Estimation is the best we have

Developing intuition- Why it isimportant?

Many great thinkers have emphasizedthe significance of intuition along with its great impacts on theirprofessional and personal lives. Many define it as ‘a priori’ knowledge. It isindispensable and essential tool for us.

An assumption behind this entire post is that X somehow fits into some linear quantitative scale.  The problem is that most people’s conception of value or good doesn’t do this.  The examples you use all have concrete linear quantitative outcomes (distance, stock price, etc.).  People get angry with this approach because of the neccessity to collapse everything down to a single metric, which ignores the complexity of the problem.  Is it really possible to collapse all good in the world down to a single metric? 

Disclaimer: I actually agree with you, but I have many friends who do not, and who make arguments such as this one so I’m interested in how you respond.  There are ways out, for example by forcing trade-offs and thus forcing a collapse onto single metrics, but I’d be interested nonetheless in how you justify this methodology being applied to hugely multi-dimensional problems such as ‘is policy x better than policy y’.

Hi,

I suppose I respond by ‘forcing trade offs’, but since I know of no way to avoid trade offs, this doesn’t seem an unnatural move. To me, to say that you value something is to say that you have a certain inclination to choose it over other things. I’m not sure what they have in mind.

So if we want to know how prepared to be for a nuclear war for instance, we should do something qualitative with scenarios and the like.

That’s an interesting example to think about. Why does scenario planning feel more useful than estimating the probabilities of various scenarios happening?

Well, the information we have is immensely biased and noisy, so the probabilities just aren’t very good. Secondly, and very importantly, the return on investment on getting better estimates is pretty low; “the risks are non-negligeable so we should do something about them because nuclear war is so horrid” is already enough to justify a lot of action; fine tuning the estimates seems to give little extra return. Thirdly, we might get too obsessed with the details of the numbers, and forget the uncertainty (even if we don’t have that failing, our institutions might). Related to this is worry about “unknown unknowns”: we have reason to suspect, based on based experience, that events in a nuclear war will turn out to be unpredictable in ways we can’t expect now. So we should build up general resilience, rather than refining probability estimates over an incomplete set of possible outcomes.

Nice post, and high-quality comments.

I totally agree that doing some impact estimation is critical, but I also think that it’s important that we are aware of its limits, and take it with a grain of salt.

There’s a fantastic, rigorous critique of relying solely on “Expected Value” estimates of impact of charitable donations here, and I think the argument cross-applies for career planning: http://blog.givewell.org/2011/08/18/why-we-cant-take-expected-value-estimates-literally-even-when-theyre-unbiased/

In addition, I know that I personally fall a little subject to “analysis paralysis” on this kind of issue, and often do underestimate the power of “unknown unknowns” (as mentioned by Stewart) or other shortcomings in my analysis. I suspect that others reading this blog might be of a similar disposition and could be prone to similar mistakes.

But what strategy should you use instead?

In 80k strategy, we often consider applying some weight to common-sense, since in some cases it might actually be optimised, or reflect some of things you miss if you’re too analytic.

You could also look at ‘meta’-strategies, like being flexible, so that you can continually update as you learn more.

It’s a good point, Katja.

Another approach quite popular in cost-benefit work is to suppose something like this:

Say you have a policy goal regarding the outcome of a highly complex system, about which you know something but nowhere near everything. First, you must determine is some defined loss value associated with the goal (like a statistical life value, etc.). Normally, you would model both the losses in your simplified model of the base case and policy case and compare the discounted differences. But this opens itself up to the sorts of criticism you spell out.

Instead of comparing one hypothetical projection of the expected loss against another (which includes the expected changes due to the policy), another approach is to look at the tuples of potential changes due to the policy and the probabilities of loss for the point at which spending breaks even. People seem better equipped to see whether these are plausible values than they are at determining potential outcomes of very complicated systems.

Interesting post Katja and interesting follow-up comments. I agree that trade-offs must come into it and in this sense all quantitative and qualitative information are effectively evaluated on a single metric (given that only one choice or decision can be made at once). As a perhaps slightly off topic aside, I have been studying cost-benefit analysis recently and came across the work of Jack Dowie who you may be interested in as he provides an (for me) illuminating framework for conceptualising how real world decisions are made with reference to different categories of belief (evidence) and preferences (values) ranging on a scale of intuition to analysis ratios. His paper is still being finalised but you should be able to find a presentation on his conceptual map called JUDEMAKIA via Google.

Statistically speaking, a good solution to these kinds of problems is to use the Bayesian approach of statistics to model uncertainty. The Bayesian approach involves both, quantitative, objective data and qualitative, subjective expert knowledge – this approach of modelling uncertainty is often termed expert system and was particular popular in the 1980s.

It can be quite powerful: Combining quantitative medical data with subjective expert knowledge for example lead to a model that could predict a disease from symptoms better than experienced medical doctors (@Benjamin: such models would also allow to integrate new information quite easily). However, as you already mentioned in your post, such models are quite difficult and expensive to construct.

Models that use newer algorithms are much easier to construct and yield even better, more accurate predictions – however they need large amounts of quantitative data and cannot deal with semantic data yet. But I am quite sure that this is how these problems will be addressed and solved in the near future.