Lesson 1: When precision fails — hits-based giving

The interventions GiveWell recommends share a common structure: deliver something (nets, drugs, cash), measure a clear outcome (deaths averted, income increased), compare cost per outcome. This works brilliantly for direct health interventions.

Now consider a different kind of grant: funding a nonprofit to run corporate campaigns for better treatment of chickens raised for meat. The causal chain is long and messy: the nonprofit pressures companies → companies change policies → farms change practices → chickens suffer less. Each arrow involves uncertainty. At the end, you’re trying to quantify something that can’t be directly measured.

Or funding a think tank to advocate for immigration policy reform. If successful, millions of people could move to higher-income countries, dramatically improving their lives. But “if successful” is doing enormous work. The political process is chaotic. Attribution is nearly impossible. The time horizon is years or decades.

Should we avoid funding such work because we can’t measure it precisely? Coefficient Giving (formerly Open Philanthropy) thinks not. Their approach — “hits-based giving” — isn’t a departure from cost-effectiveness reasoning but an extension of it: you still build the model, you still estimate expected value, you just accept that most individual grants will fail and a few could be transformative. Like venture capital, you expect most bets to return nothing, but the winners pay for the losses many times over.

It’s worth knowing where Coefficient Giving comes from. It grew out of GiveWell: Holden Karnofsky and the team spun it off (originally as the Open Philanthropy Project) to fund work that doesn’t fit GiveWell’s cost-per-life-saved mould — policy advocacy, scientific research, institutional change. Same people, same underlying logic, broader mandate. The BOTEC you’ll meet in Lesson 2 has the same shape as a GiveWell CEA. A standard CEA already rests on plenty of judgement calls — on moral weights, counterfactual impact, how much to trust any given study. The difference with a hits-based BOTEC is how many more of those calls there are, and how much wider the error bars end up.

Read: Hits-Based Giving — Coefficient Giving (20 mins)

As you read, notice: what’s the same as a GiveWell-style cost-effectiveness estimate, and what’s different? What gets built into the argument that’s harder to pin down than a bednet’s cost per life saved? Where do they explicitly accept that they might be wrong?

Accepting high variance in outcomes is the price of admission. Most hits-based grants will produce little or nothing. A few will succeed massively. If you evaluated each grant individually after the fact, most would look like failures. But the portfolio as a whole can outperform a lower-variance strategy — if the hits are big enough.

So: how do you decide which speculative bets to make? You can’t run an RCT on policy advocacy. You can’t randomise countries into treatment and control groups for immigration reform. Coefficient’s answer is to make your reasoning explicit — build a model with your best guesses for each uncertain parameter, publish it, and let others tell you where they disagree.

The rest of the week teaches you how to do that.

We’ll introduce a running example shortly: Feather Forward, a fictional nonprofit running corporate campaigns for broiler chicken welfare.