Untitled

jenroses:

geekremix:

yayfeminism:

Very informative thread -source

Oh no I’ve been duped. Shit…

Yep.

It’s pretty wild that Dr Taber’s series of tweets did not include a single piece of evidence or citation to back up her disagreement with a very well referenced report commissioned by the Global Food Security Program. (The two references most relevant to fruit waste question are a report by the UK’s Government Office for Science and a technical report from researchers at Cranfield University).

Of course her various assertions may be right or may be wrong, it’s hard to learn much from the mere fact that there exist three expert reports putatively claiming something (you’d need to dig into the details). But you can learn even less from the fact that someone with expertise in an adjacent field asserted something.

(via hoursofreading)

contingentvaluation:

slatestarscratchpad:

I’ve been trying wrap my head around Treating Stimuli As A Random Factor In Social Psychology today. I think I mostly understand what it’s saying, so I’m going write my thoughts here, and any real statisticians can tell me if I’m on the right track or not.

Suppose you’re doing one of those studies where you test racism. You get a bunch of subjects, you show them some pictures of black and white faces, and then you measure something that ought to show you if the participants are racist. Maybe their skin conductivity changes when they see the black faces, indicating that they’re stressed. Whatever.

All studies need to decide whether to analyze things as random or fixed effects. A random effect is something where your study is sampling from a larger population, and you only care about the larger population, so you better hope your study has a large enough sample size. A fixed effect is something where your study includes the entire space of possibilities that you care about and you don’t need to worry about samples. For example, in the study above, let’s say you have n = 20 subjects who are viewing these faces and having skin conductivity responses. We don’t care about your 20 subjects. We care about the population in general. So we need our models to contain a term representing the difficulty of generalizing from 20 subjects to the broader population. That’s a random effect. On the other hand, race is a fixed effect. Our study has 2 conditions, “white” and “black”, and we actually just care about white vs. black people. The wider world may have other races in it, but we’re not trying to make our study represent those other races. We’re just legitimately interested in the how subjects treat white vs. black people.

Judd and Westfall say that almost everyone correctly treats subjects as a random effect. In other words, sample size matters, because you might have accidentally picked crappy subjects. Maybe the population as a whole forms a bell-curve centered at zero racism, with a few people who are very racist against blacks, and a few other people who are very racist against whites. If by bad luck you pick ten subjects who are all very racist against blacks, your study will come to the wrong conclusion (the population is racist against blacks). So you treat number of subjects as a random effect, and then your model adjusts for this possibility, ie penalizes you appropriately for low subject sample size.

(this is different from random noise in the experiment. There will definitely be random noise in the experiment. This is about bad sampling. Even if you choose the person in the exact middle of the population, she will have random noise. But you could also fail by not choosing the person in the exact middle of the population. This distinction is the part I’m least confident in, so somebody tell me if I’m right.)

But almost nobody correctly treats stimuli as a random effect! Suppose you make each subject look at 5 black and 5 white faces. You effectively have a stimulus sample size of 10. You’re generalizing that sample size to the entire rest of the population of white and black people. But low stimulus sample size could be bad. Suppose that in reality people are not racist, so looking at white faces and black faces will produce the same skin conductivity response. But by bad luck you used pictures of some really really ugly white people, who everybody instantly hates on sight. Once again, you will come to the wrong conclusion (the population is racist against whites). In order to avoid this, you need a big enough sample size of stimuli to prevent the idiosyncratic features of each individual stimulus from swamping the effect you want (difference between races). Unless you treat number of stimuli as a random effect, your model isn’t going to adjust for this possibility and won’t penalize you for low stimulus sample size.

In other words, most experimenters are analyzing their data using models with an implicit assumption that the sample size of stimuli is infinity. But the sample size of stimuli matters just as much as the sample size of subjects does. So they are making life much easier for themselves and getting lots of spurious positive results.

How many? According to Table 2 on page 58 of the linked paper, it depends on exactly what you’re doing, but it’s often a lot. If you’re doing the racism study above, and you have 90 subjects and 10 faces, and you incorrectly use the participants-random-stimuli-fixed analysis that most people use, you will get a false positive 60% of the time.

Everything after page 58 is basically “there are ways to fix this, but they’re really complicated”, and it’s beyond my pay grade.

Am I understanding this right?

Note I’m not a real stat person, nor have I read the linked paper so this is off the top of my head.

When random or fixed effects are mentioned, it implies the model is a hierarchical one, where you have some sort of grouping involved. In a 2 level model, you would have the top level explanatory variables here it would probably be the race of the subject, which should be a fixed effect since it’s assumed to have uniform effect in the two groups. (e.g. white people are more likely to be racists by beta_white, black people are more likely to be racists by beta_black). It’s also something you can’t directly observe, and can only be inferred from other observations, this case the difference in the average racistness between group of whites vs blacks.

Last when I looked, a stackoverflow post said that random effects and fixed effects have many definitions and wikipedia more or less imply the same thing. But one definition of random effect is that it is a realization of a random variable. So in the sense it is like random noise, in that it is not constant between different realizations. And here, we could say how racists a particular person is a realization of a random variable.

A different example given by Wikipedia for random effect model is one for scholastic performance of children, y_ij, where y is the performance of the j-th pupil at the i-th school. You have a school level random effect U_i, and a student level random effect W_ij. They also gave example of the case where you add the fixed effect of parental education, beta_parents and sex, beta_sex to create a mixed effect model.

An effect can also be random on one level, but fixed on a “lower” level. Notice how U_i, the school level effect is only indexed on i. If you only look at a particular school, then U_i is not really a realization of a random variable anymore and becomes “fixed” for all the students.

I think you don’t understand random and fixed effect models very well. What gave it away is that you believe “number of subjects” here is a explanatory variable in your model in the first place. It shouldn’t be. The penalization of the estimated effects would be from a wide CI/insignificant p-value. You are correct that under the random effect model you are assumed to sample from the population at random, so that there would be no systematic bias and sampling error would always be present as usual. But sampling error would affect mixed or fixed-effect models as well.

What fixed effect model does help with is “controlling for omitted variable bias due to unobserved heterogeneity when this heterogeneity is constant over time” according to the fixed effect model Wikipedia article. Basically, what it helps is producing a correct estimation of the effects of the explanatory variables in question, because the models we use almost certainly always “wrong” in that it doesn’t include enough variables. I really don’t believe fixed effect model help with small sample size directly here because small sample size doesn’t imply there is some sort of bias in the sampling procedure, so the estimation on expectation is still unbiased, even though the CI would be wide because of sampling error.

Yeh I don’t think “Unless you treat number of stimuli as a random effect, your model isn’t going to adjust for this possibility and won’t penalize you for low stimulus sample size” is the best way to think about this. In practice, including stimuli as a random factor in your mixed/hierarchical model means that you can model the effects of the individual stimuli (e.g. white face A, B, C… non-white face A, B, C…) rather than looking at the effect only of white vs non-white faces.

To look as a perhaps simpler and more common example where you do this. Say you are looking at the effect of an educational intervention on students’ educational outcomes, but you are looking at students in many different school populations where you’d expect the intervention to have different influences. Maybe some school populations are already better or worse, so they have different intercepts, and maybe some you’d expect the intervention to produce a more or less dramatic effect (different slopes). Rather than just looking at effect of intervention on outcomes across the whole population, you can model the ‘random’ effects within each school population. Note how this doesn’t assume anything about how representative your sample of schools is, usually no-one assumes your sample of schools is representative of all schools (unless you’re only interested in a single district, for example).

In the case of stimuli sampling you are doing much the same thing but looking at the random effect of individual faces. Here, if you want to say that you’ve thereby accounted for the problem of different faces having different effects, allowing you to better estimate the fixed white/non-white effect, you need to have a decently representative sample of stimuli. But that’s not something the model knows. You don’t give it any indication of how many stimuli out of a population of possible stimuli you sampled (often the size of the possible stimuli population will be infinite).  

(via slatestarscratchpad)

worldlypositions:

A friend paid me to go to a distant party with them, involving hours of transit. Feeling bad about the huge loss of value this probably constituted for someone or another, I tried to make the best of it by offering conversational advice and practice on the way there. Hopefully increasing the chance that they could happily go solo to parties in future.

I might have been hesitant to advise anyone on conversation, given my own frequent failures at it. But if someone is sufficiently uncomfortable at a party that they are willing to pay me to come, the bar is probably low. So I tried. Here is how the conversation went. (I very much welcome corrections from anyone who actually knows about this topic.)

One thing you can do is play the ‘figure out what the other person is interested in and direct the conversation toward that’ game. Every time the other person says a thing, you try to notice anything in it that might be a clue about what they actually care about. I have never managed to play this successfully, because I always forget that I am trying and end up just having a conversation. But I hear that it might be good.

We tried it. But an immediate difficulty seemed to be that in order to direct a conversation you had to at least implicitly recognize multiple possible next answers, so that you could take the better one. My friend mostly saw one option, so the conversation was entirely out of their control.

We practiced coming up with different next answers, from the same starting point. But somehow theirs all seemed weirdly non-specific to me. If I said I was modeling an intelligence explosion, they would ask ‘how is modeling an intelligence explosion going?’ or ‘When did you start modeling an intelligence explosion?’ or ‘Do you like modeling an intelligence explosion?’, whereas I might have asked ‘What about an intelligence explosion are you modeling?’ or ‘What kind of model?’ or ‘What are you trying to learn about an intelligence explosion?’ or maybe ‘What do you mean by an ‘intelligence explosion’?’, if I didn’t know. My friend’s questions treated ‘modeling an intelligence explosion’ as a sort of black box to be passed around, whereas I would be interested in taking the box apart. I thought I was right here, and objected.

My friend said that inquiring about models and intelligence explosions and such would indicate that they were uninformed or stupid. Presumably a smart person would hear ‘modeling an intelligence explosion’ and know just what it meant, and what the point was, and what kinds of models were likely to be involved.

I objected more—to me, jumping into a topic that someone else brings up, and asking pointed questions to quickly get to the bottom of what is going on and why, is a good indication of intelligence.

They thought this was a weird idiosyncrasy of people I hang out with, and most people wouldn’t be impressed.

I disagreed, but decided that more was at stake anyway. My own procedure of probing that generates questions about the inside of the black box seemed to really be something like ‘keep asking about whatever you are confused about until you are not confused’. And I claim that this is way better for being engaged than asking about the general situation of the black box, taking its contents to be out of bounds. And being engaged is great—for seeming intelligent and informed, for not having a terrible time, for not seeming rude or weird, and even for learning anything from the conversation or getting any kind of gratification from it. I claim. Probably being engaged is easier for almost everyone than me, so maybe this isn’t such a big deal for most. But it sounded like my friend was not very engaged at all, and mostly thinking about seeming okay to a conversation partner. 

Another nice thing about a procedure like this, I realize later, is that you don’t have to consciously locate multiple options to be driving the conversation in a particular direction. You just have to search for one option each time, using some sort of confusion-seeking mental state.

My friend agreed to practice my method. Which was more entertaining, but also more like an inquisition. I suggested they add some sort of confirmation that they received the other person’s message, and maybe even encouragement, before getting into the next question. For instance, ‘ah yeah’ or ‘that makes sense’. This distracted too much from composing questions, but we compromised on assenting where appropriate through body language, such as nodding.

I don’t think any of this was helpful at the party in the end, perhaps because there were so many obstacles to party success, or perhaps because it was bad advice. But thinking through some strategic details of conversation was interesting to me.

I also wonder whether perceiving more options in general gives you more freedom and thus more control and power—in conversation, and in life. It seems like it should in the abstract. But maybe it isn’t a big consideration next to others, or maybe it is a misunderstanding because you don’t really need to perceive options—you just need to somehow look in the right place for a good one.

Fwiw general conversational conventional wisdom is that you should (like your friend) ask vague open questions to give your conversational partner more space to talk about whatever they want, not lead into conversational dead ends once you’ve answered a more narrow specific question and not to seem interrogative.

slatestarscratchpad:

taymonbeal:

fnord888:

fnord888:

Daring move for a rationalist to say “your thought experiments are unrealistic”.

@taymonbeal

The problem is the way the thought experiments are being used. They’re not trying to elucidate edge cases in our ethical or political theories; they’re trying to paint a picture of horrific consequences and suggest that these will be realized in real life if the author’s disfavored ideology is allowed any influence. It’s like if we claimed that deontology actually causes people to be run over by trolleys.

I don’t think that it’s a reasonable to read the Current Affairs article as claiming that embracing libertarian ethics will cause there to be isolated cannibal villages that require enforce compliance with religious human sacrifice on pain of universal shunning. 

Rather, the isolated village thought experiment is specifically elucidating an edge case of the claim that there’s a meaningful difference between coercion by the state and coercion by private entities. 

Obviously, they think that situations very roughly analogous to that can happen in real life: that private coercion will be used to cause people harm, and perhaps cause unnecessary deaths. But that sort of rough analogy to reality is the exact sort of thing that we believe about our thought experiments, too. Eliezer Yudkowsky explicitly said that the Torture vs. Dust Specks thought experiment is examining the same kind of scope insensitivity that causes people to ignore existential risk.

Equally obviously, the social coercion in the isolated village thought experiment is more horrific, more extreme and disgusting than anything that’s likely to happen in real life. But that too is a principle that we embrace. Moral dilemmas extreme and disgusting precisely because those are the only cases in which we can make our intuitions strong enough to be clearly detectable.

On reflection, I don’t think that “your thought experiments are unrealistic” is Scott’s actual objection to that article. But I didn’t think the late su3su2u1′s actual objection to HPMOR was “Eliezer Yudkowsky is a funny-sounding name”, but I still thought it was bad form and bad messaging (and bad epistemic practice) that he peppered his reviews with references to “Hariezer Yudotter”. I feel that SSC post in question suffers, to some degree, from a similar sort of bad messaging.

Yeah, I wasn’t really thinking about the cannibal village; that’s clearly not realistic. But I do think Nathan Robinson wants his readers to believe that libertarianism will actually, in real life, lead to a world wherein poor people have no one defending their interests and can be harmed with impunity by rich people. (The story with the Almost Infinitely Rich Man could actually happen in this world-model; I think Robinson believes that abuses at least that bad would not be unheard-of and lesser but still life-ruining abuses would be omnipresent.) The problem is that he doesn’t actually argue for this conclusion, he merely illustrates it and relies on the horrifying image to carry the moral weight of his claim. I’m not necessarily defending everything in Scott’s response, but I think his explanations of/metaphors for why this isn’t a valid argument are basically right.

Yeah, in retrospect I regret writing that (and I didn’t even tag it as things I would regret writing! Such poor foresight!). I understand what he was trying to do and it’s something I’ve tried to do and defended in the past. I think he was wrong to weakman all libertarians as natural rights libertarians but really that deserved about a once-sentence criticism, and the sentence should have ended with “but I can understand why it’s tempting”. I guess just the repetition of “libertarianism is psychotic” in the Answers section pushed my buttons, and I thought of some amusing takes on it, and I have really poor ability to resist doing things that amuse me.

The “thought experiment” angle is a bit weird, because if you view libertarianism as a good idea rather than a moral law, then it matters whether it fails globally or in extremely contrived situations. For example, if you’re trying to figure out whether a car is good, “this car would melt if placed in the center of the sun” isn’t worthy of consideration, precisely because you don’t expect your car to end up in the center of the sun. Thought experiments make good philosophy and crappy politics. But again, this should have been countered with “I see you’re treating libertarianism as a philosophy when I treat it as a political system, which is wrong, although understandable”.

It’s not a factual mistake that I can put on my Mistakes page, but spiritually I think it probably belongs there.

I think the conclusion of the CA piece makes all the difference. If they had simply said “Here are 4 puzzles for libertarianism” (with no implication that they were insuperable problems that libertarian philosophies couldn’t account for) then it would have been unreasonable to pull them apart, expose their limitations and mock them. But as they essentially stated outright that there were no possible solutions to these problems for the libertarian, that they were damning for libertarianism and that they show that libertarianism is “psychotic,” I think your response was completely proportionate and appropriate (and I don’t mean that in a ‘they did a bad thing first, so it’s OK to do an equally bad thing back’- I mean I thought your response was probably the best way to communicate how wrongheaded their stance was (and better than a more neutrally toned descriptive piece saying “These are indeed puzzles for libertarians, as they are for other philosophies, though libertarians can easily respond to them, so they have little bite”). 

slatestarscratchpad:

I’ve been reading some Discourse about the film “Mother”, and the debate seems to be “It’s kind of crazy” vs. “It has artistic merit because it’s an allegory for the Bible”.

And…I guess I don’t get how being an allegory (am I using that term right? You know what I mean) is supposed to give it artistic value. It seems kind of like a gimmick, the way it was a gimmick for Stephenson to name his main character ‘Hiro Protagonist’. “If you’re mildly intelligent and think about this for a few seconds, you can see the thing we encoded here”. Okay, I wrote Unsong, I can’t pretend I don’t like that kind of thing. But as the basis of a whole movie?

I think maybe the claim is supposed to be that the Bible contains some timeless truth and so having it retold to us in an original way is…good? But if we’re religious, we already know the Bible, and if we’re atheist, we don’t believe it contains timeless truths anyway.

So really, what’s the point of making movies or books or cartoons or whatever that are secretly the Bible (or Greek mythology, or…), if we already know the Bible, and if it’s only going to give us about three seconds of mild amusement to “crack” the “secret code”?

A lot of people seem to have been criticising it for literally being pro abuse of women or for gratuitously having the woman abused and killed for no reason whatsoever  (e.g. https://www.youtube.com/watch?v=m4xi_dpogv8) in ways which are directly undercut if you recognise that it’s an allegory. 

Being an allegory isn’t particularly artistically impressive in itself, but it does defend against the criticism ‘this is just meaningless nonsense’ by establishing that it was, at the very least, an artistic work trying to convey some meaning through an artistic device.

slatestarscratchpad:

I feel like until about one year ago I almost never heard anything from “leftists” as opposed to “liberals”. EG if I heard someone making fun of “liberals”, I would assume with 99% certainty that they were a conservative. I would occasionally see communists, but they would be clearly labeled, not identify as “leftists”, and their outgroup would be “capitalists” and not “liberals” per se.

Now I encounter leftists who attack liberals all the time.

Have I just moved up in the world, or has there been a big change in leftists’ numbers/visibility?

Self styled ‘radical’ political theorists and activists have been criticising 'liberals’ for years. This terminology only just filtered down into the chattering class on social media, c.f. the delayed infiltration of 'intersectional.’

su3su2u1:

Saw some tumblr people talking about this movement.  My biggest problem with effective altruism is that most everyone I know who identifies them self as an effective altruist donates money to MIRI.  (its possible this is more a comment on the people I know than the effective altruism movement, I guess). Based on their output over the last decade, MIRI is primarily a fanfic and blog-post producing organization.  That seems like spending money on personal entertainment. 

I’m a big believer in donating to effective charities, but I think a lot of the focus on existential risk opens the door for variants of Pascal’s wager. 

 There is a 10^(-80) probability I can open a portal to dimension-X-Earth in my basement, if you give me $80,000 a year for the next decade.  We can colonize that extra-dimensional Earth, which will safeguard humanity in the event Earth gets destroyed, so extrapolating into the future thats at least pi*3^^^^^^^^^^3 utilons, so I’m the most effective charity you can donate to ever.  

The last EA Survey (2014) found that the charities most EAs were giving to were AMF (211), SCI (114), Give Directly (101), and only then MIRI (77). Global poverty seems to have substantially more support than AI/MIRI, as the question about which causes EAs support (Global poverty 579 vs AI 332)

https://eahub.org/sites/effectivealtruismhub.com/files/survey/2014/results-and-analysis.pdf

(via su3su2u1-deactivated20160226)