In essence the researchers are hoping to test two separate hypotheses (described below in the form of research questions). Therefore, the proposed analysis has been outlined according to each hypothesis.

Hypothesis 1: Is low-dose quetiapine (50 mg/day) more effective for patients with depression who have insomnia at treatment baseline? (Stated differently: is low-dose quetiapine 50 mg/day effective as monotherapy for patients with depression regardless of whether or not they have insomnia at baseline?).

Hypothesis 2: Is high-dose quetiapine (150 - 300 mg/day) more effective for patients presenting with melancholic depression at treatment baseline?

Response rates for even the most well-established antidepressants are poor, rarely exceeding 30%. This is likely due to fact that depression is a heterogeneous disorder. Using the DSM-5, two patients diagnosed with depression may not share a single symptom. Antidepressants also vary considerably in their mechanism of action, and research has shown that antidepressants also differ in their ability to treat certain symptoms. Thus, there is a great need to understand which patients respond to which antidepressants.

Quetiapine is commonly used to treat major depression and insomnia, although there is a commonly held view that sedation underlies its antidepressant effect. Unlike most antidepressants that bind to one or two receptor sites, quetiapine has a broad mechanism of action and binds to over a dozen neurotransmitter receptors, with varying affinities, at clinically relevant doses. Because of this, quetiapine has the ability to produce different effects at different doses, in which the relative occupancy at neurotransmitter receptor sites is changes accordingly. This is evidenced by the broad dosage range for quetiapine that varies based on the disorder being treated (e.g., higher doses for schizophrenia than major depression).

At low doses (e.g., 50 mg/day) quetiapine has more affinity for histamine and adrenergic receptors, the blocking of which is presumed to result in sedation. These are the doses in which quetiapine is commonly used as a sleep aid. At higher doses (150-300 mg/day) quetiapine has greater affinity for serotonin receptors and produces a greater inhibition of norepinephrine reuptake, which is believed to underlie its antidepressant effect, than it does at lower doses at which it is more likely to bind to histamine and adrenergic receptors. Not surprisingly, doses of 150-300 mg/day are recommended for the treatment of depression. It is puzzling, however, that in one trial (NCT00320268; D1448C00001, Moonstone) 50 mg/day of quetiapine was found to be just as effective as 150 and 300 mg/day as a monotherapy for patients with major depression. Although there is evidence that the antidepressant effect of high-dose quetiapine is independent of its sedating effect, to the researchers knowledge no study has adequately assessed specifically if low-dose quetiapine is an effective antidepressant in patients with depression not currently experiencing insomnia.

High-dose quetiapine does not have a broad-spectrum antidepressant effect, but appears to be most effective at treating certain depressive symptoms, including lack of pleasure/interest, guilt/pessimistic thoughts, reduced sleep, reduced appetite, and anxiety/tension, which commonly co-exist in a depressive subtype known as melancholic depression. This is not surprising because melancholic depression is well-known to be responsive to medications that increase serotonin and norepinephrine neurotransmission, compared to selective serotonergic antidepressants. To the researchers knowledge, no study has examined if patients with melancholic depression at baseline are more responsive to quetiapine, although the fact that quetiapine seems to be most effective at treating symptoms common in melancholic depression, as well as the fact that quetiapine has an active metabolite nor-quetiapine with strong affinity for inhibiting norepinephrine reuptake, suggests this may be the case. Given that insomnia is considered a symptom of melancholic depression, it would also be pertinent to demonstrate that the effect of quetiapine on melancholic depression is not due to its sedating effect.

Given the known lag-time between antidepressant initiation and response (roughly 4-6 weeks), trial-and-error prescribing is an inevitably lengthy process. A better understanding of predictors of response to medications such as quetiapine will lead to more timely and effective treatment of patients with depression, as well as reduced financial burden of diseases to society as a whole.

Past research has demonstrated that the antidepressant effect of high-dose quetiapine (150-300 mg/day) is independent of its sleep-inducing properties (Trivedi et al., 2013), although to the researchers knowledge no study has adequately addressed this question using only low-dose (50 mg/day) quetiapine. To test this hypothesis the researchers are using data from 366 participants in D1448C00001 (50 mg/day arm [n = 182] and placebo arm [n = 184] only; Weisler et al., 2009) because as far as the researchers know it is the only quetiapine XR monotherapy trial for major depression that used doses of 50 mg/day.

The first step in the analysis for Hypothesis 1 will involve classifying participants into two groups: depression with insomnia and depression without insomnia. There should be a balanced number of patients in each group who received quetiapine or placebo, as randomization should have balanced interventions between any groups defined post-hoc.

The presence of insomnia will be determined based on responses to the baseline HAMD-17** and MADRS** according to the following criteria:

(a) Both of:

1. Reduced sleep - MADRS Q 4 (A 2-6)
2. Insomnia - HAMD-17 Q 4 (A 2) or Q 5 (A 2) or Q6 (A 2)

The researchers definition of insomnia is based on two gold-standard instruments designed to assess the DSM criterion A4 "insomnia" required for the diagnosis of a major depressive episode. The researchers definition was made deliberately broad to capture all participants with reduced sleep in contrast to those with normal sleep or hypersomnia. It is unlikely that this definition would miss any participants experiencing impaired sleep at baseline. Furthermore, participants meeting either of these criteria would be judged to be experiencing insomnia as a symptom of depression in most other research and clinical settings. The researchers have requested for this definition that participants have impaired sleep on both the MADRS and HAMD-17 so that there is agreement between participants' self-reported insomnia (MADRS) and clinician-rated insomnia (HAMD-17). This also makes the results potentially more relevant to a clinical setting in which clinicians assess the presence of depressive symptoms including insomnia using simple questions with face validity such as those on the HAMD-17, and do not typically rely solely on patient self-report questionnaires.

Based on past research (YòÈargòn et al., 1997), it is expected that up to approximately 20% of participants could have experienced hypersomnia at baseline, and others (up to 20%) will have had normal sleep patterns. Therefore one could reasonably predict anywhere between 60 and 80% of participants would meet the researchers definition of "insomnia." Should the researchers insomnia grouping variable be overly inclusive to the point that more than 90% of participants are in the "insomnia" group, or there are no responders in the "no insomnia" group, then the researchers will utilize an alternate definition that requires the MADRS Q4 score to be 4 or higher. Should the categories still be unbalanced the researchers will create a new grouping variable that defines the presence of a "high sleep disturbance" in contrast to "low sleep disturbance." The presence of a high sleep disturbance will be defined as a score of 4 or greater on the summed HAMD-17 Questions 4, 5, and 6 (summed scores range from 0-6) as well as a score of 4 or greater on MADRS Question 4 (range is 0 to 6). This definition is identical to a "high sleep disturbance" definition used in past research (Trivedi et al., 2013) although the researchers have added the stipulation that the baseline MADRS Q4 score be 4 or higher so that there is consistency between the self-report and clinical-rated assessments. It is unlikely that more than 90% of participants will be classified as having a "high sleep disturbance" based on this definition (Trivedi et al., 2013) and thus represents a potentially more balanced grouping variable. This new "high sleep disturbance" variable would only need to be used in one portion of the analysis (Analysis A, see below) as an alternate to the "insomnia" grouping variable.

Should the researchers not be able to determine an adequate "insomnia" or "high sleep disturbance" grouping variable unfortunately they will not be able to proceed with the corresponding analysis (Analysis A, as described below).

The second step in the analysis for Hypothesis 1 will involve calculating the study end points. Two primary end points will be calculated:

1. MADRS response rates, defined as a 50% score reduction from baseline
2. Modified MADRS response rates, defined as a 50% score reduction from baseline calculated without Item 4 (reduced sleep).

The test of Hypothesis 1 with modified MADRS response rates (Analysis B, see below) will only be conducted if the Cronbach alpha value for the modified MADRS at baseline is at least .80 or within .10 of the unmodified MADRS alpha and still acceptable by standard convention (at least .70). Given the hypothesis tests proposed (see below), response rates will be calculated at each time point from baseline until and including the end of Week 6.

The third step in the analysis for Hypothesis 1 will involve hypothesis testing.

First the researchers will compare MADRS response rates between participants with depression and insomnia and participants with depression and no insomnia (alternate comparison, high sleep disturbance vs low sleep disturbance) who were randomly assigned to received quetiapine 50 mg/day (Analysis A). The researchers are trying to establish that participants with insomnia had a larger response than those without insomnia, and therefore no comparison against placebo is needed. The outcome of interest will be the presence of a positive MADRS response at any time point after baseline (i.e., Day 4, Week 1, Week 2, Week 4, and Week 6). Although this analysis strategy is not equivalent to a formal interaction test of a subgroup treatment effect, which would inappropriately underpowered in this sample, the researchers feel that it is appropriate to provide preliminary support for a subgroup effect that could be evaluated with future research.

It is assumed that 80% of the sample will be in the insomnia group and 20% in the no insomnia group, and that 50% of participants in either group will have received quetiapine.

If the researchers conducted a one-sided two-sample proportion z-test to evaluate the prediction (insomnia response % > no insomnia response %) with an alpha value of .05, it would be able to detect a proportion difference of .22 with 80% statistical power (power analysis conducted with G*Power, http://www.gpower.hhu.de/en.html). This corresponds to a Cohen h effect size of .45, which is "small" bordering on "medium" in size by convention. Considering the placebo response rate was 30%, while there is no consensus on what represents a clinically meaningful subgroup effect, the researchers would consider a subgroup response rate encompassing an additional 22% to be clinically meaningful and worthy of reporting (especially in an area of medicine where response rates are typically < 50%). Furthermore, the researchers would have adequate statistical power to detect an effect size that is neither overly large nor trivial in size (i.e., within the small-to-medium range) and thus the test is appropriately powered.

To show that the antidepressant effect of quetiapine 50 mg/day was not due to its sleep-inducing properties, the researchers will conduct a one-sided test of proportions comparing modified MADRS response rates between the quetiapine 50 mg/day and placebo groups (Analysis B). Again, the outcome of interest will be the presence of a positive MADRS response at any time point after baseline (i.e., Day 4, Week 1, Week 2, Week 4, and Week 6). The insomnia grouping variable does not need to be used in this analysis because the modified MADRS response rates do not include a "reduced sleep" item, and if there is significantly higher modified MADRS response rates in the quetiapine group, one can reasonably conclude that the antidepressant effect of low-dose quetiapine was not due to its ability to induce sleep in participates who were experiencing insomnia (there are no other sleep items on the MADRS). Analysis B also provides an alternate analysis strategy for understanding the role of sleep improvement in the efficacy of low-dose quetiapine should the researchers be unable to generate a meaningful insomnia grouping variable to be used in

Analysis A.

Knowing the sample size in the original study was 366 (184 placebo), the researchers would be able to detect a proportion difference of .12 (i.e., .30 for placebo vs .42 for treatment, as per the original study results), which corresponds to a Cohen h of .26, with 80% power using an alpha value of .05, and thus this analysis would be adequately powered (power analysis conducted with G*Power, http://www.gpower.hhu.de/en.html). In essence, the researchers simply want to attempt to replicate the original study findings using modified MADRS response rates.

Strictly speaking, tests of proportions provide full information only when the period at risk for all subjects is the same. With clinical trials, censorship is common. One way of testing for an interaction of treatment group and sleep status is by conducting a stratified log rank test. The researchers will perform the stratified log rank test to examine whether the length of time to MADRS response is the same between sleep groups. A significant chi-square statistic, weighted by size of stratum, would indicate that low dose quetiapine is more effective for those with insomnia.

Hypothesis 2: Is high-dose quetiapine (150 - 300 mg/day) more effective for patients presenting with melancholic depression at treatment baseline? To test this hypothesis the researchers are using data from D1448C00001 (150mg, 300mg, and placebo arms only; Weisler et al., 2009), D1448C00002 (150mg, 300mg, and placebo arms only, excluding duloxetine arm; Cutler et al., 2009), D1448C00003 (quetiapine and placebo arms; Bortnick et al., 2011), and D1448C00004 (quetiapine and placebo arms only, excluding escitalopram; Wang et al., 2012) in order to conduct a pooled analysis. A pooled analysis is preferred as the researchers will then have enough statistical power to detect a subgroup interaction (see analysis plan below). This pooled analysis would include data from 1570 participants (n = 935 for quetiapine 150 mg/day; n = 635 for placebo).

The first step in the analysis for this hypothesis will involve classifying participants into two groups: melancholic depression and nonmelancholic depression. This definition will be applied to all participants in the four RCTs who received quetiapine (150-300 mg/day) or placebo. There should be a balanced number of patients in each group, as randomization should have balanced interventions between any groups defined post-hoc.

The presence of melancholic depression will be determined based on baseline responses to the Hamilton Depression Rating Scale (HAMD-17) and the Montgomery-Asberg Depression Rating Scale (MADRS) according to the DSM-IV-TR criteria by the presence of:

1. One of:

   1. Anhedonia - MADRS Question (Q) 8 (Answers [A] 2-6)
   2. Lack of reactive mood - MADRS Q 1 (A 6) or Q 2 (A 6)

2. At least three of:

   1. Mood worse in morning - HAMD-17 Q 18 (A 1-2 if specify in am)
   2. Terminal insomnia - HAMD-17 Q 6 (A 1-2)
   3. Psychomotor changes - HAMD-17 Q 9 (A 2-4) or Q 8 (A 2-4)
   4. Decreased appetite or weight - HAMD-17 Q 12 (A 2) or Q 16 (A 2)
   5. Excessive guilt - HAMD-17 Q 2 (A 2-4).

The researchers were not able to assess the criterion for DSM-IV melancholic depression that mood be subjectively different from guilt in the assessment because this is not measured by the HAMD-17 or MADRS. However, any participant meeting this subsample definitions would also meet the full DSM-IV-TR criteria. One grouping variable for melancholic depression (coded as 1 "present" and 0 "not present") will be created and all participants will be assigned a value.

Based on past research (Mallinckrodt et al., 2005;McGrath et al., 2008) the researchers expect that the DSM-IV definition of melancholic depression be present in anywhere between 15 to 50% of the sample. This potential imbalance in group size is not problematic for the planned analytic strategy using Cox regression models (unlike having unequal sample sizes in ANOVA, for example - please see below for a more detailed description of the planned analysis) unless there are empty cells in the model (e.g., no responders in the "nonmelancholic" group). This would preclude the estimation of a model, although it would still be clinically useful information.

Melancholia is typically associated with increased disease severity and therefore one might expect a lower prevalence in an RCT that only involves outpatients and excludes participants with significant suicidality and comorbidity. Therefore the researchers have made the criteria for defining the presence of criterion A1 B1 B2 B3 and B5 as loose as possible. Should these definition encompass more than 45% of the sample then the researchers would consider this to be too broad as it is not consistent with previous research regarding the prevalence of melancholic depression that would be expected in the sample. In this scenario the definition will be changed such that a score of 3 is needed on HAMD-17 Q8, Q9, and Q17, and a score of 4 or more is needed on the MADRS Q8. Should this new definition still be inappropriately broad (>45% of sample) the researchers will utilize a third definition that also requires a score of 4 or more on HAMD-17 Q8, Q9, and Q17, a score of 6 or more on MADRS Q8, and a score of 2 on HAMD-17 Q18 (if the specifier indicates mood is worse in the am). Between these three definitions it is likely that there will be an appropriate prevalence under 45% of the sample. If the researchers encounter the opposite problem, such that the prevalence is less than 15%, the researchers will loosen the initial definition such that only 2 of criteria B1, B2, B3, B4, and B5 are needed. If the researchers are still unable to obtain a prevalence between 15 and 45%, or a prevalence in this range but there are empty cells in the Cox model, the researchers will unfortunately not be able to proceed with the corresponding portion of the analysis (Analysis C, D, and E, as described below).

The second step in the analysis will involve calculating the study end points. Two primary end points will be calculated:

1. MADRS response rates, defined as a 50% score reduction from baseline
2. Modified MADRS response rates, defined as a 50% score reduction from baseline calculated without Item 4 (reduced sleep).

A portion of the analysis for Hypothesis 2 with MADRS response rates (Analysis D, see below) will only be conducted if the Cronbach alpha value for the modified MADRS at baseline is at least .80 or within .10 of the unmodified MADRS alpha and still acceptable by standard convention (at least .70). Given the hypothesis tests proposed (see below), response rates will be calculated at each time point from baseline until and including the end of Week 6. Data from trials that extended longer than 6 weeks will not be used.

The third step in the analysis will involve hypothesis testing. MADRS response rates between quetiapine (150-300 mg/day) and placebo groups will be evaluated with a Cox proportional hazard regression analysis. The outcome of interest will be MADRS response at any time point after baseline up until Week 6, which represents the latest trial end-point common to all four studies.

To test the hypothesis that high-dose quetiapine (150-300 mg/day) monotherapy will be more effective for participants with melancholic depression the researchers will conduct a Cox regression analyses (Analysis C) comparing MADRS response rates between the quetiapine (150-300 mg/day) and placebo groups, while including the categorical melancholic depression grouping variable as a covariate and testing for an interaction between treatment group (quetiapine 150-300 mg/day vs placebo) and melancholic depression (present vs not present).

If the researchers assume based on Weisler et al., (2012) a placebo response rate of 30% and a quetiapine 150-300 mg/day response rate of 50% (and N = 1570), a simulated Cox regression predicting overall treatment response, which the researchers ran in Stata 13, yielded a hazard ratio of 1.6 (p < .001). The Stpower command showed that a Cox regression and overall response rate of 42% (50% treatment, 30% placebo) could detect a hazard ratio of 1.6 at a=.05 with 80% power with an N value of 339. It is usually recommended that to detect a subgroup treatment effect with adequate power, the minimum required sample size to detect a treatment response be multiplied by four (Brookes et al., 2004). Therefore the researchers would consider a subgroup analysis with a sample of 1570 (> 4x 339) to be adequate to detect a potential interaction if it exists. The researchers also ran a second simulated Cox regression with a sample of 1570 (30% melancholic; 70% nonmelancholic, and assuming the ratio of 935 treatment:635 placebo would also be true of each subgroup) with an overall response rate of 42%, a 30% response rate in placebo groups (regardless of melancholia status), and a 10% response rate difference between the melancholic treatment group and nonmelancholic treatment group (which, given the predicted number of overall responses in treatment group, leaves a rate of 47% in the nonmelancholic treatment group and 57% in the melancholic treatment group) and found that the tested interaction was statistically significant such that in the treatment group the odds of response was higher for those with melancholic depression vs nonmelancholic depression (HR 1.7, p < .001). Therefore, the researchers consider the proposed analysis appropriately powered to detect a small subgroup treatment effect assuming the predicted overall treatment effect (based on Weisler et al., 2012).

Should the interaction tested in Analysis C be nonsignificant, but there be a raw proportion difference of > .10 for response rates between the melancholic and nonmelancholic groups who received quetiapine 150-300 mg/day, as a secondary analysis for Hypothesis 2, the researchers will also compare MADRS response rates between participants with melancholic depression and nonmelancholic depression who were randomly assigned to receive quetiapine with an independent-samples proportion test (Analysis D). The outcome of interest will be the presence of a positive MADRS response at any time point after baseline up until the end of Week 6. With a presumed prevalence of 30% for melancholic depression, the researchers would be able to detect a proportion difference as small as .10 (Cohen h = .20, which is "small") with 90% power (power analysis conducted with G*Power, http://www.gpower.hhu.de/en.html).

Most if not all participants with melancholic depression will have been experiencing insomnia at baseline. To show that the improved efficacy of high dose quetiapine (150-300 mg/day) for participants with melancholic depression was not simply due to the sleep-inducing properties of quetiapine, the researchers will conduct a Cox regression (Analysis E) comparing modified MADRS response rates between the quetiapine (150-300 mg/day) and placebo groups, while including the categorical melancholic depression grouping variable as a covariate and testing for an interaction between treatment group (quetiapine 150-300 mg/day vs placebo) and melancholic depression (present vs not present). Because the modified MADRS response rates do not include a "reduced sleep" item, if there is a significant interaction suggesting higher modified MADRS response rates in persons with melancholic depression, the researchers can reasonably conclude that the superiority of high-dose quetiapine for people with melancholic depression was not simply due to its ability to induce sleep in these participants.

Survival curves depicting response rates for each Cox model will also be plotted to help visualize the data. As the Cox regression model assumes that the hazard ratio between any treatment and the reference group is constant, the researchers will examine whether this assumption is violated. This will be accomplished by comparing models with and without a treatment x time interaction using a likelihood ratio test. If proportionality is violated, the researchers will use restricted mean survival time analysis as suggested by Royston and Parmar (Royston and Parmar 2013). Whether an interaction is present or not, the researchers will compare the response rates at each week of treatment.

Inclusion Criteria:

• MDD with a HAM-D score of > 21 and > 1 on Item 1

Exclusion Criteria:

• Axis I disorder in last 6 months
• Axis II disorder causing significant impact
• Current depressive episode > 12 months or < 4 weeks
• No response of current episode to > 6 weeks treatment with 2 or more classes of antidepressant medication
• Substance abuse or dependence in last 6 months
• Significant medical illness
• Conditions that may alter study drug metabolism
• Significant suicide or homicide risk
• HAM-D Item 3 score > 2
• Suicide attempt in the last 6 months
• Lab or physical exam abnormalities
• CYP34A inhibitors in last 2 weeks
• Active psychotherapy (not supportive) unless ongoing for > 3 months
• Antipsychotic, antidepressant, or mood stabilizer in last 7 days (28 days for fluoxetine), MAOI or anxiolytic in last 14 days, or deport antipsychotic within 2 dosing intervals.