I noticed that someone used my referral link to open a Capital One card and wanted to confirm for them that I’ve given $80 of the referral bonus to Malaria Consortium’s SMC program, on the understanding that it is currently GiveWell’s top recommendation.
HStencil
Credit Cards for EA Giving
[Question] Did Fortify Health receive $1 million from EA Funds?
[Question] What would “doing enough” to safeguard the long-term future look like?
Hey Max, thanks for all this explanation! One question: Has any thought been given to spinning off GWWC and EA Funds together (as a single organization), given that they share a similar focus, common users, and some degree of web integration?
Thanks for this great post! I’m curious whether you’ve looked into any of the other developing world COVID-19 initiatives for which The Life You Can Save is currently raising money (beyond Development Media International and GiveDirectly). These include programs by TLYCS top charities D-Rev, Evidence Action, Living Goods, Population Services International, and Project Healthy Children, several of which are also, as you know, highly regarded by GiveWell.
I’d also be curious about whether you’ve looked into the COVID-19 Early Treatment Fund’s work sponsoring outpatient trials of promising anti-virals as early treatments for COVID-19. Marc Lipsitch spoke favorably of its work in his recent interview on the 80,000 Hours podcast, and in a number of respects, it strikes me as similarly promising to Fast Grants.
Thanks so much! This resource has been extremely useful.
Great! Thanks so much for flagging that here! I assume this means that you consider Oxfam, PSI, DMI, and GiveDirectly to be more promising giving opportunities than the COVID-19 response programs of other TLYCS charities, like Living Goods, Project Healthy Children, etc. — is that right?
Thanks — that makes perfect sense!
Because you’re recommending Village Enterprise, I’d also flag TechnoServe, which runs similar programs and is the top-rated poverty alleviation charity by ImpactMatters. It’s worth noting that ImpactMatters only evaluated (I believe) TechnoServe’s Impulsa Tu Empresa program, which operates exclusively in Latin America, but the organization runs analogous programs in Sub-Saharan Africa. Obviously, those might not be similarly cost-effective, but a prima facie basis for making that assumption (rather than the opposite, for instance) isn’t obvious to me.
I think the right kind of feedback here depends mainly on whether you mean to propose that EA underestimates the extent to which treating people with dignity improves their welfare or you mean to propose that EA fails to consider the importance of dignity as an intrinsically and independently valuable element of a life lived well. If dignity is only important on account of its instrumental role in improving welfare, I very much doubt that a thorough evaluation of that role would lead many EAs to conclude that they should redirect their charitable giving. Even if treating someone with dignity were associated with a truly striking increase in their welfare, it seems unlikely to me that, for instance, global health interventions with such an emphasis would outperform distributing insecticide-treated mosquito nets or deworming pills. Among other things, I imagine that most parents of young children would agree to an arbitrarily large amount of undignified treatment in exchange for preventing their child from dying of malaria (revealed preferences suggest this is true). This suggests that the AMF would outperform a hypothetical charity with a dignity focus even without accounting for the positive impact of saving children’s lives unless promoting dignity were extraordinarily cheap per person affected. Similarly, I doubt that integrating concern for the role dignity may play in determining welfare into longtermist perspectives would do much to shift people’s ideas about the best giving opportunities to safeguard the long-term future of humanity.
If on the other hand you take dignity to be valuable in itself (apart from any role it might have in bringing about another good, like improving welfare), I wonder whether the philosophical foundation for your view is really fully compatible with EA. From what I’ve read, it seems as if most of the philosophers who take treating people with respect to be a good in itself view dignity as the sort of thing that each of us has a reason to accord to others when we interact with them. They do not, however, by and large view dignity as the sort of thing that we have a reason to impartially maximize (i.e. while it’s very important for me to treat you with dignity, it’s nowhere near as important—and may not even be valuable at all—for me to counterfactually enable you to treat someone else with dignity). In their view, the obligation to treat others with dignity “spring[s] from an agent’s special relationship to his own actions” and “the claims of those with whom we interact to be treated by us in certain ways” (Korsgaard 1993, emphasis mine), not from the objective value of the world having more dignity in it (or anything like that). As a result, some (see, for example, Taurek 1977) go so far as to argue that it is not necessarily any better for more people to be treated well than for fewer. Following Korsgaard, we might think of the value of treating people with dignity as similar to the value of keeping promises — while I have reason to keep my own promises, I likely do not have reason to promote a world in which more promises are kept. Doing so would suggest that I misunderstood the way in which keeping promises is valuable. If dignity is the kind of moral good that most clearly has a place in non-consequentialist moral views that oppose interpersonal aggregation wholesale, I suspect that at least our present philosophical concept of it may be unsuited to sit among what we might conventionally refer to as “EA values.”
That said, I should note: Like surprisingly many EAs, I am not a utilitarian. I am, however, some kind of consequentialist, and I would love for EA folks to invest more effort in developing a thorough conception of human flourishing, of what it means for a person’s life to go well for them. Without such a theory, we cannot ensure that we are actually improving others’ lives to the greatest extent possible (because we lack a robust understanding of what it means for a life to be improved). For that reason, I personally welcome posts like this that seek to draw attention in those kinds of directions and propose some less conventional ideas about what flourishing might involve.
[UPDATED June 30, 10:00 pm EDT to reflect substantial improvements to the statistical approach and corresponding changes to the results]
I spent some time this weekend looking into the impact of COVID-19 on the 2020 U.S. presidential election, and I figured I might share some preliminary analysis here. I used data from The COVID Tracking Project and polling compiled by FiveThirtyEight to assemble a time series of Biden’s support head-to-head against Trump in 41 states (all those with adequate polling), along with corresponding COVID-19 data. I then implemented a collection of panel models in R evaluating the relationship between Biden’s performance against Trump in state polling and the severity of the pandemic in each state. My data, code, and regression output are on GitHub, and I’ve included some interpretive commentary below.
Interpretation of Results
When appropriately controlling for state-level fixed effects, time fixed effects, and heteroscedasticity, total COVID-19 cases and deaths are not significantly associated with support for Biden, nor are the number of ongoing COVID-19 hospitalizations (see Models A, B, and C in the 6.30 R script). However, controlling for state-level fixed effects, greater daily increases in cases and in deaths are significantly associated with higher support for Biden (see Models 2 and 3 in Table I above). Breusch-Pagan tests indicate that we must also control for heteroscedasticity in those models, and when we do so, the results remain significant (see Models 2 and 3 in Table II above), though only at a 90 percent confidence level.
These results do come with a caveat. While Lagrange FF multiplier tests indicate that there is no need to control for time fixed effects in Table I models 2 and 3, F-tests suggest the opposite conclusion. I lack the statistical acumen to know what to make of this, but it’s worth noting because when you control for both time fixed effects and heteroscedasticity, the results cease to be statistically significant, even at a 90 percent confidence level.
Interestingly, the state-level fixed effects identified by Table I models 2 and 3 are strikingly powerful predictors of support for Biden everwhere except for Arkansas, Florida, Georgia, North Carolina, Nevada, Ohio, and Texas, all of which (except for Arkansas) are currently considered general election toss-ups by RealClearPolitics. This makes sense — in a society as riven by political polarization as ours, you wouldn’t necessarily expect the impacts of the present pandemic to substantially shift political sympathies on the aggregate level in most states. The few exceptions where this seems more plausible would, of course, be swing states. In the case of Arkansas, the weak fixed effect identified by the models is likely attributable to the inadequacy of our data on the state.
A Hausman test indicates that when regressing the amount of COVID-19 tests performed in a state on the support for Biden there, a random effects model is more appropriate than a fixed effects model (because the state-level “fixed effects” identified are uncorrelated with the amount of tests performed in each state). Implementing this regression and controlling for heteroscedasticity yields the result featured under Model 1 in Table II above: statistically significant at a 99 percent confidence level.
This is striking, both for the significance of the relationship and for how counterintuitive the result of the Hausman test is. One would assume that the fixed effects identified by a model like this one would basically reflect a state’s preexisting, “fundamental” partisan bent, and my prior had been that the more liberal a state was, the more testing it was doing. If that were true, one would expect the Hausman test to favor a fixed effects model over a random effects model. However, it turns out that my prior wasn’t quite right. A simple OLS regression of states’ post-2016 Cook Partisan Voting Indexes on the amount of testing they had done as of June 26 (controlling for COVID-19 deaths as of that date and population) reveals no statistically significant relationship between leftist politics and tests performed (see Model 1 in Table III), and this result persists even when the controls for population and deaths are eliminated (see Model 2 in Table III).
This is odd. Hausman tests on Models A, B, and C in the 6.30 R script favor fixed effects models over random effects models, indicating that state-level fixed effects (i.e. each state’s underlying politics) are correlated with COVID-19 cases, hospitalizations, and deaths, but those same fixed effects are not correlated with COVID-19 tests. Moreover, when applying appropriate controls (e.g. for heteroscedasticity, time fixed effects, etc.), we find that while cases, hospitalizations, and deaths are not associated with support for Biden, testing is associated with support for Biden (basically the opposite of what I would have expected, under the circumstances). We can run a Breusch-Pagan Lagrange multiplier test on Table I’s Model 1 just to confirm for sure that a random effects model is appropriate (as opposed to an OLS regression), and it is. At that point, we are left with the question of what those random effects are that are associated with support for Biden but not with COVID-19 testing, as well as it’s corollary: Why aren’t the fixed effects in Models A, B, and C associated with testing (given that they are associated with cases, hospitalizations, and deaths)? Without the answers to these questions, it’s hard to know what to make of the robust association between total COVID-19 testing and support for Biden revealed by Table I’s Model 1.
The puzzling nature of the Table I, Model 1 results might incline some to dismiss the regression as somehow erroneous. I think jumping to that conclusion would be ill-advised. Among other things that speak in its favor, the random effects identified by the model, however mysterious, are remarkably consistent with common intuitions about partisanship at the state level, even more so, in fact, than the fixed effects identified by Models 2 and 3 in Table I. Unlike those fixed effects models, Model 1′s state-level random effects explain a considerable amount of Biden’s support in Georgia and Texas. I consider this a virtue of the model because Georgia and Texas have not been considered swing states in any other recent U.S. presidential elections. They are typically quite safe for the Republican candidate. Furthermore, Model 1 identifies particularly weak random effects in a few swing states not picked up by Models 2 and 3 — notably, Wisconsin, Pennsylvania, and Arizona. Wisconsin and Pennsylvania are genuine swing states: They went blue in 2008 and 2012 before going red in 2016. Arizona has been more consistently red over the last 20 years, but the signs of changing tides there are clear and abundant. Most notably, the state elected Kyrsten Sinema, an openly bisexual, female Democrat, to the Senate in 2018 to fill the seat vacated by Jeff Flake, who is none of those things.
It’s worth noting that there is an extent to which the above is really an oversimplification of “swinginess.” As FiveThirtyEight explains, state-level opinion elasticity is not the same as being a swing state. While how close a state’s elections are is determined by the proportion of Democrat relative to Republican voters in the state (with states closer to 50⁄50 obviously being “swingier” in this sense), the extent to which events out in the world lead to shifts in polling in a given state is determined largely by how many people in the state are uncommitted to a particular partisan camp. A state being full of such voters without strong partisan commitments might well express itself in close elections, but it also might not, and by the same token, another way a state might end up with close elections is by being 50 percent composed of die-hard, party-line Republicans and 50 percent composed of die-hard, party-line Democrats. We would not expect new developments in current events to particularly shift voter sentiment in such a state. As a result, FiveThirtyEight proposes the metric of state-level opinion elasticity—measured using the extent to which shifts in national polling correspond to shifts in state-level polling in each state—as an alternative concept of “swinginess” that is potentially more appropriate for analyses such as this one. This is important because FiveThirtyEight has found that a number of states where there are frequently close elections actually exhibit extremely low elasticity. The paradigm case of this is Georgia, which has an elasticity of 0.84 (meaning a one-point shift in national polling corresponds to a 0.84-point shift in Georgia polling).
On this basis, a better way of comparing the fixed effects identified by Table I models 2 and 3 with the random effects identified by Model 1 would be to compare the average elasticities (calculated by FiveThirtyEight) of the “swing states” identified by each model. In the case of Models 2 and 3, the average is an uninspiring 0.994, or weakly un-swing-y. In the case of Model 1, however, the average is 1.025: pretty swing-y. That’s what happens when you swap out Georgia (0.84) for Arizona (1.07) and Wisconsin (1.06).
I have one outstanding technical question about Table I, Model 1. When controlling for heteroscedasticity in R’s plm package, I understand that the “arellano” covariance estimator is generally preferred for data exhibiting both heteroscedasticity and serial correlation but is generally not preferred for random effects models (as opposed to fixed effects models). The “white” covariance estimators, on the other hand, are preferred for random effects models, though “white1” should not be used to control for heteroscedasticity in data that also exhibit serial correlation. A Breusch-Godfrey test indicates that Model 1 requires an estimator compatible with serial correlation, but it is of course a random effects model, not a fixed effects model. Would it be better to control for heteroscedasticity here with “white2“ or with “arellano?” Ultimately, it doesn’t matter much because both approaches yield a result that is statistically signicant at at least a 95 percent confidence level, but only the “white2” estimator is statistically significant at a 99 percent confidence level.
Thanks so much for looking into this and posting your read of the research! I’m glad I now have a clearer sense of how these two types of interventions compare to one another. The flaws you noted in TechnoServe’s internal evaluation are certainly quite concerning, and I’m glad someone brought them to my attention.
As someone who first encountered EA through Slate Star Codex, this is also my sense.
Thanks so much! I’m thrilled to hear you liked it. To be honest, my main reservation about doing anything non-anonymous with it is that I’m acutely aware of the difficulty of doing statistical analysis well and, more importantly, of being able to tell when you haven’t done statistical analysis well. I worry that my intro-y, undergrad coursework in stats didn’t give me the tools necessary to be able to pick up on the ways in which this might be wrong. That’s part of why I thought posting it here as a shortform would be a good first step. In that spirit, if anyone sees anything here that looks wrong to them, please do let me know!
Thank you so much for putting so much thought into this and writing up all of that advice! Your uncertainties and hesitations about the stats itself are essentially the same as my own. Last night, I passed this around to a few people who know marginally more about stats than I do, and they suggested some further robustness checks that they thought would be appropriate. I spent a bunch of time today implementing those suggestions, identifying problems with my previous work, and re-doing that work differently. In the process, I think I significantly improved my understanding of the right (or at least good) way to approach this analysis. I did, however, end up with a quite different (and less straightforward) set of conclusions than I had yesterday. I’ve updated the GitHub repository to reflect the current state of the project, and I will likely update the shortform post in a few minutes, too. Now that I think the analysis is in much better shape (and, frankly, that you’ve encouraged me), I am more seriously entertaining the idea of trying to get in touch with someone who might be able to explore it further. I think it would be fun chat about this, so I’ll probably book a time on your Calendly soon. Thanks again for all your help!
Thanks! I booked a slot on your Calendly—looking forward to speaking Thursday (assuming that still works)!
At least until quite recently, there was a fairly uniform consensus in mainstream Anglo-American economics that the convergence thesis was true. I think this was mainly because it was based on fundamental theoretical insights that were believed to be relatively unimpeachable, like the Solow Model and the Stolper-Samuelson Theorem.
The Solow Model uses a formal representation of the idea that capital can be put to better use (yielding a higher economic return) in places where it is more scarce to demonstrate that, all other things being equal, places further from a given steady-state output level will grow toward that level faster than places nearer to it. In other words, ceteris paribus, places where capital stock is lower will grow faster than places where capital stock is higher because adding a marginal unit of capital in a capital-poor economy will generate a greater return than adding a marginal unit of capital in a capital-rich economy, where all the high-yielding capital investment opportunities have already been funded. (Bear in mind, though, that “ceteris paribus” is doing a lot of work in that sentence. You might reasonably claim that the traditional Solow Model holds constant nearly everything we ought to care about in trying to explain development outcomes.) To the extent that it’s true, though, in a world with open cross-border capital flows, one would expect capital to flood from low-return investment opportunities in wealthier countries to high-return investment opportunities in poorer countries. Alas, the evidence that this is actually taking place on a large scale is mixed at best, and other factors excluded from the neoclassical theories of international trade and finance likely play a large role in determining the global allocation of capital.
The productivity term in the Solow Model also often comes up in discussions of convergence. This term, representing an economy’s efficiency at deploying its factors of production to make things, is frequently treated—for the purpose of simplification—as a representation of an economy’s level of technological advancement alone. Traditional growth economists tend to treat rates of technological advancement as largely exogenous (whether this assumption is realistic is the subject of considerable debate). However, separate models of global technological advancement are typically built around the idea that it’s cheaper to copy a technology that was developed in another country and put it to use in one’s domestic industries than it is to develop a wholly new technology from scratch, thereby advancing the technological frontier. As a result, economists often conclude that countries not yet at the technological frontier will enjoy faster productivity growth than counties that are at the technological frontier, in accordance with the convergence paradigm.
The Stolper-Samuelson Theorem shows that when a national economy specializes in the production of a good in which it has a comparative advantage and then the relative price of that good rises on global markets, the return on investment in the factor of production that most contributes to making that good will rise. For example, if a country has a comparative advantage in making blue jeans, and it specializes in making blue jeans, and labor is the most important factor of production in making blue jeans, if the relative price of blue jeans on globals markets rises, then the return on investment in labor in that country will rise. This is equivalent to saying that the marginal product of labor in that country will rise, and in a competitive labor market, the price of labor (the wage) should equal its marginal product, so producer wages should rise with, for instance, a relative increase in global demand for blue jeans (which would push up the price).
There is vigorous debate over the extent to which the Stolper-Samuelson Theorem is applicable to world in which we live today. It requires making a number of assumptions in order for its conclusion to hold (constant returns to scale, perfect competition, an equal number of factors and products). One famous counterexample to Stolper-Samuelson was proposed Raúl Prebisch and Hans Singer and was embraced by the anti-trade left of the postwar years. Prebisch and Singer propose that because complex manufactured goods (like computers) exhibit greater income elasticity of demand than simple commodities (like wheat or coffee), if a country specializes in exporting wheat (consistent with its comparative advantage), and relies on imports from foreign manufacturers to get computers, as global incomes rise, it will suffer declining terms of trade (i.e. as time passes, each imported computer will cost more and more wheat). Today, the Prebisch-Singer Hypothesis, as it’s called, has received some degree of very qualified acceptance by mainstream economists. Its fundamental proposal that it doesn’t always make sense to treat comparative advantages as destiny is quite widely accepted, though more on the basis of Paul Krugman’s work in New Trade Theory (demonstrating, e.g., that comparative advantages can arise from economies of scale in addition to from initial actor endowments) than on the basis of Prebisch and Singer’s work. However, the specifics of the hypothesis are regarded as an extremely special case, an exception to what is generally true of developing countries. There are two main reasons for this. The first is that many developing countries specialize in the extraction of metals and minerals that are necessary inputs in making complex manufactured goods, like copper and silicon. These commodities likely violate Prebisch-Singer’s assumption that simple commodity goods necessarily exhibit lower income elasticity of demand than complex manufactured goods. The second reason is that many of the complex manufactured goods that the poorest countries import from wealthier countries actually probably increase those countries’ productivity in producing basic commodities (consider, for instance, the way organizations like Precision Agriculture for Development deliver scientific agricultural guidance to farmers throughout South Asia and Subsaharan Africa via their cell phones).
I’m not sure to what extent this theoretical background will be helpful to you as you think about convergence, but regarding the facts on the ground, with very few exceptions (like Botswana), almost all of the progress toward convergence in the last four decades has taken place in East Asia. While the “Asian Miracle” is very much real, it may itself prove to be a special case, specific to the region or the historical period in which it took place. As premature deindustrialization begins to take its toll on those countries that are not yet rich, there are, I think, a number of serious concerns about the continued viability of the export-led growth models that lifted countries like South Korea and Japan out of poverty. While the theoretical insights on which those models were based are robust, it remains to be seen to what extent they continue to apply in our 21st-century economy. Similarly, the traditional convergence thesis assumes increasing liberalization of international trade and capital flows, a premise that has grown increasingly untenable over the last five years.
Oops, sorry—not sure what happened to the Chase referral link, but I edited the post to add it. BofA isn’t running a referral program right now, so there isn’t a link for that card.
All of those flat-rate cards at the bottom should return 2% (or 1.8% in the case of Citizens, 1.5% in the case of Capital One, etc.) on payments through Facebook Fundraisers, given that their fixed cash back rates apply to all transactions. I can’t personally confirm that, not owning any of the 2% cards myself, but assuming they follow their policies, you should be just fine.