Personal vs. institutional decision analysis

1. The tension

• Subjective Bayes story: pick a prior that represents your beliefs, pick a utility that represents your preferences, update with data, choose the action that maximizes subjective expected utility.
• Problem: for personal decisions, that’s often overkill or even backwards. If you already kind of know what you want to do, you can always “tune” the prior or utility to get that answer → feels circular.

So for purely personal choices, formal decision analysis is only really helpful when:

• the utilities are pretty clear (e.g. years of life),
• the probabilities are pretty well sourced (e.g. medical literature),
• and you genuinely want to know: “is this extra test/treatment worth it?”

That’s why their bronchoscopy example works nicely: probabilities and utilities are well-defined, so the math clarifies that the test doesn’t help.

---

2. Institutional decision analysis

In organizations (gov’t, business, research labs), it’s different:

• Decisions must be defensible and transparent.
• You want to show: “Given these assumptions about costs and risks, and these data, this is the option with the best expected payoff.”
• That’s what they call institutional decision analysis.

It’s not about “my personal utility”; it’s about:

1. write down the probability model,
2. write down the utility/cost structure the institution is willing to use,
3. show the implied recommendations,
4. and show how sensitive the answer is if you tweak key assumptions.

So Bayes here is a clarifying tool.

---

3. How it applies to their three examples

1. Survey incentives
   • Data are noisy, lots of uncertainty.
   • They don’t pick one “optimal” incentive.
   • Instead they say: “Here is the curve: X \$ → Y% more response → Z net cost.”
   • That’s useful for an institution because it makes the trade-off explicit, even though utilities aren’t nailed down.
2. Medical screening
   • Utilities and probabilities are clearer.
   • So you can actually say: “Don’t do bronchoscopy here.”
   • Also you can open the box and ask: “What if test mortality were 2% not 5%?” → sensitivity analysis.
3. Radon
   • The iffy part is the cost–risk tradeoff (how many \$ per microlife? what action level R_action?).
   • Once you pick that (say R_action = 4 pCi/L), the model can produce concrete, map-level recommendations.
   • Are those everyone’s personal utilities? Of course not.
   • But for a government wanting to issue a uniform policy, it’s very helpful.

So: personal Bayes = only good when inputs are real; institutional Bayes = make inputs explicit so the decision is auditable.

---

4. The “objective/institutional” Bayesian stance

Their closing philosophy:

• Be explicit about:
  • model assumptions,
  • data sources,
  • utility/cost choices.
• Check that the model’s predictions and recommendations fit reality.
• Then, if needed, expand the model to include more real-world info (more covariates, better measurement model, heterogeneous utilities).
• The expansions that help most are the ones that let you use more relevant information.