You are viewing the historical archive of The Philosophy Forum.
For current discussions, visit the live forum.
Go to live forum

Factor Analysis and Realism

Marchesk November 15, 2016 at 03:38 11950 views 32 comments
In statistics, the idea behind doing a factor analysis (FA) is that there are latent variables, or factors, which are unobservables that explain the observed variables from some data set. For example, if you gave school children a bunch of assessment tests, then you end up with a bunch of data for each assessment (variable). But what you want to know is what factors explain the resulting data. The factors could be income, IQ, parental involvement, classroom size, working memory, teaching approach, etc.

If you can reduce the observed variables (assessments in the example above) to several factors that can be used to reproduce the data, and then show which variables are associated with which factors, you have demonstrated there are indeed latent (unobserved) variables responsible for the data. Of course what the latent variables actually are is a matter of theory, and thus factor analysis can be confirmatory as well as exploratory.

The philosophical question is whether FA demonstrates that realism (of whatever sort) is the case. Why would factor analysis work unless there really are unobservables explaining the results of observations? Is there another way of explaining how the statistics works out?

Comments (32)

Marchesk November 15, 2016 at 03:46 #32938
A short way of putting this is from the realist POV is that factor analysis should not work for actual data, unless there are real unobservables explaining that data. It can't simply be a useful statistical model, although the theory of what factors are doing the work is a human model. The anti-realist would need to explain why the factor analysis appears to work with real world data without resorting to unobservables as the explanation (they're useful fictions in the model for anti-realists, not the reason factor analysis works).
Wayfarer November 15, 2016 at 05:34 #32962
Reply to Marchesk How are the 'latent variables' different from 'possibilities'?

There are 'real possibilities'. Like, if you wanted to model some process, you could factor in variables that you know might effect the outcome, and run the simulation. 'Real possibilities' might be heat, pressure, and so on. Other things, like pixie dust or psychic forces, wouldn't need to be considered because they're not in 'the realm of possibility'.
Marchesk November 15, 2016 at 06:04 #32964
Reply to Wayfarer I believe possibility is an epistemic situation when we don't know for sure what the factors are. It's possible that working memory is a factor in explaining the assessment results. But whether it actually is requires further confirmation.

The question is why there would be an actuality. Why would working memory or any latent variable exist, if the anti-realist view is the correct one?
Wayfarer November 15, 2016 at 06:31 #32966
Reply to Marchesk All I'm asking is, whether another name for what you have called 'unobservables' is 'possibilities'. Isn't that what factor analysis is trying to determine?i.e. there's a range of possible factors, in factor analysis you're trying to assign them probabilities.
Marchesk November 15, 2016 at 07:05 #32969
Reply to Wayfarer Regardless, why does it work is the question. FA assumes there are such possibilities to be discovered. The math is based on that assumption.
Wayfarer November 15, 2016 at 07:21 #32972
Reply to Marchesk The reason it works is because there are real possibilities; there are things that really might happen, and other things that never will. The philosophical question is: what does it mean to say that 'possibilities exist?'
Marchesk November 15, 2016 at 07:27 #32973
Reply to Wayfarer This isn't about what might happen. Your already have your data set. The theory is that there are hidden factors explaining your data. You might be measuring one hundred different things, but they can be reduced to a dozen factors that explain those hundred measurements.

So if we want to explain why kids succeed or fail in school, and we measure a bunch of things, then we want to be able to reduce our data to what explains success in school. And then we can act on that (assuming an ideal world absent politics).

Now, on a possible anti-realist view of things, if there are no hidden factors, then some kids succeed and some fail, and there are a bunch of observations we can make. But there is nothing beyond that explaining the success and failure. Nothing beyond what we measure. This is in direct contradiction to what factor analysis assumes to be the case. There are things we cannot measure, so we have to resort to a statistical analysis to tease them out.
Wayfarer November 15, 2016 at 07:45 #32974
Marchesk:Now, on a possible anti-realist view of things, if there are no hidden factors, then some kids succeed and some fail, and there are a bunch of observations we can make. But there is nothing beyond that explaining the success and failure. Nothing beyond what we measure.


Like 'instrumentalism' in physics?

IN any case, to answer your question in the OP, I think factor analysis does try and identify 'real causes' on the basis of trends, so I presume it would generally favour a realist attitude.
dukkha November 15, 2016 at 09:45 #32981
This strikes me as the same sort of argument as "because we experience reflections, something must exist which is being reflected". I would say the mistake you are making in this argument is that there are anti realist ways of understanding the existence of unobservable variables, and 'things which are being reflected'. Your argument strikes me as somewhat question begging in that it only works if you already assume a realist understanding of unobserved things. Just take a cup for example, I don't think you'll find many antirealists who think that only the surface facing them exists, because that is the only part which is visually perceived. As in the cup is merely an outer half of a curved cylinder, and has no back. The anti realist still takes the cup to be a whole object, he just analyses the meaning of "whole object" in an experiential way, in opposition to the realist who thinks the back of the cup not perceived still exists because say there's atoms there or something.

I forget who said this but some phenomenologist wrote this idea that when one sees an object, although in terms of visual perception all one sees is a surface facing oneself, in a more mental sense one builds the three dimensionality of the object within ones mind by imagining the object as being seen from all directions, a gods eye view.

I don't think an antirealist would just outright deny the existence of unobserved variables, he'd just analyse the way in which they exist as being experiential.
Babbeus November 15, 2016 at 11:05 #32984
Reply to Marchesk Can you show a concrete example with some specific observable and unobservable variables?
Terrapin Station November 15, 2016 at 13:22 #33002
Quoting Marchesk
Why would factor analysis work unless there really are unobservables explaining the results of observations?


The easiest objection to this is the following: Why couldn't idealism posit ideas that aren't explicit/made manifest?

Or are you simply asking whether there "really are" unobservables? (That would be different than saying that it supports realism in a philosophical sense--realism in the philosophical sense refers to mind-independent existents. What I'm noting above is that the unobservables could be mind-dependent instead (to an idealist, at least))

Re asking simply whether there really are unobservables, regardless of whether they're mind-indepedent or not, I don't see why one couldn't claim that there are not, and that factor analysis is simply an instrumental means of dealing with manifest data that are otherwise difficult to explain in some systematic way, where the instrumental account posits fictions in this case.

The Great Whatever November 15, 2016 at 14:47 #33008
Dispositions to perception or subjective uncertainty in perceptual outcome.
Brainglitch November 15, 2016 at 16:25 #33018
Reply to dukkha
Exactly.

Reply to Marchesk
Latent variables are inferred, a mental construct, part of a mental model. Thus, consistent with an anti-realist metaphysics.
andrewk November 15, 2016 at 23:33 #33095
Quoting Marchesk
The philosophical question is whether FA demonstrates that realism (of whatever sort) is the case.
It's not a philosophical question until a particular factor analysis has been performed. Then one can have a philosophical discussion about the interpretation and implications of the results.

You have not specified any factor analysis, nor even indicated what sets of observed values should be used in the factor analysis, let alone performed one, so I can't see that there is, so far at least, any philosophical issue to discuss.

Also, you seem to be using an unusual meaning for 'unobservables'. My understanding of the role of unobservables is that they are well-understood concepts that cannot be directly quantified, and for which proxies are used. For example IQ test results are used as a proxy for intelligence, or life expectancy may be used as a proxy for quality of life.

Terrapin Station November 16, 2016 at 12:34 #33210
Quoting andrewk
My understanding of the role of unobservables is that they are well-understood concepts that cannot be directly quantified,


What would you call phenomena that we're not even aware of, so that there's no concept of it, etc.?

(Maybe you're implying that you're an idealist and not a realist? I don't recall your view there, although I suppose I should assume that you're an idealist, because realists seem to be by far in the minority on this board.)

Michael November 16, 2016 at 15:04 #33217
Quoting Terrapin Station
What would you call phenomena that we're not even aware of, so that there's no concept of it, etc.?


Well, assuming that there are such things, I'd assume they'd be called "unobserved and unconceived phenomena".
Terrapin Station November 16, 2016 at 15:19 #33222
Quoting Michael
I'd assume they'd be called "unobserved
That's the "unobservable" part.

Marchesk November 16, 2016 at 15:36 #33226
Quoting Michael
Well, assuming that there are such things, I'd assume they'd be called "unobserved and unconceived phenomena".


Why add "unconceived" to the mix? If it's not observed, then it's an unobservable, period. We might conceive of it, but for one reason or another, we don't observe it. They are two entirely different things.
Marchesk November 16, 2016 at 15:38 #33227
Quoting andrewk
My understanding of the role of unobservables is that they are well-understood concepts that cannot be directly quantified, and for which proxies are used. For example IQ test results are used as a proxy for intelligence, or life expectancy may be used as a proxy for quality of life.


Perhaps often this is the case, but there is exploratory factor analysis, where you might not know which concepts play the role of the factors. You can take any data set with related variables and do an exploratory factor analysis on it.
Marchesk November 16, 2016 at 15:40 #33229
Quoting Brainglitch
Latent variables are inferred, a mental construct, part of a mental model. Thus, consistent with an anti-realist metaphysics.


Yes, in the model of doing factor analysis. But the fact that it works suggest something more. You can't use the mental construct, as I mentioned above, to explain why the model works on real data.

IOW, this isn't just a mathematical concept. It's use to get at unobserved factors in real data. That's the reason statisticians came up with it. The theory being that there really are such things explaining the data.
Marchesk November 16, 2016 at 15:42 #33231
Quoting Terrapin Station
Or are you simply asking whether there "really are" unobservables?


That's stated as the motivation for the people who invented factor analysis.
Marchesk November 16, 2016 at 15:43 #33232
Quoting Babbeus
Can you show a concrete example with some specific observable and unobservable variables?


I'm just learning how to use it. You can do a search for factor analysis studies. There's plenty of scientific results and papers.
Marchesk November 16, 2016 at 15:49 #33233
I can tie this into another thread where the OP asked about aliens.

Kris Kelvin arrives aboard Solaris Station, a scientific research station hovering near the oceanic surface of the planet Solaris. The scientists there have studied the planet and its ocean for many decades, a scientific discipline known as Solaristics, which over the years has degenerated to simply observing, recording and categorizing the complex phenomena that occur upon the surface of the ocean.

https://en.wikipedia.org/wiki/Solaris_(novel)#Plot_summary


In this story, scientists have gathered their data on the Solarian ocean. However:

Thus far, they have only compiled an elaborate nomenclature of the phenomena — yet do not understand what such activities really mean.

https://en.wikipedia.org/wiki/Solaris_(novel)#Plot_summary


If anti-realism were the case, I would expect that factor analysis would never get farther than what the scientists in Lem's story have gotten with the alien ocean. Scientists and statisticians collect a bunch of data on a wide variety of phenomena which they classify accordingly, but no hidden factors can be inferred to make sense of it.
Terrapin Station November 16, 2016 at 15:58 #33234
Quoting Marchesk
That's stated as the motivation for the people who invented factor analysis.


You understand the distinction between that and realism ("Are they real?") in the "mind-independent" philosophical sense, though, right?

Those are two different questions.

You could see the issue in three stages of questions about whether they're "real:"

(1) We can treat them simply as instrumental utilities where it doesn't matter if they're real in any sense beyond being useful for the theories in question (and one might say that's "real enough"),

(2) Whether they're real in the colloquial sense of "whether they occur" versus just being useful fictions, but where no particular ontological commitment is required regarding just what sorts of occurrent things they are, and where thus the whole range of possibilities is available, and

(3) Whether they're real in the sense of mind-independent things, where to answer that positively, we are making an ontological commitment (at least in one aspect) to what sorts of things they are.
Marchesk November 16, 2016 at 16:05 #33235
Quoting Terrapin Station
(1) We can treat them simply as instrumental utilities where it doesn't matter if they're real in any sense beyond being useful for the theories in question (and one might say that's "real enough"),


Except for when it comes to school funding or medical treatments. Then you might want to know whether they are just instrumental utilities or actually real.
Terrapin Station November 16, 2016 at 16:10 #33237
Reply to Marchesk

Well, yeah, depending on the specific example and context, you might want to not just settle on stage (1) or (2) of that. I'm just noting that the question of whether unobservables are real can be taken and answered in different ways.
Marchesk November 16, 2016 at 16:13 #33238
Quoting Terrapin Station
Well, yeah, depending on the specific example and context, you might want to not just settle on stage (1) or (2) of that. I'm just noting that the question of whether unobservables are real can be taken and answered in different ways.


Yes, especially in theoretical physics.
Terrapin Station November 16, 2016 at 16:18 #33241
Quoting Marchesk
Yes, especially in theoretical physics.


Yeah, it's most common there to just stop at (A) . . . which is something I brought up just today in the recent MWI thread. Most physicists (as well as mathematicians for that matter) don't care much, in most contexts, about any ontological commitments. They're just looking for instrumental theories that work well. That's the gauge of truth for them.
Brainglitch November 16, 2016 at 16:33 #33243
Quoting Marchesk
Yes, in the model of doing factor analysis. But the fact that it works suggest something more. You can't use the mental construct, as I mentioned above, to explain why the model works on real data.

IOW, this isn't just a mathematical concept. It's use to get at unobserved factors in real data. That's the reason statisticians came up with it. The theory being that there really are such things explaining the data.


Seems to me that if you can't demonstrate, either logically or empirically, that the unobserved factors are real things rather than ideas of things, then they are just as consistent with an anti-realist as with a realist stance.
Michael November 16, 2016 at 18:19 #33252
Quoting Terrapin Station
That's the "unobservable" part.


There's a difference between being unobserved and being unobservable.

Quoting Marchesk
Why add "unconceived" to the mix?


Because Terrapin asked "what would you call phenomena that we're not even aware of, so that there's no concept of it[?]"

Quoting Marchesk
If it's not observed, then it's an unobservable, period.


A thing can be unobserved but observable, so this is false. That we don't see it is not that it can't be seen.
Terrapin Station November 16, 2016 at 18:26 #33254
Quoting Michael
There's a difference between being unobserved and being unobservable.


There can be, sure. "Unobservable" is sometimes used to denote that something is unobservable in principle. But it's also often used to simply denote what we could call "unobserveds," where it was a contingent matter that they weren't observed on that occasion.
andrewk November 16, 2016 at 21:27 #33281
Quoting Terrapin Station
What would you call phenomena that we're not even aware of, so that there's no concept of it, etc.?

To get a fix on that we can turn to that famous non-idealist philosopher Donald Rumsfeld (I'm deliberately misusing the word 'idealist' here, but why not, it's a Thursday after all):

Those phenomena are the 'unknown unknowns'. By contrast I think that the unobservables referred to in the OP are, in Rumsfeldian terms, the 'known unknowns'.