2  From question to model

ImportantStatus of this chapter

This chapter is not yet fully written. What follows is a plan, with its references. There is still a lot to develop here.

The plan is my own, not a summary of any other work. It does not reproduce the content of the corresponding chapter in Da Silva et al. (2022), which I did not consult while writing it.

2.1 Why a chapter about questions in a book about code

No amount of statistical sophistication rescues a badly posed question. The right order is: question → design → data → model → inference (one could write a whole book about that process). Tools like Julia act on the last two steps, and are of little use if the first two (question and sampling design) are not worked out with care.

This chapter exists to join the book’s two halves: the question you ask determines the model structure that appears in Chapter 8 and Chapter 9.

2.2 Multiple working hypotheses

This section will be built on Chamberlin (1890), which argued for holding several rival hypotheses in play at once rather than defending a single one — an 1890 argument that remains the best available description of what separates a productive question from a defensive one.

To develop:

  • Observation, pattern and process: the three levels, and why conflating them yields models that answer a different question than the one asked.
  • Mechanistic versus phenomenological hypotheses, and what each licenses you to conclude.

2.3 Design and dependence structure

This section will be built on Hurlbert (1984), which named pseudoreplication and catalogued its forms, and on Zuur et al. (2009), whose treatment of dependence structure connects directly to the random effects used in this book.

To develop:

  • Sampling unit, experimental unit, and why the distinction sets the denominator of your test.
  • How design structure translates into (1 | group) terms — the direct link to
  • Effect size and power before collection, not after.

2.4 Computation-limited questions

A section specific to this book. Some ecological questions are limited by computation rather than by data: null models with many permutations, individual-based simulation, fitting mechanistic models by optimization. These benefit most from the language, for the reasons discussed in Section 1.1.

This section needs to be written with runnable examples, like the rest of the book — not with unverified performance claims.

Further reading

  • Chamberlin (1890) on multiple working hypotheses.
  • Hurlbert (1984) on pseudoreplication and experimental design.
  • Zuur et al. (2009), chapter 1, on design and dependence structure.
  • Da Silva et al. (2022), for the same material treated in R.

Exercises

  1. Write your project’s central question in one sentence, without statistical jargon.
  2. Sketch the model it implies before looking at the data: what response, what predictors, what grouping structure?
  3. Identify the experimental unit in your design. How many true replicates do you have — not how many measurements?