Since model building is the essence of science, this quote has a bit of a bite to it. It is from George E. P. Box (1919 – 2013), who was not only an eminent statistician but also an eminently quotable one. Another quote from him: One important idea is that science is a means whereby learning is achieved, not by mere theoretical speculation on the one hand, nor by the undirected accumulation of practical facts on the other, but rather by a motivated iteration between theory and practice. Thus he saw science as an iteration between observation and theory. And what is theory but the building of erroneous, or at least approximate, models?

To amplify that last comment: The main point of my philosophical musings is that science is the building of models for how the universe works; models constrained by observation and tested by their ability to make predictions for new observations, but models nonetheless. In this context, the above quote has significant implications for science. Models, even those of science, are by their very nature simplifications and as such are not one hundred per cent accurate. Consider the case of a map. Creating a 1:1 map is not only impractical[2] but even if you had one it would be one hundred per cent useless; just try folding a 1:1 scale map of Vancouver. A model with all the complexity of the original does not help us understand the original. Indeed the whole purpose of a model is to eliminate details that are not essential to the problem at hand.

By their very nature, numerical models are always approximate and this is probably what Box had in mind with his statement. One neglects small effects like the gravitational influence of a mosquito. Even as one begins computing, one makes numerical approximations, replacing integrals with sums or vise versa, derivatives with finite differences, etc. However, one wants to control errors and keep them to a minimum. Statistical analysis techniques, such as Box developed, help estimate and control errors.

To a large extent it is self-evident that models are approximate; so what? Again to quote George Box: Since all models are wrong the scientist cannot obtain a “correct” one by excessive elaboration. On the contrary following William of Occam he should seek an economical description of natural phenomena. Just as the ability to devise simple but evocative models is the signature of the great scientist so overelaboration and overparameterization is often the mark of mediocrity. What would he have thought of a model with twenty plus parameters, like the standard model of particle physics? His point is a valid one. All measurements have experimental errors. If your fit is perfect you are almost certainly fitting noise. Hence, adding more parameters to get a perfect fit is a fool’s errand. But even without experimental error, a large number of parameters frequently means something important has been missed. Has something been missed in the standard model of particle physics with its many parameters or is the universe really that complicated?

There is an even more basic reason all models are wrong. This goes back at least as far as Immanuel Kant (1724 – 1804). He made the distinction between observation of an object and the object in itself. One never has direct experience of things, the so-called noumenal world; what one experiences is the phenomenal world as conveyed to us by our senses. What we see is not even what has been recorded by the eye. The mind massages the raw observation into something it can understand; a useful but not necessarily accurate model of the world. Science then continues this process in a systematic manner to construct models to describe observations but not necessarily the underlying reality.

Despite being by definition at least partially wrong, models are frequently useful. The scale model map is useful to tourists trying to find their way around Vancouver or to a general plotting strategy for his next battle. But, if the maps are too far wrong the tourist will get lost and fall into False Creek and the general will go down in history as a failure. Similarly, the models for weather predictions are useful although they are certainly not a hundred per cent accurate. However, they do indicate when it safe to plan a picnic or cut the hay; provided they are right more than by chance and the standard model of particle physics, despite having many parameters and not including gravity, is a useful description of a wide range of observations. But to return to the main point, all models, even useful ones, are wrong because they are approximations and not even approximations to reality but to our observations of that reality. Where does that leave us? Well, let us save the last word for George Box: Remember that all models are wrong; the practical question is how wrong do they have to be to not be useful.

**To receive a notice of future posts follow me on Twitter: @musquod.**

[1] Hence the foolishness of talking about theoretical breakthroughs in science. All breakthroughs arise from pondering about observations and observations testing those ponderings.

[2] Not even Google could produce that.

Tags: Philosophy of science

The Ancient Greeks were aware of this. In their theories they tried to ‘save phenomena’, no matter the abstractness of the theory.

“Again to quote George Box: Since all models are wrong the scientist cannot obtain a “correct” one by excessive elaboration.”

That would only be true if there was an infinite number of options. Philosophy don’t know this, statistics don’t know this, but physics do: there is only a finite number of laws. (Hilbert spaces.)

Hey, if elimination worked for excellent empiricist Sherlock Holmes, it should work for science too, right!? =D

Agreed on over-determination though, since statistics knows this.

But not philosophy, there isn’t a similar testable rationale for Occam’s “economy” as it is for bad statistical procedures. I think the foolishness here is to talk about philosophy, when measure theory (and hence statistics) is the science of science.

Oops. I meant measurement theory. My bad, I have studied both and I cross-wired that. :-/

I don’t think “wrong” is equivalent to “incomplete”. I prefer to reserve the word “wrong” for cases where the model actually makes a false prediction; and it is quite possible that an incomplete model which “knows” its own limitations is never actually wrong. (Though of course we may never be certain that nature will not in future contradict such a model – even within whatever limits it has imposed on itself.)

@Tobjorn Larsson, I don’t understand your comment. Although it is obvious that the number of laws that will actually ever be written down by humans is finite, that is not true of the set of laws that could conceivably be written down. And although the “number” of separable complex Hilbert spaces is (up to isomorphism) actually 1 (if we exclude finite dimensional cases), the number of inequivalent operator-valued representations of physically relevant symmetry groups and/or observable algebras may well be infinite.