All Models Are Wrong Some Models Are Useful

All models are wrong; some models are useful.

-- generally attributed to the statistician George Box


See AllAbstractionsLie, UsefulLie, MentalModel, TheMapIsNotTheTerritory.


Nonsense. Given that a model is, by definition, a simplified representation of "the real world," then it must be possible to have models that are "right." The key question is, "wrong" or "right" for what specific purpose? If I develop a model that captures enough of a problem domain to help me build good software that does what my users need it to do, then that model was not just "useful," it was also "right" for the purpose.

The saying AMAWSMAU resonates with me, so I could say that this particular saying is right for my purposes. For you, the saying AMAWSMAU doesn't resonate, so it is not right for your purposes.

I interpret "wrong" as "not = real world". You seem to be saying that, for you, the word "model" includes the concept of simplified/approximate/not = real world, so the word "wrong" is redundant.

Yes, but I'm certainly not the only one who interprets model that way. Look it up in any textbook. Here's what I find in my Webster's New World Dictionary (2nd Edition, 1978) under model: "1.a) a small copy or imitation of an existing object, as a ship, building, etc., made to scale; b) a preliminary representation of something, serving as the plan from which the final, usually larger, object is to be constructed; c) same as ARCHETYPE (sense 1); d) a hypothetical or stylized representation, as of an atom..." and so on. And that's just the generic dictionary definition. The various books by Booch that I've read, and the UML texts, make it crystal clear that each kind of model only captures some particular aspect of reality; each one has strengths and weaknesses. You need many, good models to even hope to understand a problem domain well enough to build a successful system. And should any question or doubt arise, you should always be ready to go back to the users and get them to explain their reality for you. I don't think any proponent of UML or other models would claim that models are all you need to build good software. But models can be very useful tools.

How does "All models are models; some models are useful" resonate with you?

Better, yes, but "models are models" sounds like a truism -- obvious and pointless. I suppose we could say "All models are incomplete representations of reality; some models are useful." Wordy, but more accurate...

Hmmm. Sounds promising. I guess "All models are incomplete; some models are useful" works for me as well as the original, and apparently is getting a little closer to your point of view, too.

Hmmm, hmmm. Or maybe "All models are lacking; some models are useful". I kind of like the double meaning. But I think I see a pattern. Putting trigger words like "wrong" or "lie" or "bad" in cute little sayings is not always helpful.

It's just about setting expectations, nothing more. In the context in which a model was developed it is likely to be useful. Waterfall model was. In subsequent contexts, it gets subjected to a whole different range of tests, and inevitably found defective, which leads its disappointed users to assert "The model is wrong!". This may lead to throwing out the model, flipping the bozo bit on the model, ignoring the model for all future cases. Well, this is what new generations do, so we shouldn't expect too much improvement in this area. But pointing out that while a model may fail for the current purpose, it's still useful for its original purpose may be part of a sensible conservatism that helps us leverage history into intelligence. I like AMAWSMAU, even though I know it's "wrong". -- WaldenMathews

What Walden describes is pretty much a perfect characterization of a KuhnParadigmShift.

How does "Don'tEatTheMenu" resonate with you? I was introduced to that phrase in the writings of RobertAntonWilson (so it probably originated elsewhere :-)). He has written at great length about the dangers of confusing the map with the territory, especially in science.

[How about: AllModelsAreApproximationsSomeModelsAreUseful??]


One of the reasons AMAWSMAU resonates with me is that it gives me a definition of BetterOrWorse to use when comparing two models. In other words, when comparing two models, choose the one that is more useful "for your specific purpose" (as suggested above). Don't necessarily choose the one that is more complete or more consistent or closer to the real world (unless, of course, that is really what you need).

See also SpecializationSweetSpot.


When did AMAWSMAU originate?

From http://www.norvig.com/quotations.html
 All models are wrong. Some models are useful, - George Box 

See also http://www.skymark.com/resources/leaders/box.asp

There's a brief discussion of non-Euclidian geometry in chapter 22 of ZenAndTheArtOfMotorcycleMaintenance (1974), which says:

"As well ask whether the metric system is true and the avoirdupois system is false; whether Cartesian coordinates are true and polar coordinates are false. One geometry can not be more true than another; it can only be more convenient. Geometry is not true, it is advantageous."

He attributes this idea to Jules Henri Poincare (1854-1912)


I've heard this as "All theories are wrong, but some theories are useful". I take it to mean that engineers aren't too concerned if a theory is scientifically exact, if it's close enough for AllPracticalPurposes.


WittgensteinsLadder is an example of models that are useful even though they're wrong.


It appears that the title is a deliberate over-simplification designed to provoke a reaction and subsequent discussion. It has certainly achieved that. Perhaps a more accurate, less controversial version might be

"All models are inaccurate in some aspects, some models are useful despite that, and for a given purpose some models are more useful than others."


To attempt to alter the original is to lose its power and simplicity. Attempts to "explain" it tend to lose content. One must try to underatand the original point. "All models are wrong" is a warning to expert and beginner alike that no mater how carefully or completely we model a system, the model will always be less than the reality that is modelled. To lose track of this truth is to risk deluding onesself. To substitute lesser terms than "wrong" waters down the key warning. "Some models are useful" conveys the truth that, despite the fact that all models fall short of reality, it is sometimes but not always possible to obtain useful information about reality by studying a model. The term "some" also conveys the warning that there are models which are not useful. George Box would probably have agreed that most serious attempts at a model are useful to some extent, but his succinct expression makes the point far more clearly than saying eg "Most models are useful to some extent". RM


CategoryLearningMethods

EditText of this page (last edited July 9, 2012) or FindPage with title or text search