How to check for interaction effects in ANOVA

[#1]

Hi there, 

this might be a rookie question. Nevertheless: I have performed a 2x2x2 between subjects experiment. I am running ANOVA's to test for main effects, but I also want to test for interaction effects. But how to do that? For example, what is the interaction effect of treatment X and Y together and/or the effect of gender? Any advice for how to solve this would be highly appreciated.   

[#2]

Since you have three factors and assuming that you want to use a linear model, select Analyze > Fit Model. Select the column with the response data and click Y. Select the three columns with the factor levels, click Macros, and select Full Factorial. That's it! You might also want to change the Emphasis to Effect Leverage. Click Run.

Examine the Effect Tests report for evidence to decide about the main effects and the interaction effects.

[#3]

To add to my colleague @Mark_Bailey's counsel. I also recommend using JMP's signature visualization and sharing tools in this space. To present some striking visualizations if interactions are present I recommend supplementing all the analysis Mark suggests with:

From the JMP report window, response hot spot, select Factor Profiling -> Interaction Plots, and Factor Profiling -> Profiler. The dynamic aspect of the signature JMP Prediction Profiler is a great way to show interaction effects. Just drag the factor slider bar left and right for any one factor that plays a role in a two factor interaction...and watch what happens to the prediction trace of the other factor. For people that don't understand the practical implications of interactions this can really help you make your point.

Lastly, the entire report can be saved as an interactive HTML object or Web Report. Then the dynamic interactivity extends to the HTML object for the Prediction Profiler...and anyone with a web browser can open the object and get some of JMP's look and feel with the dynamic interactivity of the Prediction Profiler within the HTML object.

[#4]

So, Mark and Peter are the experts, so here's less complex answer.

Use fit model, do what Mark said.

In the Construct model effects box you'll have

all the main effects, all the two way interactions and the three way interaction, like this

X1

X2

X3

X1*X2

X1*X3

X2*X3

X1*X2*X3

add your response and click the run button and look for this table under the parameter estimates outline bar

Parameter Estimates

Term	Estimate	Std Error	t Ratio	Prob>|t|
Intercept	0.0403779	.	.	.
X1[L1]	-0.240686	.	.	.
X2[L1]	-0.220783	.	.	.
X1[L1]*X2[L1]	0.1526094	.	.	.
X3[L1]	-0.182747	.	.	.
X1[L1]*X3[L1]	0.1110772	.	.	.
X2[L1]*X3[L1]	0.3757836	.	.	.
X1[L1]*X2[L1]*X3[L1]	0.2184517	.	.	.

X1[L1]*X2[L1]. is the interaction between X1 and X2. the [L1] means level one. I got this because my 2x2x2 was with two level categorical (nominal) variables.  if I had used numeric continuous variables, then the parameter estimates table would look like this:

Parameter Estimates

Term	Estimate	Std Error	t Ratio	Prob>|t|
Intercept	-1.89227	1.155937	-1.64	0.3491
X1	0.4813715	0.436903	1.10	0.4692
X2	0.4415666	0.436903	1.01	0.4966
X3	0.3654941	0.436903	0.84	0.5565
(X1-1.5)*(X2-1.5)	0.6104377	0.873807	0.70	0.6118
(X1-1.5)*(X3-1.5)	0.4443089	0.873807	0.51	0.7005
(X2-1.5)*(X3-1.5)	1.5031345	0.873807	1.72	0.3352

I used either 1 or 2 for my levels, so 1.5 is the middle. (your data should look really different)

(X1-1.5)*(X2-1.5). is the X1, X2 interaction

hope this helps a little.

-B