Jump to content

Science vs. Pseudoscience


Altherion

Recommended Posts

On 9/27/2018 at 12:09 AM, DMC said:

Anyway, so if the math isn't wrong what's your beef again?  We don't predict enough?  Or we don't predict accurately enough?  Cuz if that's all it is, yeah I'm done.

You make predictions about a system that is far too complex to be analyzed with the data you have so your predictions are either mostly useless, mostly inaccurate or both. Worse, they're usually politicized so you've got half the country angry at you (not without cause) and a substantial fraction of them doesn't care to make distinctions between real science and the rest.

On 9/27/2018 at 12:20 AM, DMC said:

Heh.  Yeah, "taste" is the appropriate adjective.  The first link is encapsulated by the first week of Methods 2010, and the second the first couple of weeks of 2020.  Hashtag not impressed.  

ETA:  Like seriously dude, that second link goes over OLS and confidence intervals for a ridiculous amount of pages.  Basic regression is not that hard to learn.  Because one of my faculty members is a dick, I've done it by hand, and that doesn't take as long as reading all that shit.

:lol::lmao:It's not long -- it's actually really short given the material it covers (as I said, it's a summary). Many of the statements it makes are theorems and if one were to really study this, there would be proofs (the simpler ones are left as exercises, the rest in the text). It's interesting and useful to occasionally do these things by hand, but it's not a substitute for actually understanding all of the derivations. I've seen courses like this taught as well as taught a similar course and despite the undergrads being fairly bright, it took much longer than a couple of weeks (probably a bit more than a couple of months) and this is not counting the machine learning stuff (any one of neural networks, boosted decision trees and the rest listed there easily takes over a month to get right in and of itself).

On 9/27/2018 at 12:20 AM, DMC said:

Do you know probit and logit?  Or a GLM?  Or even a GLLAMM model?

We don't use those names, but yes, I've used those two link functions before and if you mean generalized linear model, yes, I've dealt with it before -- although never for a paper because it's in between linear regression (which at least has simplicity going for it) and the far more general non-linear techniques.

On 9/27/2018 at 12:20 AM, DMC said:

How bout matching methods?  Which strategies would you prefer?

Think about it: under what circumstances would you need matching methods in physics or chemistry?

Link to comment
Share on other sites

On 9/27/2018 at 4:07 AM, OldGimletEye said:

Well this depends largely on what your idea of "science" is.

For me, its about building a model and rejecting the model if data or real world events seemingly make it false. Or at least selecting the model that seems to explain events better.
To a certain extent, I'm sympathetic to some your arguments. The fact of the matter is the field of macro has a lot of work to do to build better models. For instance, I'm not exactly happy with the standard New Keynesian model believing we need a better description of the price adjustment process than relying on the "Calvo Fairy", though it is certainly a hell of lot better at getting closer to the truth than the RBC model. Anyone that knows about the Volcker period at the  FED should know that.

Your idea of science is mostly consistent with mine except that I would add a caveat: the model must eventually converge to something that is mostly stable and agreed upon by the vast majority of the field's practitioners. It's not clear to me that this is what is happening in macroeconomics: Keynesianism was ascendant between WWII and the 1970s, then it fell out of favor for a few decades, now something similar has come back but there's a bunch of competing ideas... are we really selecting a model that explains things better or just jumping from one model to the next depending on what the most recent data shows?

Regarding neo-liberalism and identity politics: perhaps you can start a different thread? I don't think they have much to do with the this topic (or at least there are dozens more topics that are equally related).

Link to comment
Share on other sites

On 9/29/2018 at 4:40 PM, Altherion said:

You make predictions about a system that is far too complex to be analyzed with the data you have so your predictions are either mostly useless, mostly inaccurate or both. Worse, they're usually politicized so you've got half the country angry at you (not without cause) and a substantial fraction of them doesn't care to make distinctions between real science and the rest.

Anyone that tells me my results are "usually politicized" can go fuck themselves.  Seriously.  Also, I'm not supposed to generate theories and hypotheses about a system because it is "far too complex" for the current data?  So we should just give up?  Inspiring!

On 9/29/2018 at 4:40 PM, Altherion said:

:lol::lmao:It's not long -- it's actually really short given the material it covers (as I said, it's a summary).

I don't know what's up with the emojis.  If you don't realize it's laughably long you're necessarily laughably ill-equipped regarding the requisite statistics as well.

On 9/29/2018 at 4:40 PM, Altherion said:

there would be proofs (the simpler ones are left as exercises, the rest in the text).

There are frequently proofs in political science.  That's one of the easiest things to shit out.

On 9/29/2018 at 4:40 PM, Altherion said:

but it's not a substitute for actually understanding all of the derivations.

Ha!  That's rich.  As if you know all the derivations, or even what you're talking about.

On 9/29/2018 at 4:40 PM, Altherion said:

I've seen courses like this taught as well as taught a similar course and despite the undergrads being fairly bright, it took much longer than a couple of weeks (probably a bit more than a couple of months) and this is not counting the machine learning stuff (any one of neural networks, boosted decision trees and the rest listed there easily takes over a month to get right in and of itself).

Um...congrats?

On 9/29/2018 at 4:40 PM, Altherion said:

We don't use those names, but yes, I've used those two link functions before and if you mean generalized linear model, yes, I've dealt with it before -- although never for a paper because it's in between linear regression (which at least has simplicity going for it) and the far more general non-linear techniques.

What A Joke.  GLM indeed means generalized linear model, but you referring to it as a "more general" non-linear technique again simply demonstrates it's just another thing you don't have to learn so you don't really understand.  So like those immigrants your boy is scared of.

On 9/29/2018 at 4:40 PM, Altherion said:

Think about it: under what circumstances would you need matching methods in physics or chemistry?

Wait, first off, I thought physics was stupid too?  Second, my dad's a fairly prolific biochemist.  He has no fucking idea about the statistical methods I do.  Why?  Because he just has to show basic treatment effects and he'll get published.  But should he be employing more advanced methods that came around way after he was trained?  Yeah, probably.  So, to answer your stupid question - yes, maybe, fuck if you or I know.

Link to comment
Share on other sites

On 9/29/2018 at 10:40 PM, Altherion said:

...

Think about it: under what circumstances would you need matching methods in physics or chemistry?

Every single time you do an experiment, which is viable because our systems are trivially simple compared to any involving human societies. Of course the terminology is different, and the statistics involved a lot more simple.

For example you can't measure a spectrum of a compound if you don't know the properties of solvents, gasses, equipment involved. Which you all have to account for. 

Link to comment
Share on other sites

18 hours ago, DMC said:

Also, I'm not supposed to generate theories and hypotheses about a system because it is "far too complex" for the current data?  So we should just give up?

No, but if you were to be honest about what you can generate from the current data, your results would be far more limited. You'd probably still get something, it just wouldn't be as exciting.

18 hours ago, DMC said:

If you don't realize it's laughably long you're necessarily laughably ill-equipped regarding the requisite statistics as well.

If you don't realize it's short, you probably know very little about math in general. The next statement leads me to believe this is the case:

18 hours ago, DMC said:

There are frequently proofs in political science.  That's one of the easiest things to shit out.

Not political science proofs (for which I'll take your word), math proofs -- and they're not trivial for most of the statistical statements in that link.

18 hours ago, DMC said:

GLM indeed means generalized linear model, but you referring to it as a "more general" non-linear technique again simply demonstrates it's just another thing you don't have to learn so you don't really understand.

You have a tendency to misread what I wrote that's pronounced enough for this to happen nearly an order of magnitude more than for anyone else on this forum so either you're doing it on purpose or your ability to read is on par with your understanding of math. Here's what I actually wrote: "it's in between linear regression (which at least has simplicity going for it) and the far more general non-linear techniques." There is no way to read this as "referring to it as a "more general" non-linear technique". If you're still having trouble understanding, here are the three in order of complexity:

1) Linear regression. Not always applicable, but simple.

2) The generalized linear model.

3) The non-linear techniques (e.g. neural networks). More general and usually more powerful than either of the two above, but complex.

Link to comment
Share on other sites

52 minutes ago, Altherion said:

No, but if you were to be honest about what you can generate from the current data, your results would be far more limited. You'd probably still get something, it just wouldn't be as exciting.

So..your argument is political science journal articles are exciting?  M'kay.  There is something to be said for publication bias, which motivates people to overemphasize (and in the worst cases manufacture) significant relationships.  I could agree with that.  But you're asserting there's some political motivation.  Which is absolute horseshit.

1 hour ago, Altherion said:

If you don't realize it's short, you probably know very little about math in general. The next statement leads me to believe this is the case:

19 hours ago, DMC said:

There are frequently proofs in political science.  That's one of the easiest things to shit out.

Not political science proofs (for which I'll take your word), math proofs -- and they're not trivial for most of the statistical statements in that link.

HAHA!!  Man that is so hilarious.  Political science proofs?  What do you think we're writing?  You gotta be pretty ignorant to think there's any type of specific "political science" proofs.  What do you think we're doing when there's entire journals devoted to formal modeling? 

Just wow, still can't stop laughing.  What a non-credible statement that is.  If you think we don't frequently write out the logic - which is exactly what I meant in that quote - you are the one that has revealed himself as having no idea what he's talking about.  About social sciences and "math."  I love just using the word "math," it's so douchy.

1 hour ago, Altherion said:

1) Linear regression. Not always applicable, but simple.

2) The generalized linear model.

3) The non-linear techniques (e.g. neural networks). More general and usually more powerful than either of the two above, but complex.

This again displays zero understanding why GLM or non-linear models or all of it was developed.  It was precisely for the bitching that you're making.  It's to fix measurement or observational error.  Will there ever be causality definitively determined in my field?  No, probably not, I'm one of the few that will admit to it.  But you're acting like getting better is pointless, as if building knowledge through valid designs and well-executed results is somehow pointless.  And that only the normal, natural, hard, whatever sciences should be worthy of such rigor and analysis.  Your entire point is an absurd notion that any self-respecting scientist or epistemologist would reject out of hand.  And speaking of political motives, it's quite obvious this is just because of your consistently plain grudge against academics.

Link to comment
Share on other sites

On 10/2/2018 at 10:55 PM, Altherion said:

You have a tendency to misread what I wrote that's pronounced enough for this to happen nearly an order of magnitude more than for anyone else on this forum so either you're doing it on purpose or your ability to read is on par with your understanding of math. Here's what I actually wrote: "it's in between linear regression (which at least has simplicity going for it) and the far more general non-linear techniques." There is no way to read this as "referring to it as a "more general" non-linear technique". If you're still having trouble understanding, here are the three in order of complexity:

1) Linear regression. Not always applicable, but simple.

2) The generalized linear model.

3) The non-linear techniques (e.g. neural networks). More general and usually more powerful than either of the two above, but complex.

Linear regression is the specific case of GLMs for continuous outcome variables, while logistic regression uses a "logit" link function to conduct analyses for dichotomous outcome variables. I don't know that these are different orders of "complexity" or sophistication, since the more complicated the model, the harder the interpretations become. Neural networks in particular are just weird models used in machine learning and data mining, but they tend to defy ready interpretations, insofar as the "hidden units" don't really mean much at all as a model of reality. I guess you have some experience in this, but from a statistician's perspective, they're all just "wrong" models that are sometimes useful. And even though I'd only ever be 95% confident of that. 

Link to comment
Share on other sites

1 hour ago, Aemon Stark said:

Linear regression is the specific case of GLMs for continuous outcome variables, while logistic regression uses a "logit" link function to conduct analyses for dichotomous outcome variables. I don't know that these are different orders of "complexity" or sophistication, since the more complicated the model, the harder the interpretations become.

Aye, GLM was developed to merge a logit link to OLS.  I disagree the interpretations become "harder."  It's a switch to MLE, which is easily understandable and functionally the same type of statistical tests.

Link to comment
Share on other sites

On 10/3/2018 at 10:35 PM, Aemon Stark said:

Linear regression is the specific case of GLMs for continuous outcome variables, while logistic regression uses a "logit" link function to conduct analyses for dichotomous outcome variables. I don't know that these are different orders of "complexity" or sophistication, since the more complicated the model, the harder the interpretations become. Neural networks in particular are just weird models used in machine learning and data mining, but they tend to defy ready interpretations, insofar as the "hidden units" don't really mean much at all as a model of reality. I guess you have some experience in this, but from a statistician's perspective, they're all just "wrong" models that are sometimes useful. And even though I'd only ever be 95% confident of that. 

Yes, this is the reason the machine learning methods are less popular than they'd otherwise be. They almost always appear to squeeze more out of the data than any of the classical methods, but it's hard to convince people (or even one's self) that what they're doing makes sense. Fortunately, if something is real, it's usually possible to either build something out of it or at least observe it from a dozen different angles and always see what you expect.

Link to comment
Share on other sites

Post-Modernism/post-structuralism (PM/PS) can be a useful tool in helping to understand various ideas.  Deconstruction of the ideas helps see the pieces from which the idea came to be.  However, when PM/PS becomes more than just a tool.  When it becomes the framework for everything or the lens through which all ideas are viewed it becomes a problem for the hard sciences in particular.  

It turns objective reality into a world of solipsistic crazy talk in the context of the hard sciences.

Reality is not subjective.  If PM/PS is really making inroads into the hard sciences as this article suggests we have some problems:

https://quillette.com/2018/10/01/the-grievance-studies-scandal-five-academics-respond/

Link to comment
Share on other sites

4 hours ago, Ser Scot A Ellison said:

Post-Modernism/post-structuralism (PM/PS) can be a useful tool in helping to understand various ideas.  Deconstruction of the ideas helps see the pieces from which the idea came to be.  However, when PM/PS becomes more than just a tool.  When it becomes the framework for everything or the lens through which all ideas are viewed it becomes a problem for the hard sciences in particular.  

It turns objective reality into a world of solipsistic crazy talk in the context of the hard sciences.

Reality is not subjective.  If PM/PS is really making inroads into the hard sciences as this article suggests we have some problems:

https://quillette.com/2018/10/01/the-grievance-studies-scandal-five-academics-respond/

Those are not the hard sciences though, those are the soft ones. Where all the lenses, all the theories we have are subjective by definition. Because the systems they study are far too complex to comprehend without making assumptions.

Link to comment
Share on other sites

3 minutes ago, Seli said:

Those are not the hard sciences though, those are the soft ones. Where all the lenses, all the theories we have are subjective by definition. Because the systems they study are far too complex to comprehend without making assumptions.

They submitted the fake articles to journals for the soft sciences but one of the discussions mentions PM/PS making inroads into hard science discussions too.  Can we agree that when studying reality PM/PS is of extremely limited utility because if we assume reality is subjective we've fallen into solipsism?

Link to comment
Share on other sites

4 minutes ago, Ser Scot A Ellison said:

They submitted the fake articles to journals for the soft sciences but one of the discussions mentions PM/PS making inroads into hard science discussions too.  Can we agree that when studying reality PM/PS is of extremely limited utility because if we assume reality is subjective we've fallen into solipsism?

At its core postmodernism seems to be using the approach of the hard sciences on everything else, so I don't worry at all about it being implemented there. We already strive to understand our biases and try to understand in which we and our expectations form our experiments and the results we look at.

 

Link to comment
Share on other sites

5 hours ago, Ser Scot A Ellison said:

They submitted the fake articles to journals for the soft sciences but one of the discussions mentions PM/PS making inroads into hard science discussions too.  Can we agree that when studying reality PM/PS is of extremely limited utility because if we assume reality is subjective we've fallen into solipsism?

Care to explain that one?  Reality might not be subjective, but our experience of it is.  Even measuring it and observing it is.  In many cases it can be difficult to study something without altering it, and not just in social sciences (even quantum physics if I understand correctly).  Is there really much of a movement on the hard sciences arguing that reality itself is subjective?  

 

Link to comment
Share on other sites

2 hours ago, larrytheimp said:

Care to explain that one?  Reality might not be subjective, but our experience of it is.  Even measuring it and observing it is.  In many cases it can be difficult to study something without altering it, and not just in social sciences (even quantum physics if I understand correctly).  Is there really much of a movement on the hard sciences arguing that reality itself is subjective?  

 

The moon is there whether we are looking at it or not.  That argument push us toward a solipsistic worldview were nothing can be taken as true that isn’t personally experienced.  It makes flat Earthism a valid point of view.

Link to comment
Share on other sites

7 minutes ago, Ser Scot A Ellison said:

The moon is there whether we are looking at it or not.  That argument push us toward a solipsistic worldview were nothing can be taken as true that isn’t personally experienced.  It makes flat Earthism a valid point of view.

That seems like an extremely hyperbolic slippery slope argument, but ymmv.  

Link to comment
Share on other sites

Archived

This topic is now archived and is closed to further replies.

×
×
  • Create New...