🚀 go-pugleaf

I share the concern that modelling should be seen by students or
"casual" users of statistical analysis as another area where black-box
or sausage machine cookbook analyses can be applied. It is, I agree,
far more difficult to persuade people to apply their brains to data
and ask sensible questions in clear language - they think it's
anti-academic.

I addressed this problem in a conference paper (The Statistician 1986
Vol 35 pp 199-206) "Data Analysis: the need for models?" I distinguish
between models that are purely descriptive and serve to reduce a body
of data to a set of informative facts, models that are heuristic and
reflect a theoretical basis, and models that are applied to make standard
comparisons so that the same analysis will be used without confirming in
every case that it is optimal for each individual dataset.

The philosophical problem I see in model fitting is that you are always
simultaneously using the particular data to justify the model and also
the model to interpret the data. This is the distinction behind likelihood
and probability. I suspect that most users of stat methods see themselves
as using a tool to "extract the truth implicit in their data", rather than
looking for support for a hypothesis that was generated in their mind
before the data were collected.