Plausibility bias? You say that as though that were a bad thing | Hacker News Reader

Plausibility bias? You say that as though that were a bad thing

sciencebasedmedicine.org

53 pointstokenadult11 comments

The key observation here is that you're much likely to see false positives coming from the implausible than from the plausible. This is simple statistics -- if p=.05, that means there's a 5% chance that "no effect" would be mistaken for "effect". In cases where there should be no effect (prior probability near 0), 5% of the time you'll get a false positive and almost 0% of the time you'll get a real positive. In cases where it's plausible there could be a real effect (say, prior probability 50%), you'll get false positives only 2.5% of the time (5% of the 50% there isn't an effect) to go with 50% real positives.

Thus, the argument goes, in order to reduce the ratio of false positives to real signal, we should require more signal from implausible experiments. That is, we should make the required p value scale with the prior probability. This is a form of bias, but it's a very appropriately scientific form of bias -- bias based on evidence rather than error-prone forms of cognition. (This is simply a statistical reframing of the statement "extraordinary claims require extraordinary evidence".)

Isn't this the very definition of confirmation bias?

It's more like Bayesian inference.

In confirmation bias, you specifically seek out information that confirms your views and avoid/ignore information that might contradict it. The criteria you use to filter information is "do I want to believe this?" To use the example from the article, you might choose to only read articles from homeopathy sites, or to relentlessly nitpick articles from critics.

In a Bayesian framework, you compare the strength of information that contradicts your views to the strength of the information that supports those views. The criteria you use to filter information is "is this good evidence in relation to the rest of the body of evidence?" In the homeopathy example, you'd study science and then note that there's a lot of strong evidence against it, and therefore require that evidence given to support it be stronger than "5% of placebos would give this same effect".

It's the difference between ignoring contrary evidence and giving it the proper weighting.

Others can quibble over definitions, but yes, it's effectively a fudge factor in favor of whatever we believe ought to be so, and against whatever we believe ought not be so.

So on one hand, given the nature of statistics, it is necessary to cut down on the noise. On the other hand, it puts the model before reality, and necessarily obscures anything unexpected or contradictory.

It's not a matter of ought versus ought not.

It's a matter of evidence. If we have strong evidence against hypothesis X, we should require at least moderately strong evidence for hypothesis X before we accept it. We can continue to explore it, but we shouldn't give a lot of weight to weak evidence (such as likely false positives) in the face of considerably stronger evidence against.

This does, necessarily, obscure certain types of surprising results when the data supporting them is limited. That's OK. Science sometimes converges slowly.

Not so much evidence (from trials, say), rather plausibility. We do not understand how Homeopathy works, but it goes even a level deeper: There is a certain and very specific process for creating those homeopathic remedies and from what we currently know about, say, physics, that process cannot leave you with some sort of active ingredient. If homeopathy were to work, our understanding of physics (which we have lots of evidence for) would most certainly be wrong in some ways.

That doesn’t mean that homeopathy is definitely wrong, but it does mean that the signal should be very strong for us to even be considering it to be not wrong. There are also levels to this: I’m pretty sure there are remedies about which we don’t know how they work, but them working would at least not require us to overhaul the laws of physics. That’s still worse than remedies for which we already have a model of how they work in mind, but at least it’s something.

Although you can’t really quantify that, I think such plausibility arguments are vitally important for any research. In fact, since time and money is limited, all researchers (in any field) will always try to limit themselves to what is plausible. Explaining why your model and your hypotheses are plausible is a big part of any scientific work. Plausibility is already a big part of science.

The reason we say homeopathy is implausible is because we have a great deal of evidence to the contrary -- as you say, not clinical trials, but laws of physics that are very well understood and that are completely incompatible with homeopathy.

The very idea of plausibility must always go back to evidence; otherwise it's merely wishful thinking.

If we have strong evidence against hypothesis X, we should require at least moderately strong evidence for hypothesis X before we accept it.

That's a normative position: "We should require..." And a normative position with an excellent pedigree is still a normative position.

That's not a bad thing, though. Good science requires good judgement. Good hypotheses do not spring wholly formed from logical deduction, there is necessarily an element of informed speculation and guess-making. Evidence must then be interpreted, weighted even, and credibility assigned to third-party results.

So yes, it is very much a matter of ought and ought not.

but it's not a matter of what we believe ought be so vs ought not be so -- that is, it's not a matter of what we want to be true. It's a matter of the strength of evidence we ought to have in order to draw conclusions, given other sets of evidence. It's a normative position, but it's not the specific normative position (based on confirmation bias) you described above.

Think of it as a spectrum. On one end is the pure flake -- always giving too much weight to the newest evidence, no matter how weak, and therefore swinging wildly between different beliefs (I've heard this called "regressive bias"). On the other end is the pure dogmatist -- always giving too much weight to prior evidence, and therefore holding fast to a prior conclusion, ignoring or dismissing even strong contrary evidence ("confirmation bias"). Somewhere in the middle is the proper level of evidential weighting -- giving both old and new evidence the appropriate level of consideration, and therefore changing beliefs exactly as much as is warranted.

aka Bayesian inference.

Science-Based Medicine is a great blog. If you read HN, you should read SBM. Indeed, you should go read the complete SBM backlog.

Reply on news.ycombinator.com