primary expected benefit from expanded access to mental health drugs is the reduction in incidence and severity of depression in the global population, and hence improved health. Overall, around 7.99 * 106 DALYs are at stake here, with this benefit is modelled in the following way.

Morals Weight & Scale: To calculate the global disease burden of depression in DALY terms in 2023, I take the baseline 2019 GBD data on the global burden of disease of depressive disorders, both in terms major depressive disorder and dysthymia/chronic depression, and then adjust it:

  1. For growth since. Refer to discussion below on how the projection of future disease burden is made.
  2. For the public underestimating the badness of depression relative to its true badness as evaluated by actual sufferers, and
  3. To incorporate self-harm and suicides from depression.

The Happier Lives Institute raises other issues with existing GBD estimates, but I do not believe these criticisms are sustained:

  1. The issue of cultural underreporting seems overstated insofar as (i) cultural differences can plausibly cause different responses to emotional turmoil in a way that will cause differential rates of depression, and (ii) reported depression rates and its impairment correlates vary positively, which would be less likely if there were systematic underreporting of the former for cultural reasons.
  2. The issue of neglecting mental illness that is insufficiently bad to reach diagnostic threshold can be factored out here, insofar as such even-milder-than-mild depression cannot be cured through SSRIs and hence is irrelevant to an evaluation of the value of better access to mental health drugs; and 
  3. The issue of happiness/positive mental health beyond just avoiding depression being neglected is comparatively insignificant if we think the average negative human experience tends to be worse than the average positive human experience is good – but that is an issue beyond the specific scope of this piece of cause prioritization research.

In terms of the proportion of the global disease burden of depression that is actually attributable to lack of access to treatment, we can model this by making the simplifying assumption that getting treatment is a function of desire for treatment (itself a function of desire to get better, concerns over the social stigma etc) subject to affordability constraints. Per the Moitra et al systematic review and Bayesian meta-analysis, the percentage of people getting minimally adequate treatment for depression in high-income, upper-middle income, and lower middle- & low-income countries was 23%, 8% and 3% respectively. Assuming that the wealth of high-income countries means that everyone who wants treatment gets it, and assuming desire for treatment does not vary systematically with country income, 23% is the theoretical maximum number of people who would get treated given perfect healthcare access, with the remainder being people who are unwilling to seek treatment for one reason or another (e.g. social stigma). This being the case, the fraction of untreated depression sufferers who can nonetheless be helped by increased access, relative to the global population of depression sufferers, is simply (a) the difference between current treatment rates in low and lower-middle income countries and the 23% maximum just described, (b) the same difference for upper-medium income countries, (c) zero, for high income countries, all of which is (d) combined as a weighted average that factors in relative population sizes. Then, to translate this fraction of untreated depression sufferers who can nonetheless be helped by increased access, into the fraction of the global disease burden attributable to lack of access to treatment, we simply adjust for the fact that treated individuals experience a lower disease burden relative to their untreated counterfactual baseline, with this in turn calculated on the rough empirical basis that everyone is treated within a year and that the reduction in disease burden due to treatment varies linearly in time.

Finally, as for the question of how much of this remaining disease burden would be reduced by comprehensive access and use of mental health drugs – this is a function of the per patient effectiveness of SSRIs, which is taken to be the inverse of the median NNT (i.e. number of people that need to be treated to prevent one additional bad outcome).

Overall, the potential harm eliminable from fully comprehensive access to mental health drugs in DALY terms, as a function of (a) the global burden of disease of depression in DALY terms, (b) the proportion of the global disease burden attributable to lack of access to treatment, and (c) the proportion of the remaining disease burden of depression that would be reduced by comprehensive access and use of mental health drugs, is 2.81 * 106 DALYs.

Persistence: The problem of depression isn't going anywhere anytime soon, and indeed, is expected to grow. Putting in a place solution will potentially reduce the problem not just for one year but across multiple years; and in terms of how this aggregated multi-year benefit is calculated –

Firstly, I discount for the probability of the solution not persisting (i.e. treatment effect being reversed). Assuming an intervention of providing full treatment as soon as possible in a single year to all depressed individuals who want treatment but are not currently receiving it, reversal is a function of both the relapse rate as well as the replacement rate (i.e. the population experiencing turnover from some people dying out and others entering into it). The relapse rate is calculated as the average of the year-on-year relapse rate across a decade. The population turnover rate is approximated as the inverse of life expectancy. Together, this yields a very considerable reversal rate of 10% per annum.

Secondly, I calculate the proportion of disease burden that will remain after being counterfactually solved – or in fact, because the problem will grow, the increased disease burden over time. To do this, I use a simple theoretical model that treats total DALYs lost to depression as a function of DALYs lost per capita and population size.

Note that whatever efforts that agents (i.e. governments, nonprofits and businesses) are making to solve the problem (e.g. channelling more resources to treating depression in the public health system, private non-profit hospitals or private for-profit hospitals respectively), and whatever impact that structural trends are having (e.g. social atomization increasing social alienation and consequently the prevalence and severity of depression; or higher GDP allowing more resources to be spent on treatment; or increasing median age increasing DALY burden due to older people being more susceptible to depression), all this will occur through and is hence accounted for by the variable of DALYs lost per capita. The only exception is population size - increasing population size (and eventually, declining population) will, for a given per capita rate of DALYs lost to depression, increase (and eventually decrease) total DALYs lost; and this variable, of course, is handled separately in the model.

And to implement this model and project how the problem is expected to evolve over the years, I:

  1. Project future DALYs lost to depression per capita by estimating the year-on-year change through a linear regression of DALYs lost to depression per capita on discrete time.
  2. Use UN estimates of projected future population growth; and 
  3. Multiply each future year's DALYs lost per capita and population size to obtain the expected total DALY burden for each year.

That said, I limit the extrapolation to 2100 - the high-confidence UN estimate ends there, and I assume constant population after; moreover, I cap the growth of DALYs lost per capita since an important unfavourable underlying trend of increasing median age will eventually cease once population stabilizes, and after that I model the per capita rate as declining as quickly as it was rising, on the basis of the remaining drivers being net favourable (various agents trying to solve the problem, while cultural and economic structural drivers cancel out).

The projected growth of DALYS lost to depression is shown in Diagram 1.

Diagram 1: Growth of DALYs lost to depression over time

For 2101 and beyond, modelling the expected decline as equal to the DALY per capita growth rate up till 2100 yields a discount of 0.2% per annum.

Thirdly, I discount for the probability of the world being destroyed anyway (i.e. general existential risk discount). I take into account the probability of total nuclear annihilation, since the benefits of saving people from depression in one year is nullified if they had already died in a previous year. For the exact risk of total nuclear annihilation, I take it to be one magnitude lower than the risk of nuclear war itself, since nuclear war may not kill everyone. For the probability of nuclear war, I use the various estimates on the probability of nuclear war per annum collated by Luisa, but with accidental nuclear war factored in, and then calculate a weighted average that significantly favours the superforecasters. The reason for this is that (a) the estimate of the probability of intentional nuclear war based on historical frequency is likely biased upwards due to historical use being in a MAD-free context; (b) the probability of accidental nuclear war based on historical close calls is highly uncertain due to the difficulty of translating close calls to actual probabilities of eventual launch; and (c) experts are notoriously bad at long-range forecasts, relative to superforecasters. Meanwhile, I do not take into account other existential risks like supervolcano eruption and asteroid impact, since the chances of those occurring at all is very marginal per Denkenberger & Pearce, let alone the chances of such events killing everyone and not just most people. Overall, therefore, I treat the general existential risk discount to be just the risk of nuclear war but adjusted a magnitude down  (i.e. 0.07% per annum).

Fourthly, I apply a broad uncertainty discount of 0.1% per annum to take into account the fact that there is a non-zero chance that in the future, the benefits or costs do not persist for factors we do not and cannot identify in the present (e.g. actors directing resources to solve the problem when none are currently doing so).

Overall, by taking the projected trend of DALYs lost to depression up to 2100, and discounting each year's DALY burden using the other per annum discounts (i.e. solution reversal discount, existential risk discount, uncertainty discount), the total amount of DALYs due to depression that would be counterfactually solved by a total solution is shown in Diagram 2.

Diagram 2: DALYs available for solution by SSRI distribution

Value of Outcome: Overall, summing the discounted per annum relative values for 2023-2100, and using a perpetual value formula to incorporate the 2101+ values, this yields a raw perpetual value of improved health from better access to mental health drugs of 2.9 * 107 DALYs.

Probability of Occurrence: As for the probability that lack of access to mental health drugs is a problem (or, equivalently, the probability that expanding access positively impacts welfare, this is a function of four critical variables:

  • Probability that depression is an actual problem that harms people: Unlike longtermist problems, there is no uncertainty that depression is an actual problem. The World Health Organization (WHO) has demonstrated that depressive disorders are one of the leading causes of disease worldwide, with the reported prevalence throughout the world of depressive episodes being 16 per 100,000 per year for males and 25 per 100,000 per year for females, and with depression being the fourth leading cause of disease burden in the world, accounting for 4.4% of total DALY burden as of 2004. And of course, the problem has hardly been solved since then. Trivially, I assign the probability to be 1.
  • Probability that a doctor correctly identifies a patient as depressed and requiring treatment: Empirically, doctors have about a 55% success rate.
  • Probability that people given access to mental health drugs will actually use them as prescribed: Statistically, people adhere to the prescribed course only 50% of the time.
  • Probability that SSRIs work: This is given as 99.999% (i.e. 1 less the probability that the null hypothesis is true and that the observed relationship between SSRI usage and positive clinical response at post-treatment is purely due to random chance).

Overall, this yields a probability of 27.5% that lack of access to mental health drugs is a problem whose solution would be beneficial.

Expected Value: Hence, the expected value of improved health from better access to mental health drugs is 7.99 * 106 DALYs.

 

Expected Benefit: Increased Economic Output

Accompanying the reduction of the disease burden will be lower healthcare costs and improved productivity, thus promising an expected benefit of increased economic output. Overall, this economic benefit is worth the equivalent of around 5.14 * 105 DALYs, and is modelled as follows.

Moral Weights: The value of doubling consumption for one person for one year is 0.21 DALYs, which is calculated as a function of (a) the value of consumption relative to life from GiveWell's IDinsight survey of the community perspective, as adjusted for social desirability bias, and (b) CEARCH's estimate of the value of a full, healthy life in DALY terms. For more details, refer to CEARCH's evaluative framework.

Scale: To calculate the scale of the potential benefit, I first look at the economic burden of depression relative to annual income per depression sufferer. And to do this, I rely on three separate estimates.

Firstly, we have Greenberg et al's estimate of the total US economic cost of major depressive disorder (covering direct medical costs, loss of earnings due to suicide, and reduced labour supply and productivity due to depression causing both absenteeism and presenteeism), which I then adjust to account for chronic depression (assuming economic burden varies proportionally with disease burden). This is then divided by the number of depression sufferers in the US to get the economic burden of depression per depression sufferer, which is then divided in turn by annual income to yield the economic burden of depression relative to annual income per depression sufferer.

Secondly, we have Sobocki et al's estimate of the total European economic cost of depression in euros (covering direct medical costs as well as economic costs from morbidity and mortality). Converting this to USD, we can then divide it by the number of depression sufferers in Europe to get the economic burden of depression per depression sufferer, and then further divide by annual income to get the economic burden of depression relative to annual income per depression sufferer.

Thirdly, we have Hu et al's estimate of the total Chinese economic cost of depression (covering both direct as well as indirect costs), which by dividing through with the number of depression sufferers in China, we can use to get the economic burden of depression per depression sufferer. Dividing that in turn by annual income again yields the economic burden of depression relative to annual income per depression sufferer.

In creating a weighted average of the three estimates, I penalize the Sobocki et al estimate given the chances of potential misestimation, from our analysis treating Europe as the EU for simplicity. Additionally, I penalize both the Greenberg et al and Sobocki et al estimates given that the cases of depression we will be targeting with expanded access to mental health drugs will be in LMICs, for which US/EU data will be less representative compared to Chinese data.

This weighted average in turn yields the degree of consumption doubling per depression sufferer if their depression is cured.

Then, we can calculate the total number of consumption doublings achievable through better access to mental health drugs in 2023 – the base figure is just the degree of consumption doubling per depression sufferer if their depression is cured, multiplied by the 2019 prevalence of depression, but further adjusted (a) to account for increased prevalence in 2023, (b) to focus only on the economic burden attributable to a lack of access to treatment, and (c) to account for the fact that SSRIs only have a certain effectiveness. All this accounted for, we see that there is a potential total of 576,000 doublings in consumption from treating all the depressed who want treatment but do not currently receive it.

Persistence: The same per annum discounts and the same projections of the disease burden (and hence economic burden) over time, as discussed in the previous section, are used here as well.

Value of Outcome: Overall, the raw perpetual value of increased economic output is 1.89 * 106 DALYs.

Probability of Occurrence: Same overall probability as before is applied.

Expected Value: All in all, the expected value of increased economic output is 5.14 * 105 DALYs.

 

Expected Cost: Worse Health from Adverse Drug Side-Effects

With all that said, increasing access to mental health drugs is a not a risk free exercise – there are, unfortunately, potential adverse side-effects to SSRI consumption, and that is an expected cost we have to calculate. Our best estimate is that -4.84 * 105 DALYs are at stake here, with our calculations as follows.

Moral Weights: For the expected DALYs lost from adverse events per SSRI course, I calculate this by:

  1. Estimating the DALYs lost (total disease burden divided by prevalence, and adjusted for average SSRI course duration) for each potential adverse event that occurs as a result of SSRI intake, subject to those whose occurrence are statistically significant and which are severe enough to be in the GBD
  2. Multiplying by the event probability (as given by the inverse of the number needed to harm), and 
  3. Summing the expected harm of individual potential adverse events to get aggregate expected harm, of around -0.040 DALYs per SSRI course.

Scale: Taking the number of depression sufferers in 2023, and multiplying by the proportion of depression sufferers who want treatment but aren't getting it (estimated using the same model as discussed previously, where getting treatment is a function of desire for treatment subject to affordability constraints), we get the number of untreated depression sufferers who want treatment and who would potentially take SSRIs if said drugs were made available to them in 2023 - around 44.6 million people.

Persistence: In terms of the number of years that the harm persists – recall that we are modelling the solution as providing full treatment as soon as possible in a single year to all depressed individuals who want treatment but are not currently receiving it. This being so, the cost is a one-time matter, with the SSRIs prescribed for consumption within the first year and cessation thereafter. Hence, the issue of solution reversal is irrelevant, and the total across-time cost is just the amount of disease burden from adverse drug side-effects potentially causable in a given year discounted by existential risk and uncertainty: 99.83% of the baseline year's total disease burden, in this case.

Value of Outcome: Overall, the raw perpetual disvalue of worse health from adverse drug side-effects is -1.77 * 106 DALYs.

Probability of Occurrence: The probability that access to mental health drugs causes a real problem in terms of adverse drug side-effects is a function of the following three variables:

  • Probability that adverse drug side-effects are a genuine problem: For each side-effect, this probability is 1 less the estimated p-value; and to get the overall probability, we take a sum of the individual probabilities weighted by the relative share of expected harm, to get a probability of 99.6%.
  • Probability that a doctor correctly identifies a patient as depressed and requiring treatment: As discussed previously.
  • Probability that people given access to mental health drugs will actually use them as prescribed: As discussed previously.

Overall, this yields a probability of 27.39% that access to mental health drugs causes a real problem in terms of adverse drug side-effects

Expected Value: All in all, the expected disvalue of worse health from adverse drug side-effects is -4.84 * 105 DALYs.

 

Expected Cost: Decreased Economic Output

Corresponding to the health damage done by SSRI side-effects is the economic damage. This damage is probably worth around -5.01 * 104 DALYs, with the modelling as follows.

Moral Weights: As discussed in the section on increased economic output from curing depression.

Scale: To calculate the scale of the potential cost, I first look at the economic burden of adverse drug events relative to annual income per adverse drug event sufferer, using three estimates to do so.

The first estimate is Classen et al's, of the US economic cost of an adverse drug event (specifically, the excess medical costs), which divided by annual income yields the economic burden of adverse drug events relative to annual income per adverse drug event sufferer.

The second estimate is Gyllensten et al's, of the Swedish economic cost of an adverse drug event (covering both direct medical costs as well as indirect productivity costs), which divided by annual income obtains the economic burden of adverse drug events relative to annual income per adverse drug event sufferer.

The third estimate is Shi et al's, of the Chinese economic cost of an adverse drug event (as a function of total costs, covering both direct medical costs and indirect productivity costs, and of number of patients. Converting the output to USD and dividing through by annual income then produces another estimate of the economic burden of adverse drug events relative to annual income per adverse drug event sufferer.

In the weighted average, I penalize Classen et al for not taking into account indirect costs, and on top of that I penalize both Classen et al and Gyllensten et al – since US/Swedish data will be less relevant compared to Chinese data, when the people who we are targeting for expanded access to SSRIs (i.e. people who want treatment for their depression but who currently aren't being treated) will be in LMICs. This then gets us the degree of consumption doubling per adverse drug event sufferer if they had not taken SSRIs we had provided, which we can use to calculate the total number of consumption doublings prevented (by expanded access to mental health drugs causing more adverse drug events in 2023), by multiplying the first value with the number of untreated depression sufferers who want treatment and who would potentially take SSRIs if said drugs were made available to them in 2023.

Persistence: The total across-time cost as a percentage of the baseline year's total disease burden is as described previously.

Value of Outcome: Overall, the raw perpetual disvalue of decreased economic output is -1.83 * 105 DALYs.

Probability of Occurrence: Same overall probability as discussed in the health costs of adverse drug events is applied here.

Expected Value: All in all, the expected disvalue of decreased economic output is -5.01 * 104 DALYs.

 

Tractability

To summarize our tractability findings: we can solve 0.000003 of the problem per additional USD 100,000 spent on direct delivery.

Direct delivery is, of course, not the only option – but there may not be better alternatives, as we shall see. To ensure access to mental health drugs, we actually have three potential methods: policy advocacy to get rich donor governments to fund greater access to mental health drugs in LMICs; policy advocacy to get LMIC governments themselves to pony up; and direct delivery itself (i.e. a charity distributing SSRIs). Generally, policy advocacy is more effective given the enormous leverage from using low-cost regulation or mobilizing low counterfactual government resources – and of course lobbying rich donor governments in particular is especially cost-effective given the extremely low counterfactual value of spending on rich world citizens. However, in this context, lobbying rich donor governments suffers from the fact that foreign aid is never popular, and in any case, mental health is probably seen as a less important cause than saving people from physical harm (e.g. AIDS etc). As for lobbying LMIC governments themselves – the same reality holds, where governments prioritize public health other than mental health when allocating limited funding, in the context of already limited health budgets. Hence, direct delivery is probably the best solution here, tied perhaps with lobbying rich donor governments, and we examine the former as a solution. 

In terms of evaluating the proportion of disease reduction we can achieve from direct delivery of SSRIs, we have two key parameters to consider.

  1. The proportion of depressed individuals, who want treatment but who currently lack it, and that primary care providers (which our direct delivery charity will be partnering with for better access) can reach and help.
  2. The chances of convincing various primary care providers to accept the receipt of SSRIs and accompanying training

 

For (1), we have to estimate four critical sub-parameters:

(a) The maximum number of primary care physicians that our hypothetical charity will be able to support. This will depend critically on the number of drugs our charity can procure, as well as on the number of depressed patients (which will be a function of sub-parameters (b) and (c), discussed below). For now, let us focus on the quantity of drugs to be procured – this depends on both the hypothetical budget above, as well as the likely proportion of the budget that the charity can spend on core services. To estimate the latter, I use the outside view and take a weighted average of what AMFMalaria Consortium and New Incentives spend on their core services, penalizing NI in the weighing for being less procurement-oriented (n.b. I do not use an inside view as I do not believe I have sufficient experience on direct delivery logistics to make an accurate call). With the budget for core services, we can then divide by the price of generic SSRIs – which cost $174 per 6-month course, and are cheaper than their branded equivalent while being no less effective or safe – while adjusting for average SSRI course duration (around 7 weeks to really feeling better and another 6 months after that to be safe).

(b) The number of unique patients each primary care physician sees a year. To estimate this, I use developed world figures for lack of better data; note that this will be biased downwards given that developing world doctors will see a lot more patients (at the cost of lower consult times). Consequently, I choose the high end of the average figures (around 1900 per year) to balance this out.

(c) The percentage of patients that are depressed. This is taken as just global prevalence rates, on the understanding that the charity will be targeting unserved areas where people are not already getting treatment via SSRIs (let alone CBT) such that we do not have to factor in the counterfactually treated.

(d) The total size of the depressed population that wants treatment but currently do not receive it.

Overall, I expect a charity with this budget to be able to solve about 0.0008% of the disease burden suffered by depressed individuals who want treatment but who currently lack it.

 

For (2), and the operational challenge of persuading various primary care providers to accept the receipt of SSRIs and accompanying training, I evaluate this with both an outside and inside view.

For the outside view, I use multiple reference classes: (i) the general success rate of CE charities; (ii) physician attitudes towards antidepressants (since if physicians are sceptical – specifically, sceptical that antidepressants work – they would be less likely to accept the charity's offer to supply SSRIs); and (iii) the degree of support for lifelong learning amongst rural Chinese physicians (as an indicator of how willing LMIC doctors will be with respect to subjecting themselves to additional psychiatric training). That said, all these reference class have significant flaws: (i) the CE incubatee reference class is biased downwards insofar as charities can fail for other reasons (e.g. funding) than inability to deliver on their core service; (ii) the physician view of antidepressants reference class suffers from uncertainties about scepticism translating to wholesale rejection, as well as significant external validity considerations; and (iii) the lifelong learning reference class suffers from uncertainties inasmuch as positive attitudes to lifelong learning need not translate to actual likelihood of a busy doctor agreeing to and attending more training, along with similar external validity considerations. Overall, I penalize the latter two reference classes relative to the first, yielding a weighted average of a 49% probability of success.

For the inside view, I reason as follows. While doctors will certainly want to help their patients, there is also a lot of stigma against the mentally ill (even by healthcare providers themselves against their own patients), and of course the training will be a burden. On this basis, a <=10% chance of total success seems too low, and 33% or so seems about fair.

In combining the outside and inside views, it's important to note that while the inside view is subject to a lot of the usual inferential uncertainty, the outside view is particularly flawed in this case – hence, I give equal weightage out of ignorance more than anything, which yields about a 41% chance of success.

Overall, therefore, the expected proportion of disease reduction from direct delivery is around 0.000003.

 

As for costing – note that we can normalize the one year operating cost of a direct delivery charity to USD 100,000, as it does not matter for the purposes of our calculations. After all, the proportion of disease reduction from direct delivery scales with expenditure, and so tractability does not change regardless.

 

And hence, the proportion of the problem solved per additional USD 100,000 spent is around 0.000003.

 

Marginal Expected Value of Distributing Mental Health Drugs

All in all, the marginal expected value of distributing mental health drugs is 28 DALYs per USD 100,000 spent, making this around 4% as cost-effective as a GiveWell top charity.


 

32

New Comment
8 comments, sorted by Click to highlight new comments since: Today at 7:04 AM

I am super interested in this problem, but finding it a bit tricky to follow your analysis - in part because it's so spread out and lots of the information seems buried seem in large paragraphs. Is there any chance you could put in it a spreadsheet?

For instance, I'm still unsure what your hypothetical intervention is - providing training to clinics to better diagnose mental health issues and then prescribe SSRIs, is that right?

Hi Michael,

The spreadsheet/CEA is here and is also linked at the very top of the post ("Considering the... etc").

Apologies for any lack of clarity in my write up - that's right, the intervention itself is providing training and the actual SSRIs to clinics (in LMICs).

Thanks for sharing this! I think your cost of drugs is significantly too high. The cited report says that total drug cost for those "initiating" on a generic SSRI/SNRI for 6 months was $174, but just because someone initiated on a generic SSRI/SNRI doesn't mean they stayed on one. Also, US retail drug pricing is just bizarre, even for generics, so you would be better off determining wholesale cost. For example, you can get a 90-day supply of fluoxetine 10 or 20 mg for $5.70 from Mark Cuban's pharmacy here, and 2/3 of that is cost is described as (U.S.) pharmacy labor and markup. Some are slightly higher . . . but as I understand it, there is no good evidence for any one SSRI being superior any another at a population level. Maybe it would be worth paying slightly more for escitalopram if you thought the side effect burden would be lower, although I don't remember whether that is actually true (vs. theoretical reasons you might expect it to be).

I agree with Jason's comment here. Not only would the market price for medication likely be much lower than you're assuming, but also aid organizations are often able to get drugs  donated or at below market prices from manufacturers. 

Thanks for the comment above on presentation, Jason - will keep that in mind!

On the issue of pricing - I think this you/MHR make good points here, and it did slip my mind that the US market (for which we have the most data) is unrepresentative (e.g. we certainly wouldn't want to use US insulin prices if we were examining diabetes interventions!).

My quick sense is that adjusting for this (as well as any bulk buying discount) puts the naive headline estimate into the GiveWell ballpark, but not enough to warrant deeper research if your expectation that further research is likely to cause estimated cost-effectiveness to drop even further anyway - as is the typically the experience of researchers (e.g. I think Eric Hausen had a good talk on this).

I think that makes sense. I think the value of making that adjustment is the move from "rather unlikely to be viable given that 0.03x is ~ 2 orders of magnitude away from the threshold for further research" to "this is not worth further pursuing now, but keep it in the back of your mind in case you happen across new information that would change the estimate in a moderately significant way / one can envision that there might be another intervention with synergistic effects that would sufficiently increase the benefits or reduce the costs of this one to consider packaging the two interventions together."

Two more suggestions for future reference:

It might be helpful to have a short section near the topin which you discuss significant sources of uncertainty. The top-line conclusion makes it clear to me that your analysis indicates that this is not an intervention we should be prioritizing at this time. However, listing factual and other assumptions that could significantly change the results in one place does two helpful things. 

First, if the reader thinks some of those assumptions may be incorrect, listing the uncertainties upfront invites the reader to keep reading rather than just accept the top-line conclusion of 0.04x GiveWell efficiency and conclude that it isn't worth investing 20 minutes in the rest of the report given that 0.04x isn't close to justifying further review. Second, your report becomes part of the EA body of knowledge, and hopefully someone will be conducting a review of past work every few years to see if something has changed. That person may not be you, and it would be helpful for them to see, at a glance, what inputs and assumptions you think have significant uncertainty and importance. Those are the inputs and assumptions that should be re-examined every few years to decide if this should go to a deeper analysis due to changed circumstances.

Footnotes or cross-references can save the reader time spent on details that they may not find helpful. For instance, you explain that you "discount for the probability of the world being destroyed anyway (i.e. general existential risk discount)" but don't mention what the discount is until the end of a very long paragraph. I would suggest stating in a short sentence that you are applying a general existential risk discount of 0.07 percent per annum and referring the reader to a footnote (or link to your website) for details.

At the shallow stage of research, a 0.07 percent adjustment is basically noise -- understanding the details of why you chose that specific figure is unlikely to help any potential user of your report, and if there are exceptions they can read the footnote or cross-reference. Moreover, anyone who already thinks existential risk is much higher is not going to be convinced by a one-paragraph analysis to the contrary (and the adjustment is already near-irrelevant to anyone who thinks you estimated too high).

Again, the approach could depend on who the target audience is -- and in some cases, you might decide it isn't worth expending more time editing the report. But I thought these comments were worth sharing for future possible reference.

I'm not saying the analysis is wrong. I'm just curious if the analyst has ever suffered from depression. Or had someone they love suffer from depression. 

It is easy to empathize with polio or malaria, but not as much depression. And when a cheap drug (as noted by other comments) can take one from suicidal to life worth living....