Outside View as the Main Debiasing Technique

post by abramdemski · 2017-10-16T21:53:10.722Z · score: 19 (4 votes) · LW · GW · 4 comments

The inside view refers to what your explicit models tell you. For example, if you write down a plan with explicit contingencies for things going wrong and time estimates for each step, or write down a careful argument for your point of view, or solve some equations. The outside view refers to what "an independent observer" might think; the proverbial man-on-the-street; a perspective which doesn't pay attention to all the details of the situation, but places it as an instance of a class. For a plan, you have some outside-view probability that it will fail or take longer than you expect, based on unforeseen complications rather than explicitly-listed contingencies. For an argument, although the argument may conclude a belief with some high probability, your outside-view probability should usually be lower to account for your uncertainty that the argument is correct. For the equations, the outside view accounts for the probability that you've made a mathematical error.

One way of thinking about inside view vs outside view is that they are human-friendly ways of thinking about the likelihood function and the prior. Somehow, humans are frequently tempted to commit base-rate neglect: the belief in a hypothesis is just the extent to which it matches the data. For example, if a test for a rare disease comes up positive. People are not inclined by default to multiply the evidence for something by the prior odds against, except in everyday circumstances which they've experienced a lot. (I think this mistake may also account for the conjunction fallacy.) "Outside view" helps you imagine base rates by envisioning a perspective with more distance from the issue. So, you come up with an inside view and an outside view and "multiply them together" to become a proper Bayesian. This gets rid of base-rate neglect, and related biases such as the availability heuristic (where you implicitly use "number of times I've heard about this" or similar things as a substitute for base-rates).

However, I think this underestimates the usefulness of outside view. Robin Hanson discusses it almost as if it is equivalent to eliminating bias. While I do think a strong outside view needs to be combined with a strong inside view, I think there's a sense in which outside view is the first thing you reach for to debias almost anything, not just base-rate neglect.

Say you learn about a new bias, such as omission bias, and you don't have any special debiasing tools for it. What you do is you file it away under your list of important biases. Then, when you encounter a situation which pattern-matches to the bias, you think something like "Ahh, but omission bias!" and attempt to adjust for it.

In other words, you have some kind of quick pattern-matcher which looks out for signs of the bias. When it fires, you may do a more thorough check: are you in the kind of situation where the bias occurs? If so, you attempt to correct for the bias. This is outside-view reasoning; you are noticing that the situation has the general flavor of one which gives rise to bias, and attempting to follow the general policy which would correct for that bias, rather than the biased policy which you would follow by default.

If all goes well, jumping to outside view helps you subsequently develop more targeted debiasing strategies which hit the bias directly, including an inside view which doesn't produce the bias in the first place. However, I want to give outside view credit for being the quick skill that you apply all over the place which helps to generate the further skills.


Comments sorted by top scores.

comment by G Gordon Worley III (gworley) · 2017-10-16T23:30:37.055Z · score: 14 (3 votes) · LW · GW

I like the outside view a lot and mostly have nice things to say about it. But that wouldn't make a very interesting comment, so instead I'll say something about its shortcomings.

There is a sense in which the outside view doesn't exist. By this I mean any view must be the view from somewhere in the observable universe and thus is always an inside view of what we can experience. The only outside view in this sense is the one that's impossible to take because physical constraints prohibit us from getting to a vantage point where we could see it, although we can make theoretical approximations of what those views might be like (cf. Tegmark).

Luckily this is not much of a practical concern since the outside view we usually want is one outside a particular local context, such as the view from outside a social group or individual decision rather than the view from outside your lifeworld. And even the view from outside the lifeworld turns out to be, even if theoretically impossible, practically possible up to the limit of our bounded ability to construct an ontology of our lifeworld.

comment by Vaniver · 2017-10-16T23:44:23.540Z · score: 23 (5 votes) · LW · GW

I'm surprised that you jump to "the outside view doesn't exist in full purity" and not the more obvious shortcoming of "one has to use inside-view models to slot objects into reference classes before one can apply reference-class reasoning, hence the outside view always contains the inside view as a critical step."

comment by Conor Moreton · 2017-10-16T22:21:07.632Z · score: 12 (2 votes) · LW · GW

Loren ipsum

comment by John_Maxwell_IV · 2017-10-17T19:17:31.742Z · score: 3 (0 votes) · LW · GW

I'm not convinced that inside view vs outside view is a well-defined distinction. You talk about the prior vs the evidence, but the process of reasoning typically involves taking some prior and then doing a bunch of separate updates on the evidence you see. Without seeing any evidence at all, my prior that any given statement is true is 50%. At what point in my process of updating on the evidence I see do I stop and say "OK, this is the outside view, all further evidence is inside view evidence"? And how do I order evidence so that "outside view" evidence comes before this point and "inside view" evidence comes after?

Perhaps the process of noticing that you might be susceptible to a bias in a particular situation should be considered logically distinct from the prior vs evidence distinction you make earlier. Well, I suppose you could argue that outside view reasoning amounts to "deleting" certain evidence that might be susceptible to a bias? (E.g. if I'm trying to estimate how long a project is going to take: Given the planning fallacy, maybe I should choose not to update on my intuitive feeling about how long the project is going to take.) It doesn't feel to me as though this definition captures everything that people are talking about when they say "outside view", though. Maybe splitting it in to multiple terms would be useful.