[myfonj]

That final example with sugar-y (so to speak) grouped headers contradicts prior definition of "normalized table". First "normalized" example explicitly shows that grouping happens exclusively by using text repetition in adjacent headers. Row headers should then read

    Treatment, Baseline
    Treatment, Contraceptive, 2 weeks
    Treatment, Contraceptive, 12 weeks
    Treatment, Contraceptive, 20 weeks
    Treatment, Recovery phase

and column headers should read

    Hormones, FSH [IU/L]
    Hormones, LH [IU/L]
    Hormones, Testosterone [nmol/L]

Arguably, when there is single top-level header group in some axis (the "evil (0,0) cell"), it might be better used as table caption (here probably something like "Treatment - Hormones [dosage]") and then omitted in header cells.

(From (HTML) accessibility perspective, all tables should better have caption element describing them.)

[lakesare]

It does violate the perfect "Cartesian coordinates" normalization I described earlier, however I consider this is a permissible step up over the full normalization, because, while cutting down repetition (entirely), it still suggests a zero-ambiguity way to read off the values from the header cells.