**Phase-Shifting Effects of Melatonin Were Sufficient to Test the PSH.** PM melatonin caused a 0.89-h phase advance in the clock time^{}of the DLMO (*t* = 7.61, *df* = 21, *P* < 0.001) and a 0.69-h increase^{}in PAD (advance in the DLMO with respect to midsleep: *t* = 4.66,^{}*df* = 21, *P* < 0.001). (After PM-melatonin treatment, mean^{}DLMO clock time was 20:18 ± 0:14.) AM melatonin did not^{}significantly delay the DLMO (0.18 h) or decrease PAD (0.01^{}h). Placebo treatment was associated with nonsignificant trends,^{}an advance in the DLMO and an increase in PAD of both 0.17 h^{}(*t* = 1.71, *df* = 23, *P* = 0.10; and *t* = 1.76, *df* = 23, *P* = 0.09,^{}respectively), consistent with the photoperiod that lengthened^{}throughout the study. Nevertheless, because AM melatonin and^{}placebo treatments did not produce statistically significant^{}changes in circadian phase, these treatment groups were appropriately^{}collapsed in one of the analyses below (see Fig. 6). Shifts^{}in sleep times were mostly small and statistically insignificant^{}and were in a direction consistent with the treatment and the^{}constraints on bedtimes: the largest statistically significant^{}shift was an 18-min delay in sleep onset after AM melatonin^{}(*t* = 2.57, *df* = 21, *P* = 0.02). Baseline Structured Interview^{}Guide for the Hamilton Depression Rating Scale: Seasonal Affective^{}Disorder Version (SIGH-SAD) (33) ratings (27, 28, and 28, respectively)^{}were the same for the placebo, AM- and PM-melatonin treatment^{}groups; there were no significant pretreatment vs. posttreatment^{}percent change differences [ANOVA: *F* (2, 66) = 1.65, *P* = 0.20;^{}and Kruskal-Wallis *H* test: ^{2} = 4.09, *df* = 2, *P* = 0.13] among^{}the three treatment groups.^{}

**Pretreatment Phase Typing of SAD Patients and Its Implications.** In Fig. 2, the statistically significant parabola [*R*^{2} = 0.17,^{}*df* = (2, 65), *P* = 0.003] has a minimum at 5.88, which validates^{}the choice of PAD 6 (see Fig. 1) for phase typing these subjects^{}before doing change-score and treatment-response analyses. Furthermore,^{}in the present data set, neither parabolic nor absolute deviation^{}linear plots from the parabolic minimum were statistically significant^{}when any other circadian marker comprising the DLMO and/or sleep^{}times was substituted for PAD. Even the two constituents of^{}the PAD (DLMO and midsleep clock times) had nonsignificant parabolic^{}correlations [*R*^{2} = 0.006, *F* (2, 65) = 0.21, *P* = 0.81; and *R*^{2}^{}= 0.05, *F* (2, 65) = 1.64, *P* = 0.20, respectively] with depression^{}ratings.^{}

^{}

The implicit phase typing done above was explicitly done before^{}conducting all of the remaining analyses. Those who at baseline^{}had PADs 6 (*n* = 48; 71%) were designated as phase-delayed types,^{}and those who had PADs > 6 (*n* = 20; 29%) were considered^{}phase-advanced. (Unless otherwise specified, the use of the^{}terms advanced and delayed applies to these pretreatment assignments^{}and not to posttreatment phase.) With one exception (see Fig. 5),^{}when the advanced and delayed groups are separated in all of^{}the correlational analyses, statistical significance is found^{}only in the delayed group, and analyses confined to this group^{}are almost always more robust than in those with the advanced^{}and delayed groups combined.^{}

**Response to Treatment Correlates with Correcting Circadian Misalignment.** The day after treatment was discontinued, posttreatment (Fig. 3),^{}the statistically significant parabolic relationship remained^{}between depression score and PAD [*R*^{2} = 0.11, *df* = (2, 65), *P*^{}= 0.02]; a second parabola fitted to the data of the subset^{}of (delayed) subjects with pretreatment PADs 6 was also significant^{}[*R*^{2} = 0.19; *df* = (2, 45); *P* = 0.009; minimum = 5.85]. This group^{}is analyzed in more detail in Fig. 4.^{}

^{}

Linear regressions restricted to the delayed group appear in^{}Fig. 4, as well as the parabola for delayed subjects treated^{}with PM melatonin. The regression that did not exclude subjects^{}with posttreatment PADs > 6 (overshifters) did not quite^{}reach statistical significance (*r* = –0.27, *r*^{2} = 0.07,^{}*df* = 46, *P* = 0.07; however, Spearman's = –0.33, *n* = 48,^{}and *P* = 0.02 and Kendall's = –0.20, *n* = 48, and *P* = 0.04^{}were statistically significant). [These statistics are reported^{}for comparison with those of the parabolic curves presented^{}in Figs. 3 and 4 (in all of the above analyses, statistical^{}significance was lost when the data fitted to a significant^{}parabolic curve were fitted to a linear regression).] If the^{}overshifters and undershifters (those who remained PAD 6) are^{}analyzed separately, the linear regressions were statistically^{}significant, despite the reduction in sample size (respectively:^{}*r* = 0.72, *r*^{2} = 0.52, *df* = 9, *P* = 0.01; and *r* = –0.44,^{}*r*^{2} = 0.19, *df* = 35, *P* = 0.007). Thus, the data in Fig. 4 confirm^{}the therapeutic window of about PAD 6, at least for the phase-delayed^{}group: indeed, the parabolic fit of the data for the delayed^{}subjects who received PM melatonin, the treatment that caused^{}the greatest phase shifts, is remarkably impressive, particularly^{}for such a small sample [*R*^{2} = 0.65, *df* = (2, 8), *P* = 0.01, minimum^{}= 5.56].^{}

As mentioned above, analyses of the delayed group, rather than^{}the advanced group, appear to be the driving force behind most^{}of the statistically significant findings. This finding also^{}applies to the linear regression (raw data not shown) of pretreatment^{}to posttreatment percent changes in depression scores vs. shifts^{}toward or away from PAD 6 (for delayed subjects, *r* = 0.35, *r*^{2}^{}= 0.12, *df* = 46, *P* = 0.01; and for all subjects, *r* = 0.32, *r*^{2}^{}= 0.10, *df* = 66, *P* = 0.007); the three delayed subjects who^{}worsened the most received the "incorrect" treatment (AM melatonin),^{}which caused a phase delay away from PAD 6 [it is not surprising^{}that only a few subjects actually became more depressed during^{}the study, given the small, but significant, antidepressant^{}response to placebo (see Fig. 6)]. The plot for this type of^{}analysis appears in Fig. 5 for all phase-advanced and phase-delayed^{}subjects who received PM melatonin, the treatment that caused^{}the greatest phase shifts: the linear regression between percent^{}change in depression ratings and shifts toward or away from^{}PAD 6 was quite robust, despite the relatively small sample^{}size (*r* = 0.59, *r*^{2} = 0.35, *df* = 20, *P* = 0.004). {Furthermore,^{}the data in this figure make clear that most of the advanced^{}types (who, in response to PM melatonin, usually advanced further^{}away from PAD 6) had a worse clinical response than the delayed^{}types (for whom PM melatonin would be expected to be the treatment^{}of choice); notably, the subject who shifted away from PAD 6^{}more than anyone else and whose depression scores worsened more^{}than those of all but two other subjects [phase-delayed types^{}who also received the incorrect treatment (AM-melatonin); data^{}not shown] was from the advanced subgroup for whom PM melatonin^{}would have been predicted to be the incorrect treatment.} The^{}*r*^{2} of 0.35 is the largest found in the combined group of phase-advanced^{}and phase-delayed subjects (second only to the *R*^{2} of 0.65 in^{}the PM-treated delayed group alone in Fig. 4).^{}

^{}

Subjects were retrospectively (and blindly) assigned to correct^{}vs. incorrect treatments: PM melatonin is the correct treatment^{}for delayed types (*n* = 11) and AM melatonin for advanced types^{}(*n* = 6); AM melatonin is the incorrect treatment for delayed^{}types (*n* = 16) and PM melatonin for advanced types (*n* = 11).^{}Accordingly, 17 subjects received the correct treatment and^{}27 the incorrect one (24 received placebo). The correct treatment^{}decreased depression ratings by 34%, compared with 13–15%^{}for the other treatment groups, separately or combined (Fig. 6).^{}Two ways of calculating effect sizes were considered (Fig. 6);^{}the more conservative ones were based on percent differences^{}in change scores compared with the correct treatment: 0.61 (incorrect^{}treatment), 0.83 (placebo), and 0.69 (the latter two groups^{}combined).^{}