8 ms·
Too Good to Be True: How More Evidence Can Decrease Confidence
- SixSigma 11y agoIt's called equivocal data. It is known to management circles, it's the part of management that requires experience and judgement rather than facts. Computers can replace decisions based on unequivocal data.
- Homunculiheaded 11y agoThere is a similarly interesting result in ET Jaynes "Probability The Logic of Science" (chapter 5). Where Jaynes demonstrates that increased evidence actually decreases your belief in a hypothesis. Jaynes gives an example of an experiment in which a psychic predicts cards, getting n out of m correct. In classical hypothesis testing H0 would be "the psychic got lucky" and H1 is "the psychic has mystic powers". Jaynes first attempts to use Bayesian reasoning to work backward to determine your prior belief in psychics. That is, how much evidence would it take to convince you, compare that to the raw likelihood and what you have is your prior beliefs quantified. He then points out that he personally would never believe the subject was psychic, this is because there are not just H0 and H1, but H2 "the psychic is deceiving the experimenters", H3 "the experimenters are making an error" , H4 "the experiments are fudging the results", H5 ... Each of these has their own prior. If your prior for believing in psychics is low enough and your prior for "the experimenters are fraud" is high enough, the more extreme the evidence the more you will be convinced that the experimenters are disreputable con artists, and subsequently the less you will believe the subject is psychic. This is actually Jaynes' solution to a huge problem with Bayesian reasoning as a form of human reasoning: "If more data should override a bad prior, then why in the 'age of information' does nobody argee on anything!" This example shows, according to Jaynes, that while we can certainly have irrational priors we can still explain human reason in Bayesian terms and still get a situation where two people faced with plentiful information will arrive at contradictory conclusions.
- im2w1l 11y agoI don't like wording "arrive at contradictory conclusions", because it implies the creation of differences in beliefs. I'd rather frame it as "maintain contradictory preconceptions". And the cause of this I would say is a lack of information: Information capable of distinguishing between H1-H5
- im2w1l 11y agoOr... now I am not that sure anymore. What if Person A believe in H0:97%, H1: 2%, H2: 1%, and Person B believe in H0:97%, H1: 1%, H2: 2%. They have quite similar beliefs, at least by some measures. But if H0 is ruled out, then suddenly their beliefs will be very different.
- newman111 11y agoThis doesn't seem particularly illustrative without some context. or significant figures. I'm all for contrived examples when they are used to help understand, but this feels like an artificial comparison.
- marvy 11y agoHow does this contradict your earlier comment? I think it's still valid: Person A believes H1 (real psychic), Person B believes H2 (talented fraud). The question now is how to get them to agree. And the answer is: design a better experiment, such that fraud is not practical, and see what happens.
- jkldotio 11y ago"A Bayes Factor Meta-Analysis of Recent Extrasensory Perception Experiments"[0] is a good read in this area. It extends the analysis of psi with Bayes factors and underscores how the priors, and a causal mechanism, really matter. [0] http://deanradin.com/evidence/Rouder2013Bayes.pdf http://deanradin.com/evidence/Rouder2013Bayes.pdf
- Tosh108 11y agoInteresting. Derren Brown, English mentalist, describes in one of his books how he deliberately failes every now and then to make it more convincing there are pschycic abilities at play instead of a mechanic trick.
- themartorana 11y agoWell that's a convenient excuse. (Get it?)
- haruspex 11y agoH0 and H1 should always be phrased so that H0 is "not H1". In the above case that would be H1 "The psychic is not just lucky". Rejecting the data based on hypothesis testing is a misapplication of the technique. That doesn't mean that data is always right but it's just another hypothesis about the data itself, e.g. H0 "data is fine", H1 "data is not fine".
- bitwize 11y agoOne of the ways in which I can spot a scam is to look it up online. If all of the reviews of the product are glowingly positive (and use similar language), then I start looking for a pyramid scheme or something driving sales.
- jacquesm 11y agoThe error there is that people assume that online reviews are by real, independent reviewers that have bought the product with their own money. Sampling online reviews does not at all have the same kind of confidence levels as independently interviewing actual users of a product.
- csydas 11y agoI think his simple test somewhat accounts for this. Botted reviews usually don't bother with the sophistication of generating noise in their ratings, banking instead on the possible sale from a spendy shopper instead of trying to convince the already skeptical. With the assumption that most botted reviews aren't going to be any lower than the two highest options, the same suspicions the article raises should come into play here as well, botted responses or shill responses. Though, reviews in general are not really all that good. Southwark's Yelp episode was a long but cute take on it, but keep in mind that almost everyone sees reviews differently. On Newegg, it's incredibly common for a part/device to get 1 egg because it didn't work for whatever reason, even when it's the buyer's own fault (e.g., buying an incompatible CPU/mobo, which happens). DoA items, misconfigured computers (e.g., forgetting to plug in a power cable), and so on all warrant 1 star, even though they have nothing to do with the performance of the item in question. (DoA being an exception, though I would argue a device being DoA by chance is not the same as widespread DoA, and that the rating system should have a special category for DoA which gets its own tally and is independent of the ratings)
- bitwize 11y agoHere's the thing: In a multilevel marketing scheme, even the word of the product users is unreliable. This is because MLMs recruit their end users as a sales and marketing force that works, essentially, for free -- with the promise if big riches if you sell enough in the program. You know "The Secret" and all that positive thinking Aladdin's genie hogwash? That's actually the public facing side of MLM propaganda. Internally, MLMs pump their customer base up with big promises about how if you think happy thoughts you are guaranteed to fly with the hidden (or even explicit!) threat that even the tiniest sad thought will send you crashing to the ground. And this extends to what people are allowed to say. The key word is "edify". You're supposed to "edify the product" (talk about how great it is) and "edify your upline" (talk about what a great guy/gal the person who introduced you into the program is). This is often backed up with threats and verbal intimidation and abuse. (e.g., "Only losers say things like that. Losers and quitters. You're not a loser, are you?" and it gets worse from there.) So for some of the most prevalent scams, objectivity can't even be expected from the end purchasers.
- gtonic 11y agoAlso see the publications of Dr. Gerd Gigerenzer about this topic, e.g. in his book 'Gut Feelings Intelligence'.
- bsbechtel 11y ago>>What Gunn et al. show is that the accumulation of consistent (non-noisy) evidence can reverse one’s confidence surprisingly quickly. I know this comment won't be accepted very well here on HN, but I think this is one of the reasons most people who don't believe in climate change don't. Science, and the subsequent reporting on it, has been consistently reporting for the past 20 years that climate change is real, and it's worse than we thought. Every few months, there is a new study being widely reported saying that X new evidence is proving that things will be much worse than we thought. If there was a study every once in a while that said X study about climate change was wrong, and we didn't consider Y factor, but climate change still seems to be headed in a dangerous direction, then I think people might be more willing to accept the thesis. Of course, the paper doesn't suggest that we are wrong when we suspect systemic error, and skeptics of climate change may very well be wrong, but maybe this will help others understand their skepticism.
- Nutmog 11y agoIt could well be. Though I suspect it's more of contrarianism in a minority of people. We also have man-on-the-moon deniers, etc. What surprises me about climate change is that they're somehow managed to show causality between one non-repeatable event (people starting to burn fossil fuels) and one other event (global warming). Holy cow causality is notoriously hard to confirm when we can't manipulate the causes. When the cause and effect each only happen once, it's surely almost impossible. I'm suspicious of that extreme confidence.
- jessaustin 11y agoThis thread is interesting juxtaposed with the "Broken Science" piece on the frontpage now [0]. Apparently science is hard, even with laboratories and repeatable controlled experiments... [0] https://news.ycombinator.com/item?id=10960411 https://news.ycombinator.com/item?id=10960411
- Symmetry 11y agoI think you're mixing up the science itself and the reporting on it. If a paper was published saying that sea level rises were going to be 20% lower than expected I wouldn't expect most people would ever hear about it. Or if they did it would be reported as showing that sea level rise was even more of a threat than we had believed. If you want the truth the IPCC reports area great for providing the scientific consensus along with the (large) uncertainty we have about exactly what climate change will look like.
- anotherhacker 11y agoMo' Data, Mo' Problems. The New Coke flop is a perfect example of this. New Coke was the result of the largest marketing / consumer research project ever. The conclusion of this research was to change Coke's formula. Coca-Cola Chairman Roberto Goizueta claimed that the decision was “one of the easiest we have ever made”. Coca-Cola thought this way for two reasons: #1 Research was done with a sample of over 200,000 customers. #2 Coca-Cola’s researchers triangulated the validity of their data with a mixed-method approach. They used focus groups, various surveys, and individual interviews. All these data & research did the opposite of what they were supposed to. Their large sample size gave them lots of useless data. Data triangulation—which was supposed to safe guard them—did the opposite: it convinced them that their useless data, were useful. Why does this happen? In an unnatural system, variance is unbounded. As your data set grows, the unbounded variance grows nonlinearly compared to the valid data. As variance increases, deviations grow larger, and happen more frequently. Spurious relationships grow much faster than authentic ones. The noise becomes the signal.
- anotherhacker 11y agoDownvoted? I gave a perfect real world example that happens in product development and marketing research.
- Lazare 11y agoAnother way of thinking about it is if your performance metrics, error logs, or unit tests are always 100% green/perfect with no failures or spikes or anything, then it's probably a pretty good sign that your metrics/logs/tests are broken.
- peter303 11y agoMaybe later in this century we'll have the alternative of objective Trial By A.I. instead of by peers. Lawyers wont like this because they win by manipulating humans.
- jessaustin 11y agoI suspect hackers will be more supportive of this reform.
- throwaway33817 11y agoI ran across a very interesting example of this recently. In a new tab take some time to analyze this chart for yourself and what you think about it. http://geekologie.com/2009/03/percent-of-student-virgins-per.php http://geekologie.com/2009/03/percent-of-student-virgins-per... ... done yet? So, firstly I see that as you go toward more rigor, overall the rate goes up, with studio art (no rigor, it's just a humanities field) at the left and mathematics (full rigor, fully abstract) at the right. Like this - https://xkcd.com/435/ https://xkcd.com/435/ Interestingly, computer science (of interest to people here) actually isn't as far right as you'd think. But what pops out to everyone!! There is obviously one MOST interesting fact in the whole figure, the 0 for studio art. That really got me thinking. Granted, this isn't the perfect example because it's not just 0 as opposed to 5-7 - the next higher one is 20%. But what you really start thinking is stuff like - "Is it that they didn't really have many subjects, just 2 or 3? what is n? Is it that these people are lying? Is it a real fact?" and so forth. The fact that it's 0 instead of a few percent immediately raises some questions.
- muxme 11y agoI'm running into this problem on my startup website (http://muxme.com http://muxme.com). I've found that the more evidence I provide that the site is real, the more people start to question it. I.e. I post pictures of the receipt of purchase, open source the raffle code script, have a live drawing, post the winner's usernames (which reveal quite a bit about them if you google them, I encourage them to change their usernames), post USPS tracking numbers, post my phone number and address, and everyone still thinks the site is a scam. Out of every company / website I've worked on, I've never had people so skeptical of giving a name, email address and phone number. I've made websites where I charged money and had better growth! This is some real feedback I've received: "It sounds like a scam to me. If you want to try to scam people by getting them to click on your fake website, you should have enough common sense not to use the website as your username. Better luck next time!!!!" "Scamming people on Christmas Eve using someone's else's platform. Tacky." But yeah... this site isn't working. I'm thinking of scrapping it and writing an app/browser extension that automatically enters you into sweepstakes. Let's say Person A enters 10 sweepstakes on different websites. Person B enters 5 sweepstakes on different websites. When person C downloads the app, he will automatically be entered into the sweepstakes that Person A and Person B entered (via my magic backend system that I have yet to create). The more people that download the app and enter sweepstakes, the more sweepstakes everyone with the app will automatically be entered into. Now people will start winning prizes for literally doing nothing but downloading the app.
- stordoff 11y agoI can't quite put my finger on the reason, but I can see why people are sceptical. Few possible reasons, which may or may not be large factors: - Non-obvious business model - It's not obvious how you profit from the raffles, whereas charging money is generally easier/quicker to understand - Use of memes (e.g. http://muxme.com/giveaways/56-25-toys-r-us-gift-card http://muxme.com/giveaways/56-25-toys-r-us-gift-card ) doesn't inspire confidence (may be a personal quirk of mine) - Layout/design - My initial reaction when the page loaded was "Oh, it's a penny auction site", which are known for being borderline scams. Obviously it isn't, but that gut reaction is difficult for you to overcome (I'm thinking of sites like swoggi.co.uk when I say penny auctions)
- tabbott 11y agoSigh. This article describes a mathematical analysis that is more an artifact of the problem setup than a proof that more evidence decreases confidence. In particular, what's going on is that it assumes that while actual witnesses have a probability of e.g. 50% of convicting, biased witnesses have a probability of 100% of convicting(!), which is an incredibly strong assumption (and in particular, if you changed 100% to anything else, one of the headline results they advertise of unanimity being suspicious and near-unanimity not being suspicious disappears). Those values of the parameters don't make sense for many real-world scenarios like picking a subject out of a lineup (you'd have to have a really really bad lineup to have witnesses be guaranteed to pick the same wrong person!) or to the output of a panel of judges or jurors. However, the result is relevant in scenarios where it's very likely that you'd have a 100% likely error, e.g. you're doing an experiment and your apparatus isn't functioning properly or you have a bug in your computer program doing the data analysis. So it's relevant for physicists and computer programmers, but probably not to the criminal justice system. For those curious, the math is quite simple. They model a world where one of the following happens: * probability p: compromised experiment, in which case fraction y (y=1 in the paper) of witnesses report guilty * probably 1-p: normal experiment, in which case fraction x of witnesses report guilty Given that, the probability that 100% of N witnesses agree the person is guilty in each case is: prob(unanimous and compromised) = p * y^N prob(unanimous and not compromised) = (1-p) x^N ~= x^N (since p is assumed small) So the probability that the experiment is compromised, given a unanimous outcome, scales with their ratio, p y^N / x^N = p * (y/x)^N. A few things you notice: * If y <= x (as you'd hope to be the case in a lineup where the bias is less convincing than having actually seen the criminal!), unanimity is never suspicious. This is the usual world where more evidence should increase your belief in a hypothesis! * If y > x, you have exponential growth and so with sufficiently large N you'll see unanimous results are almost always the result of a compromised scenario. However, unless y/x is large and y is close to 1, the actual probability of unanimity is basically 0 so this will essentially never happen. * If y=1 and x << 1 (the case they analyze), with large N, unanimity is suspicious (but if you change y to 0.95, near-unanimity is, too, so ). This is a case very relevant in software, but less relevant in the criminal justice system. They picked y=1 and x=0.5, which means roughly you start to see a big effect of this form when N > log_2(1/p), which for their ranges of p from 0.01 to 0.0001 means N ~ 15-30 range. Unfortunately this is part of a much broader trend of people making bold claims about society based on the results of either a study with a sample a freshman psychology class that achieved 95% confidence (just like ~5% of the thousands of other studies done on psychology classes!) or a mathematical analysis of a confusingly designed problem that doesn't reflect well the broad swath of reality its authors (or the people reporting on it to a wide audience!) are claiming it describes.