Opportunities to contribute to this project
Here we list ways that you can contribute if you wish to this project.
Trialing Excel alternatives for GiveWell CEAs
One of the first steps in this project is to trial out using different interfaces for CEAs. This could help determine whether different interfaces could be valuable for representation.
Currently, Sam Nolan has created a primitive representation of the GiveDirectly model inside Causal:
There are a number of limitations to the above, that could use improvement, notably:
Formatting: The documentation of the parameters only fits on one line and takes up too much horizontal space, this can be fixed by using Freeform charts.
Calibrating the uncertainty: The uncertainty on all the parameters is guessed quite vaguely, and could be improved with careful consideration
Connecting this to empirical measures of uncertainty (challenging), and to previous and related work
Considering the plausibility of features of the implied distributions
e.g., something like 'this distribution implies a 10% chance of an average household size of 10 or more among those getting GiveDirectly direct transfers; this seems too high, as only 0.5% of the poorest households in Sub-Saharan have more than 7 members)
Further interventions for comparison: It would be great to also include/try creating AMF and DtW versions of this,
to show rich comparisons between different interventions,
to show that this can be extended to more intricate models involving health outcomes.
Testing other tools: In other work, it would be great if another representation could be done in Guesstimate for comparison (and perhaps code-based tools as well)
Improving Pedant
Pedant offers an exploration of code like languages to represent CEAs. If you wish to help develop Pedant, look at open issues on GitHub, as well as get in contact with Sam Nolan on the EA Forum, Discord, or the EA Public Interest Technologists Slack. I am currently looking for people to transcribe other models, as well as Haskell Developers.
Guessing uncertainty in GiveWell estimates
Currently, as far as I am aware, no one is yet to attempt to properly guess the amount of uncertainty in each of the parameters in GiveWell's model. I believe cole_haus modelingis the closest thing to it. Representing a GiveWell CEA (or several) with uncertainty bounds in any form would be valuable.
Last updated
