Strong upvote for multiple reasons: thoroughness and transparency of reasoning and execution, stating upfront that the cost-effectiveness was well below WHO’s threshold, prioritising reproducibility, and the “what this analysis doesn’t show” and “lessons learned” sections.
I wonder if “de-averaging the portfolio” by ballparking cost per impact by channel (school awareness packages vs workshops vs crisis hotline) might help guide resource allocation across channels (send more packets vs get more hotline volunteers etc). My naive guess is that suicide risk isn’t the same across channels, which the model as it stands implicitly assumes (which makes school awareness get ~10x as much impact credit as the crisis hotline); I’d assume that this risk for the school students is similar to that of the general population but that hotline callers are self-selected for being at much higher risk of suicide, so I’d explore the hypothesis that most of the bottomline suicide prevention impact comes from the crisis hotline even though most of the topline reach comes from the school awareness packages. This is also a channel attribution question, which as you’ve said the current model doesn’t show and is a hard one to answer.
The hotline is indeed considered the strongest channel at the moment at my nonprofit for suicide prevention impact, since people who call it do so of their own will and usually in a state of emergency. I would argue that the school workshops are very close to being effective since they also give an opportunity for a 1-on-1 conversation with a lifeline volunteer, albeit face-to-face. But I don’t know how to attribute the percentages here. The GiveWell cost-effectiveness article mentions moral weights, but they are more like philosophical moral judgments of questions about the person affected.
In data science for example the transformers are analyzing a sentence token by predicting the next token in line based on the weights, so if I follow the logical I need to assign cost effectiveness weights to decide how much each channel contributed to effectiveness. I wish there were more studies on this online.
Strong upvote for multiple reasons: thoroughness and transparency of reasoning and execution, stating upfront that the cost-effectiveness was well below WHO’s threshold, prioritising reproducibility, and the “what this analysis doesn’t show” and “lessons learned” sections.
I wonder if “de-averaging the portfolio” by ballparking cost per impact by channel (school awareness packages vs workshops vs crisis hotline) might help guide resource allocation across channels (send more packets vs get more hotline volunteers etc). My naive guess is that suicide risk isn’t the same across channels, which the model as it stands implicitly assumes (which makes school awareness get ~10x as much impact credit as the crisis hotline); I’d assume that this risk for the school students is similar to that of the general population but that hotline callers are self-selected for being at much higher risk of suicide, so I’d explore the hypothesis that most of the bottomline suicide prevention impact comes from the crisis hotline even though most of the topline reach comes from the school awareness packages. This is also a channel attribution question, which as you’ve said the current model doesn’t show and is a hard one to answer.
Thank you so much!
The hotline is indeed considered the strongest channel at the moment at my nonprofit for suicide prevention impact, since people who call it do so of their own will and usually in a state of emergency. I would argue that the school workshops are very close to being effective since they also give an opportunity for a 1-on-1 conversation with a lifeline volunteer, albeit face-to-face. But I don’t know how to attribute the percentages here. The GiveWell cost-effectiveness article mentions moral weights, but they are more like philosophical moral judgments of questions about the person affected.
https://www.givewell.org/how-we-work/our-criteria/cost-effectiveness/moral-weights
In data science for example the transformers are analyzing a sentence token by predicting the next token in line based on the weights, so if I follow the logical I need to assign cost effectiveness weights to decide how much each channel contributed to effectiveness. I wish there were more studies on this online.