Study Retrieval

Following the Freedom of Information Act (FOIA) [7], we requested from the FDA all publicly releasable information about the clinical trials for efficacy conducted for marketing approval of fluoxetine, venlafaxine, nefazodone, paroxetine, sertraline, and citalopram, the six most widely prescribed antidepressants approved between 1987 and 1999 [2], which represent all but one of the selective serotonin reuptake inhibitors (SSRIs) approved during the study period. In reply, the agency provided photocopies of the medical and statistical reviews of the sponsors' New Drug Applications. The FDA requires that information on all industry-sponsored trials be submitted as part of the approval process; hence the files sent to us by the FDA should contain information on all trials conducted prior to the approval of each medication. This strategy omits trials conducted after approval was granted.

Although sponsors are required to submit information on all trials, the FDA public disclosure did not include mean changes for nine trials that were deemed adequate and well controlled but that failed to achieve a statistically significant benefit for drug over placebo. Data for four of these trials were available from a pharmaceutical company Web site in January 2007 and were obtained from the GlaxoSmithKline clinical trial register (http://ctr.gsk.co.uk/Summary/paroxetine/studylist.asp).

We also identified published versions of the FDA trials via a PubMed literature search (from January 1985 through May 2007) using the keywords depression; depressive; depressed; and placebo; specific names of antidepressant medications; and names of investigators from the FDA trials. Potentially relevant studies were also identified through references of retrieved and review articles and from a partially overlapping list of published versions of trials submitted to the Swedish drug regulatory authority [5]. Using a standardized protocol, all retrieved abstracts and publications were compared to the FDA trials. The match between each published study and its corresponding FDA trial was independently established with 100% agreement by two investigators (BJD and a research assistant).

Selection

Forty-seven clinical trials were identified in the data obtained from the FDA. The trial flow is illustrated in Figure 1. Inclusion of a drug type for which unsuccessful trials were excluded biases overall results in favor of that drug type, in a way that is akin to publication bias. The purpose in using the FDA dataset is precisely to avoid this type of bias by including all trials of each medication assessed. Therefore, we present analyses only for those medications for which mean change scores on all trials were available.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. QUOROM Flow Chart

https://doi.org/10.1371/journal.pmed.0050045.g001

Validity Assessment

The FDA requires that rigorous standards be followed for the conduct of all efficacy trials for marketing approval [8] and also sets specific agency standards for clinical trials of antidepressant drugs [9]. In addition, the FDA independently reviews the clinical trial methods, statistical procedures, and results. The FDA dataset includes analyses of data from all patients who attended at least one evaluation visit, even if they subsequently dropped out of the trial prematurely. Results are reported from all well-controlled efficacy trials of the use of these medications for the treatment of depression. FDA medical and statistical reviewers had access to the raw data and evaluated the trials independently. The findings of the primary medical and statistical reviewers were verified by at least one other reviewer, and the analysis was also assessed by an independent advisory panel. Following FDA standards, all trials were randomized, double-blind, placebo-controlled trials. None used cross-over designs. Patients had been diagnosed as suffering from unipolar major depressive disorder using Diagnostic and Statistical Manual of Mental Disorders (DSM) criteria.

Given the above review process, we deemed it appropriate to include all studies deemed adequate and well controlled by FDA reviewers, especially as these are the data upon which the decision to approve these medications was based. Other validity criteria might yield different conclusions. In this review, some of the characteristics that may relate to the quality of trials were coded and assessed as possible moderator variables (e.g., interval of trial). The studies have similar methodological characteristics and were well controlled; therefore the methodological characteristics did not affect the final results.

Study Characteristics

In order to generalize the findings of the clinical trial to a larger patient population, FDA reviewers sought a completion rate of 70% or better for these typically 6-wk trials. Only four of the trials reported reaching this objective, and completion rates were not reported for two trials. Attrition rates were comparable between drug and placebo groups. Of those trials for which these rates were reported, 60% of the placebo patients and 63% of the study drug patients completed a 4-, 5-, 6-, or 8-wk trial. Thirty-three trials were of 6-wk duration, six trials were 4 wk, two were 5 wk, and six were 8 wk. Patients were evaluated on a weekly basis. For this meta-analysis, the data were taken from the last visit prior to trial termination.

Thirty-nine trials focused on outpatients: three included both inpatients and outpatients, three were conducted among the elderly (including one of the trials with both inpatients and outpatients), and two were among patients hospitalized for severe depression. No trial was reported for the treatment of children or adolescents.

Replacement of patients who investigators determined were not improving after 2 wk was allowed in three fluoxetine trials and in the three sertraline trials for which data were reported. The trials also included a 1- to 2-wk washout period during which patients were given placebo, prior to random assignment. Those whose scores improved 20% or more were excluded from the study prior to random assignment. The use of other psychoactive medication was reported in 25 trials. In most trials, a chloral hydrate sedative was permitted in doses ranging from 500 mg to 2,000 mg per day. Other psychoactive medication was usually prohibited but still reported as having been taken in several trials.

Meta-Analytic Data Synthesis

We conducted two types of data analysis, one in which each group's change was represented as a standardized mean difference (d), which divides change by the standard deviation of the change score (SD_c) [10], and another using each study's drug and placebo groups' arithmetic mean (weighted for the inverse of the variance) as the meta-analytic “effect size” [11].

The first analysis permitted a determination of the absolute magnitude of change in both the placebo and treatment groups. Results permitted a determination of overall trends, analyses of baseline scores in relation to change, and for both types of models, tests of model specification, which assess the extent to which only sampling error remains unexplained. The results in raw metric are presented comparing both groups, but because of the variation of the SD_cs, the standardized mean difference was used in moderator analyses in order to attain better-fitting models [12]. These results are compared to the criterion for clinical significance used by NICE, which is a three-point difference in Hamilton Rating Scale of Depression (HRSD) scores or a standardized mean difference (d) of 0.50 [1].

As known SD_cs were related to mean baseline HRSD scores, these scores were used to impute missing SD_c values, taking into account both the baseline and its quadratic form and any potential interaction of these terms with group (but in fact, there was no evidence that SD_cs depended on treatment group). One trial reported SD_cs for its drug and placebo groups that were less than 25% the size of the other trials; because preliminary analyses also revealed that this trial was an outlier, these two standard deviations were treated as missing and imputed. In total, SD_cs were known for 28 groups, could be calculated from other inferential statistics in nine comparisons (18 groups), and were imputed in 12 comparisons (24 groups) (47.38%) [13,14].

Overall analyses evaluated both random- and fixed-effects models to assess effect size magnitude; because the same trends appeared for both, for simplicity we present only the fixed-effects results. We also assumed fixed-effects assumptions in order to analyze moderators for both groups. Both Q [15] and I² [16] indices were used to assess inconsistencies from the models, not only to infer the presence or absence of homogeneity, but also (in the case of I²) to assess the degree of inconsistencies among trials [17]. We assumed fixed-effects models in analyzing moderators using meta-regression procedures [11]. Analyses examining linear and quadratic functions for baseline levels of severity used zero-centered forms of this variable [18]. A last, mixed-effects analysis for the amount of change used a random-effects constant along with fixed-effects moderator dimensions; these models provide more conservative assessments of moderation [19].

Because the same scale was used as the primary dependent variable in all of these trials, we were also able to represent results in their original metric [11]. This form of analysis makes results more easily interpretable in terms of clinical significance because mean change scores are analyzed directly, rather than being converted into effect sizes. The analytic weights are derived from the sample size and the SD_c [11]. Finally, to show directly the amount of improvement for each study's drug group against its placebo group, we calculated the difference between the change for the drug group minus the change for the placebo group, leaving the difference in raw units and deriving its analytic weight from its standard error [11,12,20]. Analyses used these weights to examine these controlled outcomes both overall and to determine the extent to which drug-related change is a function of initial severity.

Trial Flow

Mean improvement scores were not available in five of the 47 trials (Figure 1). Specifically, four sertraline trials involving 486 participants and one citalopram trial involving 274 participants were reported as having failed to achieve a statistically significant drug effect, without reporting mean HRSD scores. We were unable to find data from these trials on pharmaceutical company Web sites or through our search of the published literature. These omissions represent 38% of patients in sertraline trials and 23% of patients in citalopram trials. Analyses with and without inclusion of these trials found no differences in the patterns of results; similarly, the revealed patterns do not interact with drug type. The purpose of using the data obtained from the FDA was to avoid publication bias, by including unpublished as well as published trials. Inclusion of only those sertraline and citalopram trials for which means were reported to the FDA would constitute a form of reporting bias similar to publication bias and would lead to overestimation of drug–placebo differences for these drug types. Therefore, we present analyses only on data for medications for which complete clinical trials' change was reported. The dataset comprised 35 clinical trials (five of fluoxetine, six of venlafaxine, eight of nefazodone, and 16 of paroxetine) involving 5,133 patients, 3,292 of whom had been randomized to medication and 1,841 of whom had been randomized to placebo.

Mean Change

Baseline HRSD scores, improvement, and sample sizes in drug and placebo groups for each clinical trial are reported in Table 1. As in the FDA files, studies are identified by protocol numbers. The data from these trials can be obtained from the FDA using FOIA requests and citing the medication name and protocol number. The table also includes references to published reports of the data abstracted from the FDA files, when they could be found (using the search methods described above). Studies in which data only from selected sites of a multisite study were published are not cited in the table. We have also excluded published reports in which dropouts have been removed from the data. For each of the trials, the pharmaceutical companies had submitted to the FDA data in which attrition was handled by carrying forward the last observation carried forward (LOCF) on the patient, which was the basis in all cases of the FDA review. These data and their corresponding citations appear in the table. Even in the LOCF data, there sometimes are some minor discrepancies between the published version and the version submitted to the FDA. In some cases, for example, the N is slightly larger in the published studies than in the data reported to the FDA. Further complicating this problem is the fact that occasionally, the company has published a trial more than once, with slight discrepancies in the data between publications. Data in the table are those reported to the FDA.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1.

Baseline HRSD Scores, Sample Sizes, and Raw and Standardized Improvement with Confidence Intervals, as Reported to the FDA for Drug and Placebo Groups

https://doi.org/10.1371/journal.pmed.0050045.t001

Confirming earlier analyses [2], but with a substantially larger number of clinical trials, weighted mean improvement was 9.60 points on the HRSD in the drug groups and 7.80 in the placebo groups, yielding a mean drug–placebo difference of 1.80 on HRSD improvement scores. Although the difference between these means easily attained statistical significance (Table 2, Model 3a), it does not meet the three-point drug–placebo criterion for clinical significance used by NICE. Represented as the standardized mean difference, d, mean change for drug groups was 1.24 and that for placebo 0.92, both of extremely large magnitude according to conventional standards. Thus, the difference between improvement in the drug groups and improvement in the placebo groups was 0.32, which falls below the 0.50 standardized mean difference criterion that NICE suggested. The amounts of change for drug and placebo groups varied widely around their respective means, Q(34)s = 51.80 and 74.59, p-values < 0.05, and I²s = 34.18 and 54.47. Thus, the mean change exhibited in trials provides a poor description of results, and moderator models are indicated.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2.

Models of Improvement in Depression Scores Based on Group Assignment (Drug versus Placebo) and Initial Depression Severity (as Gauged by HRSD)

https://doi.org/10.1371/journal.pmed.0050045.t002

Drug and Initial Severity Trends in Change

Moderator analyses examined whether drug type, duration of treatment, and baseline severity (HRSD) scores related to improvement. Although drug type and duration of treatment were unrelated to improvement, the drug versus placebo difference remained significant, and amount of improvement was a function of baseline severity (Table 2, Model 1a). Specifically, the amount of improvement depended markedly on the quadratic function of baseline severity, but the linear function of baseline severity interacted with assignment to drug versus placebo (Model 1b). Specifically, as Figure 2 shows, improvement from baseline operated as a ∩-shaped curvilinear function in relation to baseline severity, with those at the lowest and highest levels experiencing smaller gains, whereas those in-between experienced larger gains; the slope for placebo declined as severity increased, whereas the slope for drug was slightly positive. The difference between drug and placebo exceeded NICE's 0.50 standardized mean difference criterion at comparisons exceeding 28 in baseline severity. Further analyses indicated that drug type did not moderate this affect. Although venlafaxine and paroxetine had significantly (p < 0.001) larger weighted mean effect sizes comparing drug to placebo conditions (ds = 0.42 and 0.47, respectively) than fluoxetine (d = 0.22) or nefazodone (0.21), these differences disappeared when baseline severity was controlled.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Mean Standardized Improvement as a Function of Initial Severity and Treatment Group

Drug improvement is portrayed as red triangles around their solid red regression line and placebo improvement as blue circles around their dashed blue regression line; the green shaded area indicates the point at which comparisons of drug versus placebo reach the NICE clinical significance criterion of d = 0.50. Plotted values are sized according to their weight in analyses.

https://doi.org/10.1371/journal.pmed.0050045.g002

For all but one sample, baseline HRSD scores were in the very severe range according to the criteria proposed by the American Psychiatric Association (APA) [21] and adopted by NICE [1]. The one exception derived from a fluoxetine trial that had two samples, one with HRSD scores in the very severe range and the other with scores in the moderate range. Because the low-HRSD condition might be considered an outlier, the analyses were performed again without it. Results continued to reveal that drug versus placebo assignment interacted with initial severity to influence improvement; yet the curvilinear function of the baseline was no longer significant, although group continued to interact with the linear component (Table 2, Model 2c). As Figure 3 shows, drug efficacy did not change as a function of initial severity, whereas placebo efficacy decreased as initial severity increased; values again exceeded NICE's 0.50 standardized mean difference criterion at comparisons greater than 28 in baseline severity. This final model comprising three simultaneous study dimensions (viz., drug vs. placebo, baseline, and the interaction) explained 51.45% of the variation in improvement. Although this model was in a formal sense incorrectly specified (Q_Residual(64) = 96.07, p < 0.01), when a random-effects constant was instead assumed, the same pattern of results remained in this more statistically conservative mixed-effects model. A final model that incorporated even the drug types for which only some trials were available confirmed these trends.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Mean Standardized Improvement as a Function of Initial Severity and Treatment Group, Including Only Trials Whose Samples Had High Initial Severity

Drug improvement is portrayed as red triangles around their solid red regression line and placebo improvement as blue circles around their dashed blue regression line; the green shaded area indicates the point at which comparisons of drug versus placebo reach the NICE clinical significance criterion of d = 0.50. Plotted values are sized according to their weight in analyses.

https://doi.org/10.1371/journal.pmed.0050045.g003

Figure 4 displays raw mean differences between drug and placebo as a function of initial severity, rising as a linear function of baseline severity levels (Table 2, Models 3a and 3b) even though, almost without exception, the scores were in the very severe range of the criteria proposed by APA [21]. Yet when these data are considered in conjunction with those in Figure 3, it seems clear that the increased difference is due to a decrease in improvement in placebo groups, rather than an increase in drug groups.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Mean Drug–Placebo Difference Scores as a Function of Initial Severity

Plotted values are sized according to their sample sizes (n); the green line represents the NICE clinical significance criterion. The solid blue regression line represents the trend across all 35 trials; the dashed red line represents the trend excluding the left-most observation.

https://doi.org/10.1371/journal.pmed.0050045.g004

A visual inspection of Figure 4 suggests that studies' effects are fairly evenly distributed above and below the NICE criterion (3) but that most small studies have high baselines and show large effects. Although sample size (N) was negatively linked to the drug-versus-placebo differences (β = −0.34, p = 0.003), when mean baseline severity values are controlled, this effect disappears and the baseline effect remains significant. The interaction of sample size with baseline severity was marginally significant, p = 0.0586, and the pattern indicated that baseline severity was somewhat more predictive for smaller than for larger studies. Yet, because simple-slopes analyses revealed that baseline scores were significantly predictive even for the largest studies, study differences in sample size would appear to qualify neither the pattern of results we have reported nor their interpretation.

Examination of publication bias often relies on inspections of effect sizes in relation to sample size (or inverse variance) [22]. A funnel plot of the data depicted in Figure 4 indicates that the larger studies in the FDA datasets tended to show smaller drug effects than smaller studies. Although such a pattern might be construed as indicating a publication or other reporting bias, our use of complete datasets precludes this possibility, unless some small trials were not reported despite the FDA Guidelines [6]. A more plausible explanation is that trials with higher baseline scores tended to be small. In any case, funnel-plot inspections assume that there is only one population effect size that can be tracked by a comparison between drug and placebo groups, whereas the current investigation shows that these effects vary widely and that the magnitude of the difference depends on initial severity values. Consequently, funnel-plot inspection is much less appropriate in the present context. Unfortunately, there are no other tools yet available to detect publication or other reporting biases in the face of effect modifiers.