16.4 Interactions

  • You need 4 times the sample size to estimate an interaction that is the same size as the main effect.

  • Implies a big problem: If you don’t find the anticipated main effect and then look to the interactions to bail you out, you are likely looking at Type M / Type S errors!

  • More commonly, you arrange code variables so that the larger comparisons appear as main effects. So we expect interactions to be smaller. If they are half the size of the main effect, you would need 16 times the samples!

  • The message is to accept that you cannot attain solid estimates of the interactions and accept the uncertainty.