3
$\begingroup$

My situation is as follows: I have built a 2-level Plackett-Burman design with 10 factors (some are actual 0/1-variables, the others are numeric and I used the minimum and the maximum) and 32 experiments. The experiments are done, the results are at hand and the resulting (linear) models are ok but not quite satisfying.

So far those models only contain main effects and no interactions, because in the current design with 10 factors in 32 experiments the interactions are highly correlated und therefore using them in the models would not give meaningful results (right?).

The assumption now is that it's probably the interactions, which are missing to make satisfying models, and to explore them I need to do more experiments. But what kind of design should I use here? My best guess so far is extending the 2-level design to a (face-centered) central composite design, because it's said to be for "exploring non-linear effects". So is this the way to go or are there better methods?

edit: context is a plc controlled machine in which we are trying to detect whether or not the outcome of a production cycle (i.e. the product) will be satisfying or not. So the factors we are changing in the experiments are different machine settings and differences in the processed workpieces that influence the production process and the outcome.

$\endgroup$
1
$\begingroup$

That might be a good way to go, but (depending on context you did not tell) it might be wise to run the new points as a new block$^\dagger$. Another way of extending the design might be to use software for optimal design, also called algorithmic design (D-optimality, search this site).

Then you would have to define a set of candidate points, and the software will choose among them. Going that way, be sure to use more than two experimental points (among the candidate points) for the continuous variables (maybe the extremes + at least the center point), and quadratic effects + interactions in the model, so you get a design that can actually detect nonlinearities.

$^\dagger$This means to make a categorical variable (factor, in R-speak) with values old run, new run (possibly more) and include that in the analysis. That accounts for possible change in conditions between the runs.

| cite | improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ You're rigth, context might be helpful. I edited the question. What exactly do you mean with "run the new points as a new block"? Not reuse the already done experiments in the central composite design? Why? D-optimal plans ware something I also considered but scaped because I read you need to know the excact model you want to use beforehand. This is not the case since we are trying out different kinds of models (linear regression, random forests, ...) and also we are building different models for different target variables (i.e. different characteristics of the outcome). $\endgroup$ – NoThanks93330 Jul 27 at 8:37
  • $\begingroup$ Please see my edits $\endgroup$ – kjetil b halvorsen Jul 27 at 14:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.