# Understanding two-way repeated ANOVA: timing of intervention

I would like to perform a two-way repeated measures ANOVA test for my research. In my data, I have six time periods and two groups (intervention and control group) - I am interested in the interaction effect time*group. In the control group, there is no treatment at all and in the intervention group, the treatment takes place after period 3 (i.e. there are three periods before and three periods after the treatment).

I can run the model in SPSS, however I don't quite understand yet how the timing of the treatment influences this model/the interpretation. In the model, I do not specify when the treatment takes place, but surely it must make a difference when it takes place. Should I leave out the first two time periods in this case, so that I only have a status quo and three post-treatment observations? Or do I only need to take into account the timing of the treatment for the interpretation?

• What is "interaction effect time"? It seems like that is a constant, in which case, you can't really look at it. Jul 12 at 12:56

If so, you could treat the values for an individual's time periods 1 through 3 as technical replicates, and similarly for an individual's time periods 4 through 6. The time variable would then have only 2 levels, "before" and "after" period 3, that could be combined with the two treatment groups in a 2-way repeated-measures ANOVA.
If you care about possible trajectories in time before and after the intervention, however, then you need to be careful in how you model time. I would be simplest if you treated time as a categorical predictor with 6 levels. Then your hypothesis would presumably be that there are no time by treatment interactions at time = 3 or earlier (prior to treatment), but there is some such interaction at one or more later times. That could be evaluated by joint tests (e.g., Wald tests) on the corresponding sets of interaction coefficients.
• @clemens_s ANOVA is just a name for one type of linear regression model, in which the predictor variables are categorical (e.g., groups) and, in a multi-way design, in which you also evaluate how interactions among them are associated with a continuous outcome. I suspect that you could evaluate your data with either a DiD design or by doing post-modeling tests on the time:treatment interaction coefficients, and get pretty much the same results.