2

Is there a rule for selecting which experimental design one should carry out? I am new to statistics/ED and sometimes I find myself asking which design is best to use. My general guidelines have been (please correct if wrong):

  • Use a Plackett Burman (PB) for screening the number of variables
  • Use a fractional factorial if you have a large number of variables and cost/money is a factor as well as estimating effects
  • Use an orthogonal design if your variables have different number of levels
  • Use a full factorial if you do not have a large number of variables

Another sort of guideline I have is from here. I would like to have some input from members on their approach on this and if they do things differently.

EDIT: As per @Scortchi advice. Assuming you have a study which has 12 variables and another which 6 variables both with 3 levels for each variable. What type of design would you choose for each and why? The way i would approach this is by maybe screening the variables first as a full factorial would be too many cases to run then apply a full or fractional factorial. However I would then ask well why not do a orthogonal straight away. From the link i linked to earlier the objective of the design is a consideration so in the context of obtaining an optimal fit if you have +5 you would screen and then run a design, but why not run an orthogonal straight away. Hopefully this may be a bit clearer (apologies if the terminology I am using isn't correct)

John
  • 121
  • 2
  • 3
    Welcome to Cross Validated! This is a rather broad question - whole books have been written to answer it (see https://stats.stackexchange.com/questions/1815/recommended-books-on-experiment-design for some of them). Perhaps you could give a specific example, explaining what designs you're considering & what you're unsure about. – Scortchi - Reinstate Monica Jun 27 '17 at 13:31

1 Answers1

1

Your approach may be good as a start, until you get more experience, at least for your application area (industry/processes?). But it is impossible to reduce experimental design to such a set of rules, especially if you want them to cover all application areas! So you should strive for an understanding where you base design on a few important principles, such as

  1. replication

  2. blocking

  3. factorial design/fractional factorials

  4. orthogonality

  5. confounding

then you will be able to construct a design for your needs. Then find a good book from the many mentioned here: Recommended books on experiment design?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467