Lingyan,

> I have two factors, a and b, each has two levels (a1,a2 and b1,b2). In total I have four conditions (a1b1,a1b2,a2b1,a2b2).
> In each of these conditions, I test 9 different touch frequencies stimulation

Actually your experiment seems to be a 2 x 2 x 9 design. The specifications for the two main effects look fine, but the one for interaction

+a1b1_f1 +a1b1_f2 ... +a1b1_f9 -a1b2_f1 -a1b2_f2 ... -a1b2_f9 +a2b1_f1 +a2b1_f2 ... +a2b1_f9 -a2b2_f1 -a2b2_f2 ... -a2b2_f9

should be

+a1b1_f1 +a1b1_f2 ... +a1b1_f9 -a1b2_f1 -a1b2_f2 ... -a1b2_f9 -a2b1_f1 -a2b1_f2 ... -a2b1_f9 -a2b2_f1 -a2b2_f2 ... -a2b2_f9

> I saw in some other posts that these analyses were done within each frequency, combined with the \ sign

That would be a different inference. For example,

+A \
+B

is intended to set up a null hypothesis

H₀: A = 0 and B = 0

and to find out whether there is strong evidence for at least one of the effects.

> I can also add a coefficient in front of each condition like 0.0056.

You could use a different weight (e.g., 0.0556) other than 1, and its impact is only on the interpretability of the effect estimate (beta), not the statistic value. Use 0.0556 if you care about the interpretability of the effect estimate.

Gang