> as long as all possible linear combination of all rows can cover any
> contrast of any two conditions, and these rows are orthogonal to
> each other, then the f-statistics is fine?

You can set up the rows so that they are orthogonal with each other, but it's not really necessary for the tests you're interested.

> For example, if I have 4 levels, then a matrix as follows would be equivalent
> to an ANOVA on individual level, right?
> 1 -1 0 0
> 0 1 -1 0
> 0 0 1 -1

Yes, that's correct.

> I mainly used -gltsym in the past. If using -glt and specifying matrix, is the
> order of each regressor the same as the order that 3dDeconvolve
> parameters are written?

No, don't use -glt. Instead continue to use -gltsym. Just use the weights we just discussed in -gltsym. See the specific usage here:

[afni.nimh.nih.gov]

Gang