It's better mockery of the P
I am due to be converted to Bayesianism, presumably by Jaynes, but haven't gotten there yet.
Duh. Of the P < 0.05 than of frequentism per se.
(But even I know that the Bayesians believe we know that the possibility the sun has exploded is minuscule, minuscule, so that if the detector says Yes you bet it's wrong. Besides, if the sun has exploded saving the money won't do much good.)
The Origin of Statistics in the Breakdown of Prior Expectations?
That cartoon moved me into the neb view of xkdc.
That and the continuing wisdom emitted by his brain.
(Also, if the bayesian is so confident that the sun hasn't exploded, why are they consulting the machine in the first place?)
Andrew Gelman doesn't like the cartoon.
It's a really stupid cartoon. He should add dinosaurs.
Gelman doesn't address neb's question about the calculation, though, nor has anyone in this thread.
I am confident that Cosma will be along in due course and set me straight.
Does he have grading to do or something?
To answer the OP question, though, the frequentist is supposed to set ((the sun has exploded and it rolls at most one six) at 0 since it hasn't happened and or(the sun hasn't exploded) at 1.Then 1 in 36 falls out after that.
Of course, if the chance of the sun exploding is really 0, there's no need to ask the machine, and the final answer contradicts the premise, but that's not really a defect in the cartoon.
Sorry:
(the sun has exploded and it rolls at most one six) at 0 and (the sun hasn't exploded) at 1.
I can't tell if the sun has exploded because it's night.
While we're criticizing Munroe's presentation of statistics, this post is good on how this comic is also misleading.
Of course, if the chance of the sun exploding is really 0, there's no need to ask the machine, and the final answer contradicts the premise, but that's not really a defect in the cartoon.
I think if the cartoon is internally incoherent in multiple respects, that is a flaw.
I think if the cartoon is internally incoherent in multiple respects, that is a flaw.
It's supposed to be a reductio ad absurdum of frequentist statistics.
I mean it occurred to me as well that if you set p(the sun hasn't exploded) to 1 then you can arrive at 1/36; that's why I said, in the post, that we're assuming that we don't know whether the sun has exploded (and I am assuming that if we are willing to set p(X) to 1 then we at least consider ourselves to know that X). After all, as the cartoon header says: "we're not sure". P(X) = 1 sounds like being sure.
Surely one can be a frequentist without believing that p = 0.05 is a magical value.
… and the absurdity of the frequentist's position seems to depend on his treating that way: since he knows already that the sun hasn't exploded, p(yes) = 1/36, which is less than 0.05, therefore, we must be in the has-exploded-no-lie case. Everything turns on the frequentist's taking being less than 0.05 as somehow decisive, so an equally good case against frequentism could have been constructed thus:
A frequentist playing with dice rolls two sixes. Since the probability of his having done so is 1/36, he concludes that he didn't do so.
22
The null hypothesis is that the sun hasn't exploded. Since in that case there is only 1/36 of observing a "yes" and we observe a "yes" we can reject the null hypothesis (using a 5% significance level).
Ok, so when the frequentist says that there's a 1/36 chance of "that happening" he doesn't mean there's a 1/36 chance of observing a "yes" sans phrase, just that there's a 1/36 chance of observing a "yes" given that the following hypothesis, that the sun hasn't exploded.
I can accept that.
I remain confused about the use of the 5% significance level.
The sun may have exploded, but we won't know about it for another 8 minutes.
the author clearly thinks the unthinking use of 5% significance as dispositive is moronic and that this is widespread. Why are you pretending not to get this? This is pose is stupider by far than the comic, and seems to be your sole lens on xkcd. That seems a shame.
The unthinking use of 5% significance as dispositive is moronic. Is it a key point in the dispute between bayesians and frequentists, with the former opposing and the latter supporting it? That would be surprising.
I'm not sure what "this is pose" means, so I don't know what is supposed to be my sole lens on xkcd, but I've certainly linked/alluded positively to it in the past.
There is no difficulty in proving any statistical approach whatever to be absurd, if we suppose universal idiocy to be conjoined with it. … Men really ought to leave off talking a kind of nonsense on this subject.
is
pose=pretending not to get it. I guess I may be incorrect to assume you have sufficient good will to go along with your more than sufficient intelligence.
Just to make sure that I get it (I have enough good will, it's my knowledge that's lacking), the connection between the P value and frequentist statistics is that the frequentist has to use such tests, since she can't appeal to prior odds for unprecedented events?
Oh, I wasn't sure if "pose" was a typo for "post", you see, in addition to the extraneous "is". Anyway, I've certainly been critical of xkcd in the past without pretending not to get, or actually not getting (I believe), what he's on about, and I wouldn't even say that in this case I've been pretending not to get what he's on about. It really seems to me as if he's just conjoining unrelated things. Perhaps I am, in fact, simply not getting it, and mindlessly applying a 5% cutoff is the province of all and only frequentists.
26
I remain confused about the use of the 5% significance level.
By longstanding convention observations of some purported effect are called significant if there is a less than 5% probability that they would have occured by chance absent the effect.
So if for example you want to test whether painting classrooms purple improves student performance you take a bunch of classrooms and paint half of them purple at random. You then look at how the students do. If the students in the purple classrooms do better to an extent that would occur less than 5% of the time by chance (if purple paint made no difference) then you have a significant result which you can publish and get tenure and government grants. If the students in the purple classrooms do better but by an amount that would occur by chance 6% of the time then you don't have a significant result, your paper will probably be rejected and you won't get tenure.
This creates various bad incentives which are widely deplored but not to the extent of doing much of anything about them.
33: Bayesians tend to argue against any kind of threshold for meaningfulness in statistical analysis. Frequentists tend to be less worried about that as a problem. But these are mostly sociological descriptions, as opposed to strongly methodological ones, I think.
34: I mean I know that; the question is why it's being trotted out here.
Anyhow Shearer isn't broadly right; in many cases you have to do a fair bit more justification than just pointing to a significant result in order to get through peer review. Unless you're Daryl Bem.
Isn't it 1/36 because the cartoon was already written to pick on a p-value of less than .05 and no other value than 1/36 derived from a pair of dice could meet that requirement? This isn't science or math here. It's xkcd.
"Already written"? It's not as if he got the cartoon complete with the two dice and had only to write in a fraction somewhere in the lower left panel.
I just mean that if you've got two dice and want to make fun of the .05 thingy, you've got only 1/36 to work with.
I'm going to have to conclude that my knowledge of statistics is insufficient to understand why the meaning of this cartoon isn't as clear & obvious as it appears.
I have no knowledge of the baseball stat world, but I thought article on Nate Silver was interesting.
Although what I know of Macleans makes me doubt what it publishes.
37
34: I mean I know that; the question is why it's being trotted out here.
As has been stated by others XKCD is making fun of people who use a (totally arbitrary) 5% cutoff for significance.
I have already responded to those others, James, even in comments prior to the one you're answering. As JS Mill pointed out, a reductio of frequentists is ineffective if it only works if the frequentist happens to hold additional absurd positions.
The article in 43 is indeed interesting. Maybe we should declare Sam Wang our new god instead.
46
I have already responded to those others, James, even in comments prior to the one you're answering. As JS Mill pointed out, a reductio of frequentists is ineffective if it only works if the frequentist happens to hold additional absurd positions.
So you are expecting a cartoon to be fair and balanced?
I think people fetishizing statistical significance too much is a problem, but I don't really think it's a frequentist vs Bayesian thing. The comic amused me, but it seems like it's targeting people who don't think when they apply statistics, not frequentists.
I agree with Shearer, essear, and Turgid Jacobian: the point is to make fun of unthinking incoherent ways of thinking about statistics, not about frequentism per se.
That said, my understanding of frequentism is that it's an incoherent theory-less cookbook-style approach to statistics. I've often found that people who are frequentists hold absurd positions. (e.g. Sash/a Vol/okh's old argument that if the margin of error of a poll is 3 and it shows one candidate up 2.9% then you literally haven't learned anything and you can't say that one candidate is up.)
50.2: But I feel like the debate among real statisticians is fundamentally different. Doesn't Cosma identify more as a frequentist?
I had a discussion with another physicist recently where he was saying that he found the following problem clarified his thinking on Bayesian vs. frequentist: you are told that a coin has been flipped 14 times and landed heads 10 times. The coin is now going to be flipped two more times. Would you take a bet at even odds that the coin will land heads both times? Supposedly the frequentist will judge that the probability of heads, based on the observed trials, is ~71% and so it has a slightly better than 1/2 chance of landing heads in two trials, and will take the bet. The Bayesian will sum over all assumptions about the probability of heads, weighted by the existing information, and decide not to take the bet.
My objection to this is that it sounds like we're dealing with a stupid frequentist and a stupid Bayesian. A reasonable frequentist might also say "I know that most coins land heads 50% of the time, so let's call that the null hypothesis. On the basis of the observed information, I can't reject the null hypothesis at 95% confidence level. So I'll proceed assuming the null hypothesis is true, which means I expect the chance of winning the bet is 25% and I won't take it." A reasonable Bayesian would probably take not a uniform prior but one peaked around 50/50 heads/tails.
I don't think these kinds of toy examples are very useful-- I think probably people can do reasonable things from a frequentist point of view or a Bayesian point of view, and on a case-by-case basis it's going to involve making judgements that can't be discussed purely as airy abstractions.
I think probably people can do reasonable things from a frequentist point of view or a Bayesian point of view
Nice prior.
I'm never able to understand what a "null hypothesis" is other than a strongly peaked prior...
The other xkcd foray into this area was the one that showed 20 studies, one coming up with p<0.05 and a headline touting that result. No mention of false detection rate correction.
I'm curious about whether there was any historical link between the ideas, but it seems like there is some kind of cocktail of Karl Popper and old-fashioned frequentist cookbook reject-the-null-hypothesis stuff that strongly influences the way people (including scientists) talk about science in a really annoying way. Lots of "science can never prove things, only disprove them" or "we're allowed to rule out hypotheses but never to rule one in" or whatever. It's all obvious nonsense-- the same people will turn around and talk about what the data is actually telling us is true in their next breath-- but it still seems kind of pernicious to me.
Here is the link to the comic mentioned in 55
55: On the one hand, you're obviously right that it's unfair to ignore the more sophisticated methods that people actually use. On the other hand, there's something deeply unsatisfying about the approach where you notice that you're getting the wrong answers so you add a new ad-hoc fix on top of that, rather than scrapping the original approach because it produced nonsense.
56
... It's all obvious nonsense ...
It doesn't seem like nonsense (much less obvious nonsense) to me.
58: Constructing and testing a reasonable hypothesis isn't a ad hoc fix to frequentist statistics.
55
... No mention of false detection rate correction.
Possibly because most people (even people who know about 5% significance levels) have never heard of it (I had to google it).
Anyway it would be a bit hard to apply if the 20 studies were done by 20 different groups.
an equally good case against frequentism could have been constructed thus:
A frequentist playing with dice rolls two sixes. Since the probability of his having done so is 1/36, he concludes that he didn't do so.
That would make a good cartoon!
I have twice found $20 bills on sidewalks. Am I not allowed to joke about economics-inspired oversimplifications?
62: Why not? There are less facile arguments to make, but for one thing, at a purely logical level "ruling out A" and "ruling in not-A" mean basically the same thing.
I still haven't finished my stupid grant proposal so I should stop commenting here today.
I found four twenty dollar bills in Lake Michigan. I got out of the water because there must have been a dead guy in the water whose pockets were being washed empty.
67
62: Why not? There are less facile arguments to make, but for one thing, at a purely logical level "ruling out A" and "ruling in not-A" mean basically the same thing.
OK, you have to be talking about the kind of hypothesis that can be falsified by a single observation (ignoring errors). Like "there are no black swans".
64: d'oh. Do your research, Munroe!
I'm never able to understand what a "null hypothesis" is other than a strongly peaked prior...
Me neither!
In fact, I often feel like I should really go take a statistics class and learn some of the basics.
56: I assume the historical connection is that Fisher was the one who originally suggested 0.05 as a rule of thumb.
I think the issue is that frequentist statistics are the methods most frequently taught to non-statisticians and to people who have the attitude 'give me the recipe so I don't have to think about it'. They are the people most likely to do cartoon-able things. If everyone was taught Bayesian statistics I'm sure laughable things would happen based on only ever using the prior you got taught that one time in your one required stats class for your major.
The thing to remember about the null hypothesis is that almost nobody appreciates it when you look at H0 and say "Ho." And they get really angry when you imply anything about their mother related to Hos.
The trouble with the comic isn't that it's not getting across any kind of point, it's that one of the perks of xkcd has been that it holds up to at least casual scrutiny. The one linked in 55 might be about an error which well-established methods exist to correct, but it does reflect real life (when lots of similar small studies are done all the time, considered as separate studies, then have publication bias and media bias stacked on). I do get that the current comic is about p and rejecting the null, but beyond that it makes about as much sense as something an editorial cartoonist would make if p-standards somehow became a national issue. The frequentist knows there's a dice-roller in the machine.
I have more sympathy for him against the critique in 18. There really were pundits explicitly rejecting the concept of poll aggregation and basing their forecasts on feelings and momentum, over whom all poll aggregators and analysts, including Nate Silver, collectively triumphed. (And the comic doesn't mention Nate Silver - it wasn't changed, was it?)
56: I assume the historical connection is that Fisher was the one who originally suggested 0.05 as a rule of thumb.
As I understand it the original 0.05 rule of thumb was proposed as a rule of thumb for the particular problem for which it was introduced, where 5% made sense for calling some result "significant", not as a general rule about what statistical significance is.
Not that it is directly relevant to the cartoon under discussion, but I've been seeing things that seem to imply that Bayesian:frequentist::Nate Silver:Karl Rove. That's just bizarre since Rove was adjusting based on priors also. He was just picking really stupid priors.
56: agreed on the annoying cocktail, but surely crude positivism ('the data tell us XXX') is no better. I think the reason Popper is so popular among working scientists (even though philosophers of science have long since moved past) is that he lets you do a nod to the problems with crude positivism without really coming to grips with the role unproven assumptions perform in science in rationalizing contradictions, ignoring inconsistencies, etc. Which if a science is predicting well most working scientists do get to ignore in their ordinary work.
77: I don't know the story but that sounds very plausible. I have been told that he really did not intend for it to be a general rule used by everyone.
77 is incorrect---the 5% level was very much a general rule about what significance is. But he did reconsider later in life, and especially in his last book, as suggested in 80.
65- I don't think "most people have heard of it" is a criteria for inclusion in an xkcd strip. And as can clearly be seen by the detailed goggles on the stick figure, they're definitely all the same research group.
73 56: I assume the historical connection is that Fisher was the one who originally suggested 0.05 as a rule of thumb.
But what does Fisher have to do with Popper, if anything?
The passage I had in mind is from, I believe, The Design of Experiments: "it is usual and convenient for experimenters to take 5 percent as a standard level of significance, in the sense that they are prepared to ignore all results which fail to reach this standard." But his writings on the topic were not at all consistent.
Would love to follow the thread, but I must be going...
84: oh, I see. Yeah, dunno. Somebody wake up ttaM...
86: somewhat differently, but apparently consistently, stated here:
If one in twenty does not seem high enough odds, we may, if we prefer it, draw the line at one in fifty (the 2 per cent. point), or one in a hundred (the 1 per cent. point). Personally, the writer prefers to set a low standard of significance at the 5 per cent. point, and ignore entirely all results which fail to reach this level. A scientific fact should be regarded as experimentally established only if a properly designed experiment rarely fails to give this level of significance.
I'm going to wait for 36 comments by other people before I accept that Kreskin is really gone.
I wonder where I got the idea in 77, then.
83
65- I don't think "most people have heard of it" is a criteria for inclusion in an xkcd strip. And as can clearly be seen by the detailed goggles on the stick figure, they're definitely all the same research group.
OK, the comic is a little incoherent in that the blame is implicitly being placed on the media instead of on the scientists for failing to present their results properly but the end result is the same.
89-That's certainly how newspapers report polls. Within MOE so they're really tied! Who knew political pundits were so knowledgable about the history of science.
The quote in 89 is interesting because it shows how frequentist methodology is designed for repetition of experiments (see the last sentence in the quote). Lots of softer science areas don't place much institutional emphasis on repeating experiments, use historical data points as though they were repeated experiments on the same underlying process, etc. Also, it makes little sense to talk about statistical significance outside of a theory -- every fact you pull from a full population survey will be statistically significant at the highest possible level but your theory is the only thing telling you if they will still be true a week from now. This is obvious in one sense but worth repeating as most of the real uncertainty is usually in the relationship between the finding and the theory.
[Gah, previous attempt at any entry eaten]
0. To answer the question in the OP, it's really
Pr("Yes"|Sun) = 1/36
so the measurement is relatively improbable under the hypothesis that the Sun is still there. Since 1/36 < 1/20, the conventional threshold, the frequentist idiot is supposed to blindly reject the hypothesis that the Sun is still there, and conclude that it is not there.
The 0.05 p-value threshold owes its prominence, I think, to three factors. (I) Fisher's use of it, though as I recall he thought of this more or less as the highest acceptable threshold, not an especially compelling one. (II) 5% corresponds very nearly to 2 standard deviations with a Gaussian distribution, and even doctors can multiply by two. (III) Self-reinforcing conventions, especially outside the community of statisticians. I'm quite sure that no actual frequentist statistician today would say this was anything more than a convention, and few of them would say it was a wise convention.
To treat the cartoon more seriously than it deserves, I'd actually say that it would show some, very weak, evidence for the sun having blown up. The probability of "Yes" when the Sun's blown up is fairly high (35/36), but low otherwise (1/36), so it's sensitive to what we're trying to infer (it's not just a Gygax test), and while the error probabilities are really quite high, it's not worthless. One can elaborate on this.
Ideally, null hypotheses are actually models which try to account for apparently interesting patterns in the data by means of boring or already-known processes. (E.g., might this apparent clumping of events really be due to mere fluctuations? Can we explain the spread of this mutant through selectively-neutral processes?) Fetishizing parameter values of zero is stupid.
Turning to the Bayesian side, the Bayesian statistician actually is much more convinced that the Sun has blown up after seeing the machine say "Yes" than they were before. Specifically, the ratio of belief in "Sun kaboom" to belief in "Sun fine" has increased by a factor of 35. According to the scale published by some aggressively-Bayesian forensic scientists, this is "moderate support" for the Sun having blown up.
Of course, there is nothing in the machinery of Bayesian inference which tends towards true, reliable answers, except under assumptions where there are also reliable frequentist methods (and sometimes no Bayesian method can work, while there are reliable frequentist ones). This has a lot to do with why the sensible practices of Bayesian data analysts have so little to do with the ideology they profess.
Shorter me: Randall Munroe is wrong on the Internet.
So "due course" for Cosma is about eleven and a half hours. Good to know!
You'd really need more trials to establish a reasonable CI on that.
84: The statistical theorists of hypothesis testing (Fisher, Neyman and Pearson) and Popper seem to have completely ignored each other; they may not even have been aware of each other's existence. The first author I know of to find this puzzling was R. B. Braithwaite (Scientific Explanation, 1953).
Gerg Gigerenzer has some papers on the theme of how so much of the social sciences got caught up in a weird Fisher/Neyman-Pearson/positivist/Popperian mish-mash of rejecting hypotheses that parameters are exactly 0, at the 5% level. (See e.g. the group volume The Empire of Chance.) Whether other historians disagree, I don't know.
My 2/3 completion of the statistics 101 course on Udacity.com has left me unable to participate meaningfully in this discussion. The Internet is a pack of lies and false promises.
||
Anyone have restaurant recs for New Orleans for places that might have a few vegetarian options? If not, I'm open to a citywide beignet crawl.
|>
Dammit, now I'm being lured back into procrastination-land by all of Cosma's links.
Also by the basic cognitive dissonance that: I'm applying for a grant so I can hire a postdoc; in my field good postdocs are expected to be independent and so the person I hire might not collaborate with me at all; but to decide whether this is a good use of their money the government requires me to write 15 pages about what I will do, rather than asking the hypothetical future postdoc what s/he will do. I guess things make more sense in labs or whatnot.
101: "The postdoctoral fellow will assist in all aspects of this work."
103: Yeah, but, that's almost certainly false.
96, 97: So far as I can tell, this is a pretty good account of my response times.
104: Distinguishing it from any other sentence in a grant proposal how, exactly?
Hey, Cosma, I see you got tenure. Congrats.
107: Thanks, but here promotion to associate professor is a pre-tenure up-or-out step. I come up for tenure in 2014. (And if that happens, promotion to full professor a few years later.)
Hey, Cosma, I see you work at a bats institution.
Congrats on getting associate professorhood.
106: Hmm. I guess I'm not yet cynical enough.
Yeah, but, that's almost certainly false.
What's the p-value on that?
re that link of Cosma's in 105, now I want to know more about the "numbers of men in ten Prussian army corps killed by the kick of a horse..."
Congrats Cosma.
Thanks, all!
109: I have never seen any bats around campus in the seven years I've been here. But there are raptors (falcons, maybe?) which sometimes perch on the edge of the building opposite my office window, looking about with infinite disdain.
||
A thing I did not formerly know: Hubert Sumlin's funeral was paid for by Mick Jagger and Keith Richards.
|>
We have hawks that perch on our building and the buildings nearby and swoop around outside my window. It's pretty great.
||
Today's cultural confusion: why do all the female mannequins on display here have such prominent nipples?
|>
Isn't it cold where you are? I can't keep track of your travels.
And it only hurt for little bit at first.
The link in 102 is quite devastating.
102, 124: I've now removed Udacity Stats 101 from my list of things I feel like I should do but probably never will.
Although in an odd way, I am now more likely to investigate the course, because I often need examples of how not to do online education.
To Udacity's credit, Thrun responded pretty quickly on the Udacity blog and linked to that critical review acknowledged lots of problems with the course, said they'd be revising it, and I think the revisions might already be in the course now.
I haven't actually done the stats course, but I get Udacity's email announcements since I've done/am doing other courses. I really thought the CS101 was done well.
126: Huh, so he did. The author of the original critique also responded.
Speaking of doctors and statistics--and since people who tend to know things are around--I've been a bit curious about something I saw a short report on in Science News a few months back. It discussed work that purported to describe how to build an "evidence" scale which was analogous to temperature. This seems to be the key paper: "Measurement of statistical evidence on an absolute scale following thermodynamic principles" - V. J. Vieland, J. Das, S. E. Hodge, S.-C. Seok.
I'm going to presume that it is either wrong or trivial/not very informative or useful (and based on my quick peruse of the paper I'd guess the latter). So having done the hard work of noticing it in Science News and finding the paper in arXiv, I leave it to others to judge and report back.
I was kind of hoping for a figure that started with "Likert scale where the coders kept forgetting whether 5 or 1 was 'strongly agree'" and ended with "atomic clock."
Is there anything funnier than Likert scales?
1) Definitely no.
2) Probably no.
3) I don't care.
4) Probably yes.
5) Definitely yes.
130: It's horribly organized, so it's very hard for me to say what they're actually doing. Looking at their Table 1, their "evidence" looks like it's basically the ratio
(likelihood at its maximum)/(integrated likelihood with a flat prior)
What they are taking this to be evidence for; why they think that statistics should look just like the ideal gas law, when most of physics doesn't look like the ideal gas law; why they make a big deal about having an equation of state for variables they make up so it has the form they like, as opposed to physical quantities which can be independently measured; God alone knows.
I'd say, not worth anyone's time.
136: Thanks. As I suspected.
But before I die I want our whole understanding of the universe to unify around some information/entropy formulation that explains everything and it's the end of science and everyone gives up and goes home and that's that. So I keep my eye out.
Oh, and is simple enough for me to understand.
I'd say, not worth anyone's time.
Roger that!
Hey Cosma, I heard a reasonably spirited defense of Bayesian cognitive models (as, basically, just as good an approach as any other on the computational level) the other day, you will no doubt be thrilled to hear.
Boy, I killed all the threads. Look!
just as good an approach as any other on the computational level
So, lacking evidence, solid theoretical support, and a realistic implementation?
So much for that attempt to start conversation.
Sorry, Sifu already killed all the threads.
My trollery is too crude for thread resuscitation.
Maybe people have started going to bed or something? I guess not everyone has tomorrow off.
142: that would be my take, sure. But so is it best to ignore that level of analysis entirely?
(I would have responded sooner but I got distracted trying to figure out why the New Yorker published an enormous, fawning profile of Kid Rock.)
148: That is a puzzler. Is he trying for a comeback or something?
Well, apparently he has become very popular as a country rocker, and wrote Romney's campaign song. Still weird, though.
But so is it best to ignore that level of analysis entirely?
It's central to our existence and the only area of science still at the alchemical stage. I wouldn't.
Have scientists really made no inroads at all into the study of Kid Rock?
Presumably, everyone was off watching the Bears play a shitty football game.
151: well, right. So then, the argument goes, Bayesian models of inference are at least analytically and computationally tractable and relatively intuitive, even if they're not immediately, obviously plausible.
(I would have responded sooner but I got distracted trying to figure out why the New Yorker published an enormous, fawning profile of Kid Rock.)
Was it written by SF/J?
139: Unless they have a good fix for the problem pointed out by Eberhardt and Danks, I'd say they're in very deep trouble even as computational models. (Whereas one can do perfectly respectable Bayesian statistics with, say, ACT-R models.)
Seems to be by someone named Kelefa Sanneh. That abstract is like the perfect distillation of the ridiculousness of this kind of article.
And speaking of distillation, I'm currently polishing off the bottle of Evan Williams I bought a while back. Anyone have recommendations on other whiskeys I should try?
Bulleit, and their rye's not too shabby either.
ISTR that the Hudson Baby Bourbon is quite nice, but it's v. spendy. (I just had it once at a bar.)
But before I die I want our whole understanding of the universe to unify around some information/entropy formulation that explains everything and it's the end of science and everyone gives up and goes home and that's that.
There's no shortage of people claiming to do things like that. You can just pretend like one of them isn't a crackpot. We'll lie to you as you're dying, if you want.
Don't listen to him, JP. Physicists have their reasons for delaying the time when there's a grand unified theory of everything.
re that link of Cosma's in 105, now I want to know more about the "numbers of men in ten Prussian army corps killed by the kick of a horse..."
You ask, we answer. It's from Ladislaus Bortkiewicz "The Law of Small Numbers" (1898); not obviously available online, but according to this biography (http://www-history.mcs.st-andrews.ac.uk/Biographies/Bortkiewicz.html) over twenty years, fourteen army corps of the Prussian army suffered between 0 and four soldiers killed per year from the kick of a horse.
You'd think that, after a bit, they would have figured it out and shot the horse.
I doubt I'll learn enough statistics to fully understand a thread like this, much less the stuff linked from it, but I'd really like to have a better understanding of Bayesian statistics. It seems like I've been running into Bayes rule everywhere lately.
But before I die I want our whole understanding of the universe to unify around some information/entropy formulation that explains everything
Check the quantitative biology section of the Arxiv regularly. I've found that several times a year someone deposits a paper claiming to have discovered the secret of Life, The Universe, And Everything. Said papers invariably involve information theory.
Oh, and bonus points if it's a cellular automaton. Maybe like a new kind of shit.
170: Yeah, that's what I was thinking of with my comment about crackpots above.
I kind of see how information is important, but it also seems like a case of it becoming a dominant metaphor in our time because of the prevalence of computers. Just like people used to have a "clockwork universe" and then in the 19th century thinking in terms of heat engines became popular. Metaphors driving the way we think about the world track technology. I'm sure I've said this before. Probably someone wrote a book about it or something.
159: I mean, I think in this case they weren't even talking about being tied to the data at a level that would privilege the bayesian explanation particularly; it was more that, hey, these models are easy to create and simulate, and they could be a synecdoche for a replicator process or whatever, so let's see if people behave at least consistently with hierarchical versions of these models because maybe that'll give us some information about representational structure.
The "synecdoche for a replicator process or whatever" is my gloss; the person making this case skipped the "a replicator process or" part of the phrase.
If recent history has taught us nothing, its that fawning biographers are usually banging their subjects.
Oh, now I think I get "null hypothesis".
If 175 is to 174, you've got a great stats text.
175 was indeed to 174. I was being extremely hilarious.
So you can't help me find weird porn?
Be the weird porn you want to see in the world, Mobes.
170, 171: I was also thinking of Edward Fredkin.
172: Yes, I think there is a tendency to go all in* on metaphors derived from the predominant technology** of the age. Somewhere I read (and have frustratingly not found again) a Twain essay where he gives his late 19th century view. There probably is a book, I first saw it discussed in a critique of Minsky back in the 80s.
*In the pre-Petraeus sense.
why they think that statistics should look just like the ideal gas law, when most of physics doesn't look like the ideal gas law;
Because so many in the professoriate are full of hot air?
(The above came to me in a dream.)
I actually owned the book in 170 at one point. That fucker weighed like 6 pounds.
Statistics are annoying because my college library appears to have stopped buying physical books in 2004 in favor of all-electronic offerings. So I don't know where to go to familiarize myself with the basics of bioinformatics. Presumably whatever books are available would be largely collections of glorified software package manuals, but at least it would be a book instead of a collection of unprintoutable computer files accessible from a small number of computers.
Is there not a Bioinformatics for Dummies?
Well I'll be damned. I guess that would be where to start.
Latest edition published in 2006.
I'm sure nothing has changed in the field over the past 6 years.
At least nothing new that a dummy can understand.
Thanks to 114 and 167 everything I know about Poisson distributions I know from reading "Gravity's Rainbow" and links posted in Unfogged threads.
everything I know about Poisson distributions
It's just about counting. You learned it in kindergarten.
Right. One fish, two fish. Rouge fish, bleu fish.
